chore: auto-commit (1 archivos)

- logs/ardour_mcp_server.log

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitignore

@@ -54,6 +54,13 @@ reports/*
 !reports/.gitkeep
 projects/*/reports/
 
+# Papers — artefacto local: papers académicos reproducibles. En fase interna viven
+# local y gitignored (como los reports); al promocionar a fase publishable se
+# vuelven sub-repo Gitea propio (como apps/analyses). Solo el marcador .gitkeep se
+# versiona. Convención: docs/capabilities/papers.md
+papers/*
+!papers/.gitkeep
+
 # Node / pnpm
 **/node_modules/
 

bash/functions/io/next_numbered_dir.md

@@ -0,0 +1,58 @@
+---
+name: next_numbered_dir
+kind: function
+lang: bash
+domain: io
+version: "1.0.0"
+purity: impure
+signature: "next_numbered_dir(parent_dir: string, [width: int]) -> string"
+description: "Calcula el siguiente prefijo numerico NNNN- para un directorio numerado incremental. Escanea los subdirectorios directos de parent_dir cuyo nombre empiece por NNNN- (4+ digitos seguidos de guion), toma el maximo, le suma 1 y lo imprime con zero-padding al ancho width (default 4). Si parent_dir no existe o no tiene subdirs que matcheen, imprime 0001."
+tags: [papers, io, scaffold]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+params:
+  - name: parent_dir
+    desc: "directorio padre cuyos subdirectorios numerados (NNNN-...) se escanean; obligatorio"
+  - name: width
+    desc: "ancho del zero-padding del numero impreso (default 4); opcional"
+output: "el siguiente numero como string con zero-padding a width digitos a stdout (ej. 0003); usage a stderr y exit 1 si falta parent_dir"
+tested: false
+tests: []
+test_file_path: ""
+file_path: "bash/functions/io/next_numbered_dir.sh"
+---
+
+## Ejemplo
+
+```bash
+source bash/functions/io/next_numbered_dir.sh
+
+# Sobre un papers/ que ya contiene 0001-foo y 0002-bar
+mkdir -p /tmp/papers/{0001-foo,0002-bar}
+next_numbered_dir /tmp/papers
+# -> 0003
+
+# Directorio vacio o inexistente -> primer numero
+next_numbered_dir /tmp/papers_nuevo
+# -> 0001
+
+# Ancho de padding distinto
+next_numbered_dir /tmp/papers 6
+# -> 000003
+```
+
+## Cuando usarla
+
+Cuando scaffoldees un artefacto numerado incremental (papers/, reports/, issues/) y necesites el siguiente NNNN sin colision: escanea lo que ya existe en disco y te da el numero libre listo para crear `<NNNN>-<slug>`.
+
+## Gotchas
+
+- **Impura**: lee el filesystem (estado del directorio en el momento de la llamada). No crea nada — solo calcula e imprime el numero.
+- **Octal**: los numeros con cero a la izquierda (`08`, `09`) se interpretan como octal en aritmetica bash y romperian el calculo. La funcion fuerza base 10 con `10#$num` para evitarlo.
+- **Solo subdirectorios**: cuenta unicamente subdirs directos. Archivos sueltos (`.gitkeep`, `notas.md`) y subdirs que no matcheen el patron se ignoran. No es recursivo.
+- **Patron estricto**: el prefijo debe ser `NNNN-` (minimo 4 digitos seguidos de guion). Un subdir `12-foo` o `0001foo` (sin guion) NO se cuenta.
+- No hay deteccion de huecos: devuelve `max+1`, no el primer numero libre intermedio. Si tienes `0001` y `0003`, devuelve `0004`, no `0002`.

bash/functions/io/next_numbered_dir.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# next_numbered_dir — Compute the next NNNN- prefix for a numbered directory.
+#
+# Scans the DIRECT subdirectories of <parent_dir> whose names start with a
+# numeric prefix of the form `NNNN-` (4+ digits followed by a hyphen), takes
+# the maximum number, adds 1, and prints it zero-padded to <width> (default 4).
+# If <parent_dir> does not exist or contains no matching subdir, prints the
+# first number (0001 at default width).
+
+next_numbered_dir() {
+    local parent_dir="${1:-}"
+    local width="${2:-4}"
+
+    if [[ -z "$parent_dir" ]]; then
+        echo "usage: next_numbered_dir <parent_dir> [width]" >&2
+        return 1
+    fi
+
+    local max=0
+    local entry base num
+
+    if [[ -d "$parent_dir" ]]; then
+        # Iterate only over direct subdirectories. The trailing slash in the
+        # glob ensures files (e.g. .gitkeep) are skipped — only dirs match.
+        for entry in "$parent_dir"/*/; do
+            # If the glob matched nothing it stays literal; guard with -d.
+            [[ -d "$entry" ]] || continue
+            base="$(basename "$entry")"
+            # Require a prefix of 4+ digits followed by a hyphen.
+            if [[ "$base" =~ ^([0-9]{4,})- ]]; then
+                num="${BASH_REMATCH[1]}"
+                # Force base 10 so leading zeros (08, 09) are not read as octal.
+                num=$((10#$num))
+                if (( num > max )); then
+                    max=$num
+                fi
+            fi
+        done
+    fi
+
+    printf "%0*d\n" "$width" $(( max + 1 ))
+}
+
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+    next_numbered_dir "$@"
+fi

bash/functions/pipelines/init_paper.md

@@ -0,0 +1,69 @@
+---
+name: init_paper
+kind: pipeline
+lang: bash
+domain: pipelines
+version: "1.0.0"
+purity: impure
+signature: "init_paper(slug: string, [--title <t>] [--domain <d>] [--tags <csv>]) -> void"
+description: "Scaffold de un paper académico reproducible en papers/<NNNN-slug>/. Calcula el siguiente número incremental escaneando papers/, crea las subcarpetas (experiments data figures reviews out), copia las plantillas paper.md (IMRaD) + preregistration.md (anti-HARKing) rellenando el frontmatter (title, slug, date de hoy, phase=question, status=draft) y crea references.md. NO hace git init: el paper arranca en fase interna local (papers/ gitignored). Grupo de capacidad papers."
+tags: [papers, scaffold, paper, pipeline, bash, launcher]
+uses_functions:
+  - next_numbered_dir_bash_io
+  - slugify_ascii_py_core
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+params:
+  - name: slug
+    desc: "identificador legible del paper; se slugifica a ASCII (espacios/acentos se normalizan) y se prefija con el siguiente NNNN incremental"
+  - name: "--title"
+    desc: "título del paper (string); si se omite, usa el slug limpio. No debe contener el carácter '|'"
+  - name: "--domain"
+    desc: "dominio del paper escrito en el frontmatter (default datascience)"
+  - name: "--tags"
+    desc: "tags CSV que se escriben en el frontmatter de paper.md (opcional)"
+output: "sin salida directa; crea papers/<NNNN-slug>/ con paper.md, preregistration.md, references.md y las subcarpetas experiments/ data/ figures/ reviews/ out/. Imprime el resumen y los pasos siguientes a stdout."
+tested: false
+tests: []
+test_file_path: ""
+file_path: "bash/functions/pipelines/init_paper.sh"
+---
+
+## Ejemplo
+
+```bash
+# Scaffold de un paper nuevo (numera 0001, 0002, ... automáticamente)
+fn run init_paper mi-primer-paper --title "Mi primer paper"
+fn run init_paper reactive-loop-calls --domain datascience --tags registry,telemetria
+
+# El slug se slugifica: "Áreas de Mejora" -> papers/0003-areas-de-mejora/
+fn run init_paper "Áreas de Mejora"
+```
+
+## Cuando usarla
+
+Cuando empiezas un paper académico nuevo dentro de `fn_registry` y necesitas el esqueleto del artefacto (`papers/<NNNN-slug>/`) con las plantillas IMRaD y de pre-registro listas para rellenar. Es el paso 1 del grupo de capacidad `papers` (ver `docs/capabilities/papers.md`), antes de la revisión de literatura y del pre-registro de la hipótesis.
+
+## Flujo
+
+1. Parsea `<slug>` (posicional) + flags `--title` / `--domain` / `--tags`. Falla con exit ≠ 0 si falta el slug.
+2. `slugify_ascii` — normaliza el slug a ASCII lowercase sin diacríticos (reutiliza la función del registry, solo stdlib).
+3. `next_numbered_dir papers/` — calcula el siguiente NNNN de 4 dígitos sin colisión.
+4. Crea `papers/<NNNN-slug>/` con las subcarpetas `experiments/ data/ figures/ reviews/ out/`.
+5. Copia `docs/templates/paper.md` + `docs/templates/preregistration.md` y rellena el frontmatter por clave de línea (title, slug, date de hoy, domain, tags; phase=question y status=draft vienen de la plantilla).
+6. Crea `references.md` vacío.
+
+## Gotchas
+
+- **NO hace `git init`.** El paper arranca en fase interna local; `papers/` está gitignored en el repo padre (solo `papers/.gitkeep` se versiona). Promocionar a sub-repo Gitea (fase publishable) es manual.
+- **El `--title` no debe contener el carácter `|`** (se usa como delimitador de sed al rellenar el frontmatter; los `&` y `\` sí se escapan).
+- **No indexa el paper en `registry.db`** — los artefactos `papers/<slug>/` no se indexan en esta fase (KISS); sí se indexa este pipeline.
+- Requiere `python3` (del venv del registry o del sistema) para slugificar; `slugify_ascii` solo usa stdlib, así que el venv no es obligatorio.
+- Idempotencia: si el directorio destino ya existiera, aborta con exit ≠ 0 en vez de sobrescribir.
+
+## Notas
+
+Cada paper es un artefacto independiente (mismo patrón que `apps/` y `analysis/`, pero para investigación). El pipeline usa `set -euo pipefail`: cualquier fallo detiene la ejecución. Parte del grupo de capacidad `papers` — diseño completo en `reports/0001-2026-06-30-papers-system-design.md`.

bash/functions/pipelines/init_paper.sh

@@ -0,0 +1,177 @@
+#!/usr/bin/env bash
+# init_paper
+# ----------
+# Scaffold de un paper académico reproducible en papers/<NNNN-slug>/.
+#
+# Calcula el siguiente número incremental escaneando papers/, crea el
+# directorio con todas las subcarpetas (experiments data figures reviews out),
+# copia las plantillas paper.md + preregistration.md rellenando el frontmatter
+# (title, slug, date de hoy, phase=question, status=draft) y crea references.md.
+#
+# NO hace `git init`: el paper arranca en fase interna local (papers/ está
+# gitignored en el repo padre, solo .gitkeep se versiona). La promoción a
+# sub-repo Gitea (fase publishable) es un paso posterior MANUAL.
+#
+# Compone: next_numbered_dir (helper de numeración del registry) +
+#          slugify_ascii (slug ASCII del registry).
+#
+# USO:
+#   ./init_paper.sh <slug> [--title "..."] [--domain <d>] [--tags a,b,c]
+#
+# EJEMPLOS:
+#   ./init_paper.sh mi-primer-paper --title "Mi primer paper"
+#   ./init_paper.sh reactive-loop-calls --domain datascience --tags registry,telemetria
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REGISTRY_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+
+# Funciones atómicas del registry
+source "$REGISTRY_ROOT/bash/functions/io/next_numbered_dir.sh"
+
+# ── Parsing de argumentos ────────────────────────────────────
+
+SLUG_RAW=""
+TITLE=""
+DOMAIN="datascience"
+TAGS=""
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --title)
+            TITLE="$2"; shift 2 ;;
+        --domain)
+            DOMAIN="$2"; shift 2 ;;
+        --tags)
+            TAGS="$2"; shift 2 ;;
+        -h|--help)
+            grep "^#" "$0" | sed 's/^# \?//' ; exit 0 ;;
+        -*)
+            echo "Flag desconocido: $1" >&2 ; exit 1 ;;
+        *)
+            if [ -z "$SLUG_RAW" ]; then
+                SLUG_RAW="$1"
+            else
+                echo "ERROR: argumento posicional inesperado: '$1' (solo se admite un <slug>)." >&2
+                exit 1
+            fi
+            shift ;;
+    esac
+done
+
+if [ -z "$SLUG_RAW" ]; then
+    echo "ERROR: falta el argumento <slug>." >&2
+    echo "Uso: $0 <slug> [--title \"...\"] [--domain <d>] [--tags a,b,c]" >&2
+    echo "  Ejemplo: $0 mi-primer-paper --title \"Mi primer paper\"" >&2
+    exit 1
+fi
+
+# ── Slugificar (reutiliza slugify_ascii del registry; solo stdlib) ──
+
+PYBIN="$REGISTRY_ROOT/python/.venv/bin/python3"
+[ -x "$PYBIN" ] || PYBIN="$(command -v python3 || true)"
+if [ -z "$PYBIN" ]; then
+    echo "ERROR: no se encontró python3 para slugificar el slug." >&2
+    exit 1
+fi
+
+SLUG_CLEAN=$("$PYBIN" -c '
+import sys, os
+sys.path.insert(0, os.path.join(sys.argv[2], "python", "functions"))
+from core.slugify_ascii import slugify_ascii
+print(slugify_ascii(sys.argv[1], default="paper"))
+' "$SLUG_RAW" "$REGISTRY_ROOT")
+
+# ── Resolver número incremental y directorio destino ─────────
+
+PAPERS_DIR="$REGISTRY_ROOT/papers"
+mkdir -p "$PAPERS_DIR"
+
+NUM=$(next_numbered_dir "$PAPERS_DIR")
+SLUG_FULL="${NUM}-${SLUG_CLEAN}"
+PAPER_DIR="$PAPERS_DIR/$SLUG_FULL"
+
+if [ -d "$PAPER_DIR" ]; then
+    echo "ERROR: el directorio del paper ya existe: $PAPER_DIR" >&2
+    exit 1
+fi
+
+TODAY=$(date +%Y-%m-%d)
+[ -n "$TITLE" ] || TITLE="$SLUG_CLEAN"
+
+TAGS_YAML="[]"
+if [ -n "$TAGS" ]; then
+    TAGS_YAML="[$(echo "$TAGS" | sed 's/,/, /g')]"
+fi
+
+echo ""
+echo "════════════════════════════════════════════════════════════"
+echo "  INIT PAPER: ${SLUG_FULL}"
+echo "  Título:     ${TITLE}"
+echo "  Directorio: ${PAPER_DIR}"
+echo "════════════════════════════════════════════════════════════"
+echo ""
+
+# ── Crear estructura ─────────────────────────────────────────
+
+echo "[1/3] Creando estructura..."
+mkdir -p "$PAPER_DIR"/experiments "$PAPER_DIR"/data "$PAPER_DIR"/figures \
+         "$PAPER_DIR"/reviews "$PAPER_DIR"/out
+echo "  experiments/ data/ figures/ reviews/ out/"
+
+# ── Copiar plantillas + rellenar frontmatter ─────────────────
+
+echo "[2/3] Escribiendo paper.md + preregistration.md..."
+
+# Escapa caracteres especiales del RHS de sed (delimitador |)
+sed_escape() { printf '%s' "$1" | sed -e 's/[\\&|]/\\&/g'; }
+TITLE_ESC="$(sed_escape "$TITLE")"
+DOMAIN_ESC="$(sed_escape "$DOMAIN")"
+
+PAPER_MD="$PAPER_DIR/paper.md"
+PREREG_MD="$PAPER_DIR/preregistration.md"
+
+cp "$REGISTRY_ROOT/docs/templates/paper.md" "$PAPER_MD"
+cp "$REGISTRY_ROOT/docs/templates/preregistration.md" "$PREREG_MD"
+
+sed -i \
+    -e "s|^title:.*|title: \"${TITLE_ESC}\"|" \
+    -e "s|^slug:.*|slug: ${SLUG_FULL}|" \
+    -e "s|^date:.*|date: ${TODAY}|" \
+    -e "s|^domain:.*|domain: ${DOMAIN_ESC}|" \
+    -e "s|^tags:.*|tags: ${TAGS_YAML}|" \
+    "$PAPER_MD"
+
+sed -i \
+    -e "s|^paper_slug:.*|paper_slug: ${SLUG_FULL}|" \
+    "$PREREG_MD"
+
+echo "  $PAPER_MD"
+echo "  $PREREG_MD"
+
+# ── references.md ────────────────────────────────────────────
+
+echo "[3/3] Escribiendo references.md..."
+cat > "$PAPER_DIR/references.md" << EOF
+# References — ${TITLE}
+
+<!-- Una entrada por referencia. Formato libre (o BibTeX) hasta promocionar a publishable. -->
+EOF
+echo "  $PAPER_DIR/references.md"
+
+# ── Resumen ──────────────────────────────────────────────────
+
+echo ""
+echo "════════════════════════════════════════════════════════════"
+echo "  PAPER '${SLUG_FULL}' LISTO (fase: question, status: draft)"
+echo "════════════════════════════════════════════════════════════"
+echo ""
+echo "  Pasos siguientes:"
+echo "  1. Revisión de literatura (skill /deep-research) → Related work."
+echo "  2. Pre-registro: congela H0/H1 + plan en preregistration.md (preregister_hypothesis)."
+echo "  3. Experimentos en experiments/ → análisis (grupo eda) → escritura IMRaD en paper.md."
+echo "  4. render_paper_pdf → out/paper.pdf. Peer review adversarial → reviews/."
+echo ""
+echo "  papers/ está gitignored: este paper vive local hasta promocionar a publishable."
+echo ""

docs/automatic_eda_contract.md

@@ -41,12 +41,13 @@ reconocido se degrada a `Note`, nunca lanza).
 | `Heading(text, level=1)` | título de sección, `level` 1 (grande) … 3 (chico) | una o varias líneas en negrita; nivel 1 lleva subrayado de acento |
 | `Markdown(text)` | texto markdown ligero | ver subset abajo; **nunca corta a media línea** |
 | `KVTable(rows, title=None)` | `rows = [(clave, valor), ...]` | tabla de 2 columnas etiqueta/valor; el valor se envuelve |
-| `DataTable(header, rows, title=None, note=None)` | `header=[...]`, `rows=[[...],...]` | tabla con cabecera; **se parte por filas repitiendo cabecera**; las celdas largas se envuelven dentro de su columna |
+| `DataTable(header, rows, title=None, note=None)` | `header=[...]`, `rows=[[...],...]` | tabla con cabecera; **si cabe** como texto se parte por filas repitiendo cabecera; **si NO cabe** (demasiadas columnas) se rasteriza entera como imagen de alta resolución para hacer zoom. Ver §11.4 |
 | `Figure(fig=None, make=None, caption=None, height_in=None)` | una `matplotlib.figure.Figure` ya construida (`fig`) o un callable `make()->Figure` (perezoso) | se rasteriza y escala para caber entera (nunca recortada) |
 | `Image(path, caption=None, height_in=None)` | ruta a PNG/JPG | se escala para caber entera |
 | `Caption(text)` / `Note(text)` | texto auxiliar pequeño | pie/nota en gris; `Note` es además el fallback de lo desconocido |
-| `Group(blocks, title=None)` | unidad **keep-together**: sus bloques se mantienen juntos | el renderer mide el grupo entero y lo mueve completo a la página/slide siguiente si no cabe; encoge la figura para dejar sitio al título+texto. Ver §11 |
+| `Group(blocks, title=None, page_break_before=False, layout="stack")` | unidad **keep-together**: sus bloques se mantienen juntos | el renderer mide el grupo entero y lo mueve completo a la página/slide siguiente si no cabe; encoge la figura para dejar sitio al título+texto. `layout="side_by_side"` coloca tabla+figura en dos columnas (solo PPTX). Ver §11 y §11.4 |
 | `GlossaryEntry(key, label, definition)` | una entrada del glosario (destino clicable) | la genera el capítulo `glosario`; registra su posición como destino de los términos marcados. Ver §11 |
+| `TocEntry(label, target_id)` | una entrada de **índice clicable** en la portada | la genera el capítulo `portada`; el renderer la cablea como salto al inicio del capítulo cuyo `id` o `title` coincide con `target_id`. Ver §11.4 |
 
 `Figure`/`Image` aceptan `height_in` (hint): el renderer **clampa** la figura a esa altura máxima (lo usa `Group` para encoger la figura). Toda figura escala dejando sitio a su caption en la misma página/slide; en PPTX el caption es **siempre** visible (si no se da `caption`, cae al último heading o a "Figura").
 
@@ -397,6 +398,65 @@ cabecera con su fondo propio. Es automático en PDF y PPTX; el patrón se mantie
 cuando una tabla larga se parte y repite cabecera (el índice de fila es lógico, no por
 página). No hay nada que hacer en los capítulos.
 
+### 11.4 Calidad de render global: DPI alto, tabla ancha → imagen, figura al lado, índice clicable
+
+Cuatro capacidades transversales del motor, **todas automáticas salvo `layout`** (que un
+capítulo activa explícitamente). Aplican a PDF y PPTX salvo donde se indique.
+
+**(a) DPI alto (automático).** Toda figura/imagen embebida se rasteriza a **220 dpi**
+(constante `_RASTER_DPI` en ambos renderers; en PDF se aplica también al `savefig` de la
+página, porque matplotlib re-rasteriza cada `imshow` al escribir la página). Objetivo:
+ampliar en el móvil y leer detalle (ejes, celdas) sin pixelar. El texto sigue siendo
+vectorial y seleccionable. No hay nada que hacer en los capítulos.
+
+**(b) Tabla ancha → imagen de alta resolución (automático).** Cuando un `DataTable` tiene
+**demasiadas columnas para ser legible como texto** en el ancho útil (criterio
+`_table_fits_as_text`: ancho mínimo legible por columna × nº de columnas > ancho útil; en
+la práctica salta sobre tablas tipo `df.head` con muchas columnas), en vez de comprimir las
+columnas hasta hacerlas ilegibles, la tabla se dibuja **entera como una imagen de alta
+resolución** (función `render_table_as_figure_py_datascience`: cabecera sombreada + zebra)
+escalada para caber completa, de modo que el lector hace **zoom** y la lee sin perder datos.
+Si la tabla **sí cabe**, se mantiene como texto seleccionable (PDF) / tabla nativa (PPTX).
+Las `KVTable` (2 columnas) caben siempre y se quedan como texto. No hay nada que hacer en
+los capítulos.
+
+**(c) Figura al lado de la tabla — `Group(layout="side_by_side")`.** Hint de layout que un
+capítulo activa para que su **tabla quede a la izquierda y su figura a la derecha** en la
+misma diapositiva, en lugar de apiladas:
+
+```python
+model.Group(
+    layout="side_by_side",
+    blocks=[
+        model.Heading(text=str(name), level=2),       # va a ancho completo arriba
+        model.DataTable(header=..., rows=...),         # columna IZQUIERDA (~55%)
+        model.Figure(make=_grafico_perezoso(...)),     # columna DERECHA (~45%)
+        model.Markdown(text="explicación…"),           # va a ancho completo abajo
+    ])
+```
+
+Contrato exacto del campo:
+
+| Campo | Valor | Efecto |
+|---|---|---|
+| `layout` | `"stack"` (por defecto) | comportamiento histórico: apilado vertical (keep-together). |
+| `layout` | `"side_by_side"` | **PPTX**: la tabla (rasterizada a imagen) ocupa la columna izquierda (~55% del ancho útil) y la figura la derecha (~45%); cualquier otro bloque (heading, markdown) va a ancho completo arriba/abajo. Si no hay un par tabla+figura, o no caben lado a lado en una slide, **cae automáticamente a apilado**. **PDF**: se trata **igual que `stack`** (el ancho A5 móvil no admite dos columnas legibles). Valores desconocidos degradan a `"stack"`. |
+
+Es **retrocompatible**: un `Group` sin `layout` (o `layout="stack"`) se comporta exactamente
+como antes. El capítulo `cat_distr` es el consumidor previsto (gráfico a la derecha de la
+tabla de categorías en PPT); este motor solo provee el soporte.
+
+**(d) Índice clicable en la portada — `TocEntry`.** La portada emite un `Heading("Índice")`
+seguido de un `TocEntry(label, target_id)` por capítulo. El renderer registra la
+página/slide de inicio de **cada** capítulo (indexado por `id` **y** por `title`) y cablea
+cada `TocEntry` como un salto real a ese inicio: en **PDF** vía
+`add_pdf_internal_links_py_datascience` (link GOTO de PyMuPDF), en **PPTX** vía
+`pptx_link_run_to_slide_py_datascience` (salto a slide nativo). Como la portada solo conoce
+los **títulos** de los capítulos, el `target_id` se hace coincidir contra el `title` (o el
+`id`) de destino. Si un destino no resuelve, la entrada se muestra igualmente como texto
+(en color de enlace), nunca se corta. Es el mismo mecanismo que los términos clicables del
+glosario (§11.1), reutilizado en sentido portada → capítulo.
+
 ---
 
 ## 10. Integración futura con `profile_table` (siguiente fase)

docs/capabilities/INDEX.md

@@ -39,6 +39,7 @@ Indice de grupos de capacidades del registry. Cada grupo agrupa >=3 funciones qu
 | [cpp-tables](tql.md) | 9 | Table Query Language C++ puro: filter, group, agg, sort, join, stats, formulas Lua, round-trip emit/apply |
 | [data-table-renderers](data_table_renderers.md) | 1 | API declarativa de cell renderers para data_table: Badge, Progress, Duration, Icon via TableInput.column_specs |
 | [scheduler](scheduler.md) | 4 | Cron expression parsing, matching, next-run y traduccion humana (consume `apps/dag_engine`) |
+| [papers](papers.md) | — | Papers académicos reproducibles en `papers/<NNNN-slug>/`: scaffold del artefacto (`init_paper` + helper `next_numbered_dir`), plantillas IMRaD + pre-registro anti-HARKing, y (en construcción por la flota) congelar hipótesis, funciones estadísticas (effect size/CI/corrección múltiple), render md→PDF y peer-review adversarial. Reutiliza `deep-research`, grupo `eda` y el motor PDF de `datascience`. Diseño: `reports/0001-2026-06-30-papers-system-design.md` |
 | [extractor](extractor.md) | 15 | Funciones que leen datos de fuentes externas (BD, API, archivos, web). Nodos input de `data_factory` |
 | [transformer](transformer.md) | 15 | Funciones que clean/dedup/aggregate/feature-engineer datos. Nodos intermedios de `data_factory` |
 | [sink](sink.md) | 11 | Funciones que escriben datos a destino externo (BD, dashboard, alerta, email). Nodos output |

docs/capabilities/papers.md

@@ -0,0 +1,82 @@
+# papers — papers académicos reproducibles
+
+Grupo de capacidad para producir **papers académicos** dentro de `fn_registry`: investigación con hipótesis falsables, experimentos reproducibles, análisis estadístico honesto y escritura en formato IMRaD. Cada paper es un artefacto nuevo en `papers/<NNNN-slug>/` que reutiliza infraestructura existente (skill `deep-research` para la revisión de literatura, grupo `eda` para el análisis, motor md→PDF de `datascience`, patrón de verificación adversarial del orquestador) y añade lo que falta como funciones del registry.
+
+Diseño completo y decisiones: `reports/0001-2026-06-30-papers-system-design.md`.
+
+> **Regla de oro anti paper-mill:** una hipótesis que **podía** fallar + un experimento con riesgo real de refutación + estadística que no es teatro. Si no hay riesgo de refutación, no es un paper. Los claims nunca superan a la evidencia. El antídoto al HARKing es el **pre-registro**: el plan de análisis se congela *antes* de mirar los datos.
+
+## Estructura del artefacto
+
+```
+papers/0001-mi-paper/
+  paper.md            # frontmatter (title, slug, authors, date, status, phase, tags, domain, hypothesis_id) + cuerpo IMRaD
+  preregistration.md  # H0/H1 + plan de análisis CONGELADO (frozen_at + content_hash) antes de correr
+  references.md       # bibliografía
+  experiments/        # código / notebooks por experimento (exp01_*, exp02_*)
+  data/               # crudos + procesados (gitignored si pesa)
+  figures/            # gráficos generados
+  reviews/            # outputs del peer-review adversarial
+  out/                # paper.pdf — entregable final
+  .git/               # SOLO cuando promociona a fase publishable (sub-repo Gitea)
+```
+
+`papers/` está gitignored en el repo padre (solo `papers/.gitkeep` se versiona): un paper en fase interna no contamina el repo. Al promocionar a `status: publishable` se vuelve sub-repo Gitea `dataforge/<slug>` (como apps y analyses).
+
+### Fases (campo `phase` de `paper.md`)
+
+```
+question → review → hypothesis → design → running → analysis → writing → internal-review
+  → [DONE interno]  → polish → submitted          [solo en fase publishable]
+```
+
+## Funciones
+
+| ID | Pureza | Estado | Qué hace |
+|---|---|---|---|
+| `init_paper_bash_pipelines` | impure | ✅ disponible | Scaffold de `papers/<NNNN-slug>/`: calcula el siguiente NNNN, crea las subcarpetas, copia `paper.md` + `preregistration.md` con el frontmatter relleno (slug, title, date de hoy, `phase: question`, `status: draft`) y `references.md` vacío. NO hace `git init` (el paper arranca en fase interna local). |
+| `next_numbered_dir_bash_io` | impure | ✅ disponible | Dado un directorio, devuelve el siguiente número incremental de 4 dígitos (`0001`, `0002`, …) escaneando los subdirs con prefijo `NNNN-`. Helper de numeración de `init_paper` (reutilizable por reports/issues). |
+| `preregister_hypothesis` | impure | 🚧 en construcción (flota) | Congela el `preregistration.md` (H0/H1 + plan de análisis) con `frozen_at` + `content_hash`, pasa `status` a `frozen` y escribe `hypothesis_id` en `paper.md`. Mata el HARKing: tras congelar, el plan no se edita. |
+| `cohens_d` (effect size) | pure | 🚧 en construcción (flota) | Tamaño del efecto (Cohen's d) entre dos grupos. Reporta magnitud, no solo significancia. |
+| `confidence_interval` | pure | 🚧 en construcción (flota) | Intervalo de confianza de una métrica (media/diferencia). |
+| `holm_bonferroni` | pure | 🚧 en construcción (flota) | Corrección de comparaciones múltiples (Holm-Bonferroni / FWER) para el plan de análisis. |
+| `render_paper_pdf` | impure | 🚧 en construcción (flota) | Markdown IMRaD (`paper.md` + figuras) → `out/paper.pdf`, reutilizando el motor md→PDF del grupo `eda`/`datascience`. |
+
+> Las funciones estadísticas reutilizan lo que ya exista en `datascience` (p.ej. `fdr_correction_py_datascience` cubre la corrección de comparaciones múltiples por FDR; el agente del rigor experimental decide si añade Holm-Bonferroni o reusa lo existente). Buscar antes de duplicar: `mcp__registry__fn_search query="effect size" domain="datascience"`.
+
+### Peer review (no es función del registry)
+
+El agente adversarial `.claude/agents/paper-reviewer.md` (🚧 en construcción por la flota) puntúa novedad, rigor, reproducibilidad y validez, e intenta **refutar** cada claim. Default a "failed" si la evidencia no soporta. Escribe su veredicto en `reviews/`. Es el equivalente al verificador adversarial del orquestador aplicado al paper.
+
+## Ejemplo canónico (end-to-end)
+
+```bash
+# 1. Scaffold del paper (fase question, local). Crea papers/0001-mi-paper/.
+./fn run init_paper mi-paper --title "¿El bucle reactivo reduce las calls inline?" --domain datascience --tags registry,telemetria
+
+# 2. Revisión de literatura → llena Related work (skill deep-research, fase review).
+#    /deep-research "..."
+
+# 3. Pre-registro: congela H0/H1 + plan de análisis ANTES de mirar datos (fase hypothesis).
+./fn run preregister_hypothesis papers/0001-mi-paper      # 🚧 en construcción
+
+# 4. Experimentos en papers/0001-mi-paper/experiments/ (fase running) →
+#    análisis con el grupo `eda` + funciones de effect size / CI / corrección múltiple (fase analysis).
+
+# 5. Escritura IMRaD en paper.md (fase writing) → render del entregable PDF.
+./fn run render_paper_pdf papers/0001-mi-paper            # 🚧 en construcción → out/paper.pdf
+
+# 6. Peer review adversarial (fase internal-review).
+#    Agent(subagent_type="paper-reviewer", prompt="Revisa papers/0001-mi-paper ...")  # 🚧 en construcción
+```
+
+## Fronteras
+
+- **NO es para reports de trabajo.** Un report (`reports/`) es el entregable escrito de una tarea (resumen + evidencia + gaps); un paper es investigación con hipótesis falsable y experimento. Ver `.claude/rules/reports.md`.
+- **NO se indexa en `registry.db` en esta fase.** No hay tabla `papers` ni `entity_type` `paper` (KISS); se añadiría con migración propia si se decide. Las *funciones* del grupo sí se indexan (viven en `bash/functions/`, `python/functions/`), pero los artefactos `papers/<slug>/` no.
+- **NO hace `git init` en el scaffold.** El paper arranca en fase interna local y gitignored. La promoción a sub-repo Gitea (fase publishable) es un paso manual posterior.
+- **NO soporta LaTeX/arXiv todavía.** Formato elegido: Markdown como fuente + PDF como entregable. El soporte LaTeX se añadiría al promocionar un paper a fase publishable.
+
+## Estado
+
+Fase de scaffolding. Disponible: estructura del artefacto, plantillas (`docs/templates/paper.md`, `docs/templates/preregistration.md`), pipeline `init_paper` + helper `next_numbered_dir`, esta página y el bloque gitignore de `papers/`. En construcción por la flota: `preregister_hypothesis`, funciones estadísticas (effect size / CI / corrección múltiple), `render_paper_pdf` y el agente `paper-reviewer`. Validación end-to-end con un paper piloto real: pendiente.

docs/templates/paper.md

@@ -0,0 +1,94 @@
+---
+title: "TITULO DEL PAPER"
+slug: NNNN-slug
+authors: [Enmanuel]
+date: 2026-01-01
+status: draft          # draft | internal | publishable
+phase: question        # question -> review -> hypothesis -> design -> running -> analysis -> writing -> internal-review -> polish -> submitted
+tags: []
+domain: datascience
+hypothesis_id: ""      # lo rellena preregister_hypothesis al congelar el preregistro
+---
+
+<!--
+Paper académico reproducible (formato IMRaD). Esta es la FUENTE editable en Markdown;
+el entregable PDF se genera con render_paper_pdf (grupo `papers`).
+
+Regla de oro anti paper-mill: una hipótesis que PODÍA fallar + un experimento con
+riesgo real de refutación + estadística que no es teatro. Si no hay riesgo de
+refutación, no es un paper. Los claims nunca superan a la evidencia.
+-->
+
+# {{título del paper}}
+
+## Abstract
+
+<!--
+Resumen estructurado en 4-6 frases: contexto -> gap -> método -> resultados -> conclusión.
+Sin citas, sin abreviaturas sin definir. Es lo único que mucha gente leerá: que se sostenga solo.
+-->
+
+## 1. Introduction
+
+<!--
+Embudo en cuatro movimientos:
+1. Contexto — el área y por qué importa.
+2. Gap — qué NO se sabe todavía (el hueco que este paper llena).
+3. Pregunta / hipótesis — formulada de forma falsable (ver preregistration.md).
+4. Contribución — lista explícita de lo que aporta este trabajo ("Contributions:").
+-->
+
+## 2. Related work
+
+<!--
+Qué existe ya y por qué no basta. Agrupa por enfoque, no por autor. Cada cita debe
+justificar por qué el gap sigue abierto. Output de la fase de revisión (skill deep-research).
+-->
+
+## 3. Methods
+
+<!--
+Diseño REPRODUCIBLE: otra persona lo corre y obtiene lo mismo.
+- Variables: independiente(s), dependiente(s), control.
+- Diseño: N, condiciones, muestreo, aleatorización.
+- Métricas y cómo se miden.
+- Protocolo paso a paso + dónde vive el código (experiments/) y los datos (data/).
+Debe ser coherente con el preregistration.md congelado (no se cambia el plan tras ver datos).
+-->
+
+## 4. Results
+
+<!--
+Datos SIN interpretar. Tablas y figuras (figures/) con su lectura literal.
+Reporta effect size + intervalos de confianza, no solo p-valores.
+Incluye también los resultados negativos / no significativos (anti cherry-picking).
+-->
+
+## 5. Discussion
+
+<!--
+Interpretación de los resultados a la luz de la pregunta. Claims <= evidencia.
+-->
+
+### 5.1 Limitaciones
+
+<!-- Qué no cubre el estudio, supuestos, datos faltantes. Honestidad explícita. -->
+
+### 5.2 Amenazas a la validez
+
+<!--
+- Validez interna — ¿la causa es lo que decimos o hay confusores?
+- Validez externa — ¿generaliza fuera de esta muestra/condiciones?
+- Validez de constructo — ¿la métrica mide lo que dice medir?
+- Validez estadística — ¿N suficiente, supuestos del test cumplidos, comparaciones múltiples corregidas?
+-->
+
+## 6. Conclusion + Future work
+
+<!--
+Cierre en 2-4 frases: qué se aprendió (sin overclaiming) + las siguientes preguntas que abre.
+-->
+
+## References
+
+<!-- Ver references.md. -->

docs/templates/preregistration.md

@@ -0,0 +1,59 @@
+---
+paper_slug: NNNN-slug
+frozen_at: ""          # timestamp ISO — lo rellena preregister_hypothesis al congelar
+content_hash: ""       # hash del contenido congelado — lo rellena preregister_hypothesis
+status: draft          # draft -> frozen (preregister_hypothesis lo pasa a frozen; tras congelar NO se edita)
+---
+
+> **⚠️ ESTE DOCUMENTO SE CONGELA ANTES DE MIRAR LOS DATOS (anti-HARKing).**
+> El plan de análisis se fija aquí *antes* de ejecutar el experimento. Una vez congelado
+> (`status: frozen`, con `frozen_at` + `content_hash`), **no se edita**. Inventar o ajustar
+> la hipótesis después de ver los resultados (HARKing) invalida el paper. Si el plan cambia
+> tras ver datos, eso es análisis exploratorio y se reporta como tal, no como confirmatorio.
+
+# Pre-registro — {{título del paper}}
+
+## 1. Pregunta de investigación
+
+<!-- La pregunta concreta, en una frase. Debe poder responderse con un experimento. -->
+
+## 2. Hipótesis
+
+<!-- Falsable (Popper): una predicción que PODRÍA fallar. -->
+
+- **H0 (nula):** <!-- no hay efecto / no hay diferencia. Es lo que el test intenta rechazar. -->
+- **H1 (alternativa):** <!-- el efecto esperado, con dirección si la hay. -->
+
+## 3. Variables
+
+- **Independiente(s):** <!-- lo que se manipula. -->
+- **Dependiente(s):** <!-- lo que se mide (la métrica de resultado). -->
+- **Control:** <!-- lo que se mantiene fijo / se cubre estadísticamente. -->
+
+## 4. Diseño
+
+<!--
+- N: tamaño de muestra (y justificación / power analysis si aplica).
+- Condiciones / grupos.
+- Muestreo y aleatorización.
+- Criterios de inclusión / exclusión de datos (definidos AHORA, no después).
+-->
+
+## 5. Plan de análisis
+
+<!--
+El plan estadístico EXACTO, decidido antes de ver los datos:
+- Test estadístico concreto (p.ej. t-test de Welch, Mann-Whitney U, regresión...).
+- Métrica de effect size (p.ej. Cohen's d, diferencia de medias, odds ratio).
+- Criterio de decisión (umbral alpha, qué resultado confirma/refuta H1).
+- Corrección por comparaciones múltiples (p.ej. Holm-Bonferroni) si hay >1 contraste.
+- Manejo de supuestos (normalidad, varianzas) y qué se hace si no se cumplen.
+-->
+
+## 6. Predicción cuantitativa
+
+<!--
+La predicción numérica concreta que el experimento pondrá a prueba.
+P.ej. "esperamos d >= 0.5 con IC95% que no cruza 0" o "una reducción >= 15% en la métrica X".
+Cuanto más específica, más falsable.
+-->

python/functions/datascience/__init__.py

@@ -59,6 +59,9 @@ from .acf_pacf import acf_pacf
 from .stl_decompose import stl_decompose
 from .to_returns import to_returns
 from .fdr_correction import fdr_correction
+from .effect_size_cohens_d import effect_size_cohens_d
+from .confidence_interval_mean import confidence_interval_mean
+from .preregister_hypothesis import preregister_hypothesis
 from .suggest_reexpression import suggest_reexpression
 from .exploratory_caveats import exploratory_caveats
 from .render_eda_pdf import render_eda_pdf, render_eda_pdf_relational
@@ -73,9 +76,15 @@ from .resample_timeseries import resample_timeseries
 from .add_pdf_internal_links import add_pdf_internal_links
 from .suggest_intratable_fk_candidates import suggest_intratable_fk_candidates
 from .render_paper_pdf import render_paper_pdf
+from .draw_join_graph_figure import draw_join_graph_figure
+from .generate_synthetic_eda_table import generate_synthetic_eda_table
+from .generate_synthetic_eda_folder import generate_synthetic_eda_folder
 
 __all__ = [
+    "generate_synthetic_eda_table",
+    "generate_synthetic_eda_folder",
     "render_paper_pdf",
+    "draw_join_graph_figure",
     "suggest_intratable_fk_candidates",
     "detect_time_column",
     "extract_timeseries_raw",
@@ -92,6 +101,9 @@ __all__ = [
     "stl_decompose",
     "to_returns",
     "fdr_correction",
+    "effect_size_cohens_d",
+    "confidence_interval_mean",
+    "preregister_hypothesis",
     "suggest_reexpression",
     "exploratory_caveats",
     "render_eda_pdf",

python/functions/datascience/automatic_eda/__init__.py

@@ -29,6 +29,7 @@ from .model import (  # noqa: F401
     KVTable,
     Markdown,
     Note,
+    TocEntry,
     as_blocks,
     as_chapters,
     merge_manifest,
@@ -52,6 +53,7 @@ __all__ = [
     "Group",
     "GlossaryEntry",
     "GlossaryCollector",
+    "TocEntry",
     "Chapter",
     "as_blocks",
     "as_chapters",

python/functions/datascience/automatic_eda/build_document_only_test.py

@@ -0,0 +1,109 @@
+"""Tests del filtro `only` de build_document (selección de capítulos).
+
+Verifican que:
+  - only=None mantiene el comportamiento histórico (todos los capítulos).
+  - only=[ids] restringe el CUERPO a esos ids, pero portada (primera) y glosario
+    (última) están SIEMPRE presentes.
+  - only=[] produce el documento mínimo (solo portada + glosario).
+  - la selección también viaja por la clave reservada ctx['_only_chapters']
+    (el canal que usan los renderers, que llaman build_document sin `only`), y
+    esa clave nunca se filtra a los capítulos.
+"""
+
+import os
+import sys
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_FUNCTIONS = os.path.abspath(os.path.join(_HERE, "..", "..", ".."))  # python/functions
+if _FUNCTIONS not in sys.path:
+    sys.path.insert(0, _FUNCTIONS)
+
+from datascience.automatic_eda import build_document  # noqa: E402
+
+
+def _profile_with_cat_and_num():
+    """Perfil mínimo que hace construir cat_distr y num_distr (cuerpo no vacío)."""
+    return {
+        "table": "ventas", "n_rows": 120, "n_cols": 2, "quality_score": 91,
+        "duplicate_pct": 1.5, "null_cell_pct": 0.8,
+        "columns": [
+            {"name": "region", "inferred_type": "categorical",
+             "categorical": {
+                 "top": [{"value": "norte", "count": 50, "pct": 0.42},
+                         {"value": "sur", "count": 40, "pct": 0.33},
+                         {"value": "este", "count": 30, "pct": 0.25}],
+                 "mode": "norte", "n_distinct": 3, "entropy": 1.55,
+                 "imbalance": 0.1}},
+            {"name": "importe", "inferred_type": "numeric",
+             "numeric": {"mean": 50.0, "median": 48.0, "std": 10.0,
+                         "min": 10, "max": 99, "iqr": 15,
+                         "histogram": [{"lo": 0, "hi": 50, "count": 40},
+                                       {"lo": 50, "hi": 100, "count": 80}]}},
+        ],
+    }
+
+
+def test_only_none_is_full_document():
+    """Retro-compat: sin `only`, salen todos los capítulos aplicables."""
+    chs = build_document(_profile_with_cat_and_num(), ctx={"dataset_name": "v"})
+    ids = [c.id for c in chs]
+    assert ids[0] == "portada"
+    assert ids[-1] == "glosario"
+    # El cuerpo trae las distribuciones (cat/num), no solo portada+glosario.
+    assert "num_distr" in ids
+    assert "cat_distr" in ids
+
+
+def test_only_restricts_body_but_keeps_cover_and_glossary():
+    # cat_distr registra el término "entropía" en el glosario, así que el
+    # glosario (destino del término clicable) aparece — demuestra el contrato
+    # "portada primera + capítulo + glosario última".
+    chs = build_document(_profile_with_cat_and_num(),
+                         ctx={"dataset_name": "v"}, only=["cat_distr"])
+    ids = [c.id for c in chs]
+    assert ids[0] == "portada", f"portada no es la primera: {ids}"
+    assert ids[-1] == "glosario", f"glosario no es la última: {ids}"
+    assert "cat_distr" in ids
+    # num_distr quedó fuera de la selección.
+    assert "num_distr" not in ids
+
+
+def test_only_empty_yields_minimal_document():
+    # only=[] -> cuerpo vacío. La portada está siempre; el glosario solo aparece
+    # si algún capítulo registró términos (patrón preexistente: glosario vacío se
+    # omite). Sin cuerpo no hay términos → documento mínimo = solo portada.
+    chs = build_document(_profile_with_cat_and_num(),
+                         ctx={"dataset_name": "v"}, only=[])
+    ids = [c.id for c in chs]
+    assert ids == ["portada"], \
+        f"only=[] debe dar el documento mínimo (solo portada), no {ids}"
+
+
+def test_selection_via_reserved_ctx_key():
+    """La selección viaja por ctx['_only_chapters'] cuando no se pasa `only`."""
+    chs = build_document(_profile_with_cat_and_num(),
+                         ctx={"dataset_name": "v",
+                              "_only_chapters": ["cat_distr"]})
+    ids = [c.id for c in chs]
+    assert "cat_distr" in ids
+    assert "num_distr" not in ids
+    assert ids[0] == "portada" and ids[-1] == "glosario"
+
+
+def test_explicit_only_arg_wins_over_ctx_key():
+    """Si se pasan ambos, el argumento `only` manda sobre la clave del ctx."""
+    chs = build_document(_profile_with_cat_and_num(),
+                         ctx={"dataset_name": "v",
+                              "_only_chapters": ["cat_distr"]},
+                         only=["num_distr"])
+    ids = [c.id for c in chs]
+    assert "num_distr" in ids
+    assert "cat_distr" not in ids
+
+
+def test_reserved_key_not_leaked_to_caller_ctx():
+    """build_document no muta el ctx del caller (copia interna)."""
+    ctx = {"dataset_name": "v", "_only_chapters": ["num_distr"]}
+    build_document(_profile_with_cat_and_num(), ctx=ctx)
+    # La clave reservada sigue en el dict del caller (no se mutó su copia).
+    assert ctx["_only_chapters"] == ["num_distr"]

python/functions/datascience/automatic_eda/chapter_deps.py

@@ -0,0 +1,205 @@
+"""chapter_deps — mapa central de dependencias de cómputo por capítulo del EDA.
+
+Fuente de verdad ÚNICA de qué necesita cada capítulo de ``CHAPTER_ORDER`` para
+computarse COMPLETO (sin caer en su rama degradada "datos insuficientes"). Lo
+consume el pipeline ``render_automatic_eda`` cuando se le pide renderizar un
+SUBCONJUNTO de capítulos (kwarg ``only_chapters``): antes de perfilar, resuelve
+los requisitos de los capítulos pedidos y activa SOLO el cómputo que esos
+capítulos necesitan, de modo que un capítulo suelto siempre llegue poblado y a la
+vez no se malgaste CPU/LLM en piezas que ningún capítulo pedido usa.
+
+Diseño: el mapa es CENTRAL (este módulo), NO una constante por capítulo. Así se
+evita tocar los ``chapters/<id>.py`` (cada agente es dueño de su capítulo) y se
+elimina el riesgo de colisión entre ramas. Si un capítulo cambia lo que lee del
+``profile``/``ctx``, se actualiza ESTE mapa — es donde el motor mira.
+
+Dos clases de dependencia, derivadas inspeccionando qué lee cada capítulo:
+
+  - ``profile_flags``: flags de coste de ``profile_table`` que hay que ACTIVAR
+    para que el ``profile`` traiga el bloque que el capítulo lee. Son los caros:
+      * ``run_models``  -> ``profile['models']`` (KMeans/IsolationForest/PCA).
+        Lo leen ``outliers`` (fallback del multivariante) y ``modelos``.
+      * ``run_series``  -> ``profile['series']`` (análisis de serie temporal).
+        Lo lee ``timeseries``.
+      * ``run_llm``     -> ``profile['llm']`` (interpretación del modelo).
+        Lo lee ``analisis_llm``.
+
+  - ``ctx``: etiquetas de las piezas de DATOS CRUDOS que construye
+    ``build_eda_render_ctx`` y que el capítulo lee del ``ctx``. Si la lista está
+    vacía, el capítulo no necesita datos crudos y el pipeline puede saltarse
+    ``build_eda_render_ctx`` por completo cuando ningún capítulo pedido los pide.
+    Etiquetas y claves reales que mapean (ver ``CTX_LABEL_TO_KEYS``):
+      * ``head_rows``      -> ``ctx['head_rows']``      (overview: df.head real).
+      * ``raw_numeric``    -> ``ctx['raw_numeric']``    (outliers/modelos/
+        correlacion/missingness/geospatial: muestra numérica alineada por fila).
+      * ``timeseries_raw`` -> ``ctx['timeseries_raw']`` (timeseries: serie cruda).
+      * ``geo_points``     -> ``ctx['geo_points']`` (+ ``raw_numeric``)
+        (geospatial: lat/lon).
+      * ``db_path_table``  -> ``ctx['db_path']`` + ``ctx['table']`` (agregacion/
+        text_distr/missingness/relaciones: push-down de queries propias).
+
+``portada`` y ``glosario`` NO son opcionales: el pipeline los incluye SIEMPRE
+(la portada resume el documento y el glosario es el destino de los términos
+clicables), así que aquí se declaran sin requisitos de cómputo.
+
+Todas las funciones de este módulo son PURAS (no I/O, deterministas): se prestan
+a test unitario directo.
+"""
+
+from __future__ import annotations
+
+# Mapa central. Una entrada por id de CHAPTER_ORDER. ``profile_flags`` lista los
+# flags de coste a activar; ``ctx`` las etiquetas de datos crudos que lee. Las
+# claves vacías significan "no necesita ese tipo de dependencia".
+CHAPTER_DEPS = {
+    # Portada y glosario: SIEMPRE presentes, sin cómputo propio (la portada lee
+    # el document_summary que arma build_document; el glosario lee los términos
+    # que el resto registró). Se declaran para que el mapa cubra CHAPTER_ORDER
+    # entero y la validación los reconozca.
+    "portada":      {"profile_flags": [], "ctx": []},
+    "overview":     {"profile_flags": [], "ctx": ["head_rows"]},
+    "analisis_llm": {"profile_flags": ["run_llm"], "ctx": []},
+    "num_distr":    {"profile_flags": [], "ctx": []},
+    "cat_distr":    {"profile_flags": [], "ctx": []},
+    # text_distr empuja su propia query de texto (no usa raw_numeric); necesita
+    # db_path/table en el ctx para hacerlo.
+    "text_distr":   {"profile_flags": [], "ctx": ["db_path_table"]},
+    "calidad":      {"profile_flags": [], "ctx": []},
+    # missingness lee la muestra numérica cruda (co-ocurrencia de ausencias) y
+    # puede empujar una query de patrón de nulos con db_path/table.
+    "missingness":  {"profile_flags": [], "ctx": ["raw_numeric", "db_path_table"]},
+    # outliers corre IsolationForest EN VIVO sobre ctx['raw_numeric']; run_models
+    # asegura además el fallback profile['models']['outliers'] si el ctx faltara.
+    "outliers":     {"profile_flags": ["run_models"], "ctx": ["raw_numeric"]},
+    "correlacion":  {"profile_flags": [], "ctx": ["raw_numeric"]},
+    "relaciones":   {"profile_flags": [], "ctx": ["db_path_table"]},
+    "modelos":      {"profile_flags": ["run_models"], "ctx": ["raw_numeric"]},
+    "timeseries":   {"profile_flags": ["run_series"], "ctx": ["timeseries_raw"]},
+    "geospatial":   {"profile_flags": [], "ctx": ["geo_points", "raw_numeric"]},
+    "agregacion":   {"profile_flags": [], "ctx": ["db_path_table"]},
+    "glosario":     {"profile_flags": [], "ctx": []},
+}
+
+# Capítulos que el documento incluye SIEMPRE, independientemente de only_chapters.
+ALWAYS_PRESENT = ("portada", "glosario")
+
+# Flags de coste reconocidos (el orden no importa; se devuelven como set).
+KNOWN_PROFILE_FLAGS = ("run_models", "run_series", "run_llm")
+
+# Mapeo de cada etiqueta de ctx a las claves REALES que produce
+# build_eda_render_ctx. ``db_path_table`` es especial: db_path/table siempre se
+# ponen para un backend válido y son inofensivos, por eso no se podan nunca (no
+# aparecen en DATA_CTX_KEYS). El resto (head_rows/raw_numeric/timeseries_raw/
+# geo_points) son las piezas de datos podables.
+CTX_LABEL_TO_KEYS = {
+    "head_rows":      {"head_rows"},
+    "raw_numeric":    {"raw_numeric"},
+    "timeseries_raw": {"timeseries_raw"},
+    "geo_points":     {"geo_points", "raw_numeric"},
+    "db_path_table":  set(),  # db_path/table siempre presentes; nunca se podan.
+}
+
+# Claves de datos crudos del ctx que se pueden podar cuando ningún capítulo
+# pedido las necesita (las que cuestan muestreo). db_path/table NO entran aquí.
+DATA_CTX_KEYS = ("head_rows", "raw_numeric", "timeseries_raw", "geo_points")
+
+
+def _as_id_list(chapter_ids):
+    """Normaliza la entrada a una lista de ids string, defensiva. None -> []."""
+    if chapter_ids is None:
+        return []
+    if isinstance(chapter_ids, str):
+        return [chapter_ids]
+    return [c for c in chapter_ids if isinstance(c, str)]
+
+
+def validate_chapter_ids(chapter_ids, order):
+    """Separa los ids pedidos en válidos y desconocidos respecto a ``order``.
+
+    Args:
+        chapter_ids: lista (o str) de ids de capítulo pedidos.
+        order: lista canónica de ids válidos (CHAPTER_ORDER).
+
+    Returns:
+        dict ``{"valid": [...], "unknown": [...]}`` preservando el orden de
+        aparición de la entrada. Función pura.
+    """
+    valid_set = set(order or [])
+    valid, unknown = [], []
+    for cid in _as_id_list(chapter_ids):
+        (valid if cid in valid_set else unknown).append(cid)
+    return {"valid": valid, "unknown": unknown}
+
+
+def resolve_requirements(chapter_ids):
+    """Une los requisitos de cómputo de los capítulos pedidos.
+
+    Es el corazón de la resolución de dependencias: dado el subconjunto de
+    capítulos a renderizar, devuelve TODO lo que hay que activar/construir para
+    que esos capítulos lleguen COMPLETOS, y solo eso.
+
+    Los capítulos ``ALWAYS_PRESENT`` (portada/glosario) se añaden implícitamente
+    porque el pipeline siempre los incluye; como no tienen requisitos, no alteran
+    el resultado, pero se contemplan para que el conjunto sea coherente.
+
+    Args:
+        chapter_ids: lista (o str) de ids de capítulo. Ids desconocidos se
+            ignoran silenciosamente (la validación estricta es de quien llama).
+            None o lista vacía -> requisitos vacíos.
+
+    Returns:
+        dict ``{"profile_flags": set[str], "ctx_keys": set[str]}`` donde
+        ``ctx_keys`` son las ETIQUETAS de ctx (no las claves reales). Función
+        pura.
+    """
+    ids = set(_as_id_list(chapter_ids)) | set(ALWAYS_PRESENT)
+    profile_flags = set()
+    ctx_keys = set()
+    for cid in ids:
+        dep = CHAPTER_DEPS.get(cid)
+        if not isinstance(dep, dict):
+            continue
+        for f in dep.get("profile_flags", []) or []:
+            if f in KNOWN_PROFILE_FLAGS:
+                profile_flags.add(f)
+        for k in dep.get("ctx", []) or []:
+            ctx_keys.add(k)
+    return {"profile_flags": profile_flags, "ctx_keys": ctx_keys}
+
+
+def resolve_profile_flags(chapter_ids):
+    """Atajo: solo el set de profile_flags a activar para los capítulos pedidos.
+
+    Función pura. Devuelve un set ⊆ KNOWN_PROFILE_FLAGS.
+    """
+    return resolve_requirements(chapter_ids)["profile_flags"]
+
+
+def needs_render_ctx(chapter_ids):
+    """True si algún capítulo pedido necesita datos crudos del ctx.
+
+    Cuando es False, el pipeline puede saltarse ``build_eda_render_ctx`` entero
+    (ahorro real de CPU/I/O): los capítulos pedidos no leen ninguna pieza de
+    datos crudos. Función pura.
+    """
+    return bool(resolve_requirements(chapter_ids)["ctx_keys"])
+
+
+def resolve_ctx_data_keys(chapter_ids):
+    """Claves REALES de datos del ctx a CONSERVAR para los capítulos pedidos.
+
+    Traduce las etiquetas de ctx a las claves concretas que produce
+    ``build_eda_render_ctx`` (head_rows/raw_numeric/timeseries_raw/geo_points).
+    El pipeline poda del ctx las claves de datos que NO estén en este set, para
+    que un capítulo suelto no arrastre piezas de datos que no usa. db_path/table
+    nunca se podan (no aparecen aquí). Función pura.
+
+    Returns:
+        set[str] subconjunto de DATA_CTX_KEYS.
+    """
+    req = resolve_requirements(chapter_ids)
+    keep = set()
+    for label in req["ctx_keys"]:
+        keep |= CTX_LABEL_TO_KEYS.get(label, set())
+    # Solo claves de datos podables (db_path/table se gestionan aparte).
+    return {k for k in keep if k in DATA_CTX_KEYS}

python/functions/datascience/automatic_eda/chapter_deps_test.py

@@ -0,0 +1,160 @@
+"""Tests del mapa central de dependencias por capítulo (chapter_deps).
+
+Todas las funciones bajo prueba son PURAS (sin I/O): se ejercitan directamente
+sin DuckDB ni renderizado. Cubren la resolución de requisitos (golden + edges),
+la validación de ids y los helpers de eficiencia (qué cómputo se salta).
+"""
+
+import os
+import sys
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_FUNCTIONS = os.path.abspath(os.path.join(_HERE, "..", "..", ".."))  # python/functions
+if _FUNCTIONS not in sys.path:
+    sys.path.insert(0, _FUNCTIONS)
+
+from datascience.automatic_eda.chapter_deps import (  # noqa: E402
+    ALWAYS_PRESENT,
+    CHAPTER_DEPS,
+    DATA_CTX_KEYS,
+    needs_render_ctx,
+    resolve_ctx_data_keys,
+    resolve_profile_flags,
+    resolve_requirements,
+    validate_chapter_ids,
+)
+from datascience.automatic_eda.chapters_registry import CHAPTER_ORDER  # noqa: E402
+
+
+# --------------------------------------------------------------------------- #
+# El mapa cubre CHAPTER_ORDER entero (sin huecos ni claves de más).
+# --------------------------------------------------------------------------- #
+def test_chapter_deps_covers_every_chapter_in_order():
+    assert set(CHAPTER_DEPS) == set(CHAPTER_ORDER), (
+        "CHAPTER_DEPS debe declarar exactamente los ids de CHAPTER_ORDER")
+    # Cada entrada tiene la forma esperada.
+    for cid, dep in CHAPTER_DEPS.items():
+        assert isinstance(dep.get("profile_flags"), list), cid
+        assert isinstance(dep.get("ctx"), list), cid
+
+
+# --------------------------------------------------------------------------- #
+# resolve_requirements — golden: outliers exige run_models + raw_numeric.
+# --------------------------------------------------------------------------- #
+def test_resolve_outliers_requires_run_models_and_raw_numeric():
+    req = resolve_requirements(["outliers"])
+    assert "run_models" in req["profile_flags"]
+    assert "raw_numeric" in req["ctx_keys"]
+    assert "run_series" not in req["profile_flags"]
+    assert "run_llm" not in req["profile_flags"]
+
+
+def test_resolve_timeseries_requires_run_series():
+    req = resolve_requirements(["timeseries"])
+    assert req["profile_flags"] == {"run_series"}
+    assert "timeseries_raw" in req["ctx_keys"]
+
+
+def test_resolve_analisis_llm_requires_run_llm():
+    assert resolve_requirements(["analisis_llm"])["profile_flags"] == {"run_llm"}
+
+
+def test_resolve_union_of_several_chapters():
+    req = resolve_requirements(["outliers", "timeseries", "analisis_llm"])
+    assert req["profile_flags"] == {"run_models", "run_series", "run_llm"}
+
+
+# --------------------------------------------------------------------------- #
+# Eficiencia: capítulos que NO necesitan flags caros no los activan.
+# --------------------------------------------------------------------------- #
+def test_resolve_geospatial_needs_no_cost_flags():
+    """geospatial sale de geo_points/raw_numeric del ctx, NO de los modelos."""
+    req = resolve_requirements(["geospatial"])
+    assert req["profile_flags"] == set(), \
+        "geospatial no debe activar run_models/run_series/run_llm"
+    assert "geo_points" in req["ctx_keys"]
+
+
+def test_resolve_correlacion_needs_raw_numeric_but_no_models():
+    req = resolve_requirements(["correlacion"])
+    assert req["profile_flags"] == set()
+    assert "raw_numeric" in req["ctx_keys"]
+
+
+def test_always_present_chapters_add_no_requirements():
+    """portada y glosario están siempre, pero no arrastran cómputo."""
+    for cid in ALWAYS_PRESENT:
+        req = resolve_requirements([cid])
+        assert req["profile_flags"] == set()
+        assert req["ctx_keys"] == set()
+
+
+def test_resolve_profile_flags_shortcut():
+    assert resolve_profile_flags(["modelos"]) == {"run_models"}
+    assert resolve_profile_flags(["num_distr"]) == set()
+
+
+# --------------------------------------------------------------------------- #
+# needs_render_ctx — cuándo se puede saltar build_eda_render_ctx por completo.
+# --------------------------------------------------------------------------- #
+def test_needs_render_ctx_true_when_chapter_reads_raw_data():
+    assert needs_render_ctx(["outliers"]) is True
+    assert needs_render_ctx(["agregacion"]) is True  # db_path/table push-down
+    assert needs_render_ctx(["timeseries"]) is True
+
+
+def test_needs_render_ctx_false_for_purely_aggregated_chapters():
+    """num_distr / cat_distr / calidad solo leen el profile agregado."""
+    assert needs_render_ctx(["num_distr"]) is False
+    assert needs_render_ctx(["cat_distr", "calidad"]) is False
+
+
+# --------------------------------------------------------------------------- #
+# resolve_ctx_data_keys — poda: qué claves de DATOS conservar (db_path/table no).
+# --------------------------------------------------------------------------- #
+def test_resolve_ctx_data_keys_outliers_keeps_only_raw_numeric():
+    assert resolve_ctx_data_keys(["outliers"]) == {"raw_numeric"}
+
+
+def test_resolve_ctx_data_keys_geospatial_keeps_geo_and_numeric():
+    assert resolve_ctx_data_keys(["geospatial"]) == {"geo_points", "raw_numeric"}
+
+
+def test_resolve_ctx_data_keys_aggregation_keeps_nothing_prunable():
+    """agregacion usa db_path/table (siempre presentes), 0 claves podables."""
+    assert resolve_ctx_data_keys(["agregacion"]) == set()
+
+
+def test_resolve_ctx_data_keys_subset_of_data_keys():
+    keep = resolve_ctx_data_keys(["overview", "timeseries", "geospatial"])
+    assert keep <= set(DATA_CTX_KEYS)
+    assert {"head_rows", "timeseries_raw", "geo_points", "raw_numeric"} == keep
+
+
+# --------------------------------------------------------------------------- #
+# validate_chapter_ids — separa válidos de desconocidos preservando orden.
+# --------------------------------------------------------------------------- #
+def test_validate_separates_known_and_unknown():
+    out = validate_chapter_ids(["outliers", "nope", "timeseries", "ghost"],
+                               CHAPTER_ORDER)
+    assert out["valid"] == ["outliers", "timeseries"]
+    assert out["unknown"] == ["nope", "ghost"]
+
+
+def test_validate_all_known():
+    out = validate_chapter_ids(["portada", "glosario"], CHAPTER_ORDER)
+    assert out["unknown"] == []
+
+
+# --------------------------------------------------------------------------- #
+# Robustez: entradas raras nunca lanzan.
+# --------------------------------------------------------------------------- #
+def test_resolve_handles_none_and_empty():
+    assert resolve_requirements(None)["profile_flags"] == set()
+    assert resolve_requirements([])["profile_flags"] == set()
+    # ids desconocidos se ignoran silenciosamente en la resolución.
+    assert resolve_requirements(["no_existe"])["ctx_keys"] == set()
+
+
+def test_resolve_accepts_single_string():
+    assert resolve_requirements("outliers")["profile_flags"] == {"run_models"}

python/functions/datascience/automatic_eda/chapters/cat_distr.py

@@ -5,28 +5,32 @@ page (PDF) / slide (PPTX)**: every column is wrapped in a keep-together
 ``model.Group`` with ``page_break_before=True`` (except the first, which may share
 the intro's page), so its chart sits next to its tables and no column is split.
 
-A short intro names the clickable **[[term:entropia]]entropía[[/term]]** term —
+Per column the Group is laid out ``side_by_side`` (PPTX: cardinality table LEFT,
-the full definition lives in the GLOSARIO chapter, so it is NOT repeated inline
+chart RIGHT; PDF: stacked) and contains, in order:
-here (one click jumps to the glossary entry). The intro also carries the dataset
-row total used as a comparison baseline.
 
-Per column the Group contains, in order:
+1. The column name plus, when the LLM layer ran, its business **description** and
-
+   **unit** (read from ``profile['llm']['dictionary']``, matched by column name).
-1. A cardinality key/value table: distinct values, ``% distinct`` (distinct /
+2. A cardinality key/value table: distinct values, ``% distinct`` (distinct /
    total rows), total dataset rows, singleton values (frequency 1), entropy with
    its theoretical maximum and the normalized ratio, mode, imbalance and
    string-length stats.
-2. A short note flagging problematic cardinality (id-like ≈100% distinct, or a
+3. A short note flagging problematic cardinality (id-like ≈100% distinct, or a
    single dominating category).
-3. A ``top-k`` table (value / count / %).
+4. A ``top-k`` table (value / count / %).
-4. A **donut pie chart** of the most common categories (top-k + an "Otros"
+5. A **horizontal bar chart** of the most common categories (top-k + an "Otros"
    bucket), drawn lazily so the renderers scale it to fit entirely.
 
+A short intro names the clickable **[[term:entropia]]entropía[[/term]]** and
+**[[term:pagina_categorica]]page-layout[[/term]]** terms — their full
+definitions live in the GLOSARIO chapter, so they are NOT repeated inline here
+(one click jumps to the glossary entry). The intro also carries the dataset row
+total used as a comparison baseline.
+
 Data comes from the ``eda`` group: each ``columns[i]['categorical']`` is the
 output of ``summarize_categorical`` (``top[{value,count,pct}]``, ``mode``,
 ``n_distinct``, ``entropy``, ``imbalance``, ``len_min/mean/max``). The derived
-cardinality metrics and the pie figure are delegated to two registry functions
+cardinality metrics and the bar figure are delegated to two registry functions
-(``categorical_cardinality_block`` and ``categorical_top_pie_figure``); both are
+(``categorical_cardinality_block`` and ``categorical_top_bar_figure``); both are
 imported lazily and degrade to a minimal inline fallback so this chapter never
 raises even if they are unavailable.
 
@@ -39,10 +43,21 @@ import math
 
 from .. import model
 
-CHAPTER_VERSION = "1.2.0"
+CHAPTER_VERSION = "1.3.0"
 CHAPTER_ID = "cat_distr"
 CHAPTER_TITLE = "Distribuciones categóricas"
 
+# Key under which eda_llm_insights stores its interpretive block in the profile.
+LLM_KEY = "llm"
+
+# Second glossary term this chapter names: "how each categorical page is laid
+# out". The long paragraph that used to describe it inline in the intro now lives
+# in the GLOSARIO chapter (canonical definition in ``glosario._BASELINE_TERMS``);
+# the intro only names the clickable term, relocating the explanation, not losing
+# it. The chapter only needs to register key+label here.
+_TERM_PAGINA_KEY = "pagina_categorica"
+_TERM_PAGINA_LABEL = "Cómo se organiza cada página categórica"
+
 # Glossary term this chapter explains. Registered in the shared collector and
 # marked clickable on its first appearance (end-to-end glossary example —
 # mejora 6). Other chapters hook their own terms the same way (see the contract).
@@ -59,14 +74,14 @@ _TERM_ENTROPIA_DEF = (
 # Cap the number of categorical columns rendered to keep the document bounded;
 # the rest are summarized in a closing note (no silent truncation).
 MAX_COLS = 40
-# Rows shown in each top-k table and explicit slices in the pie. Kept moderate so
+# Rows shown in each top-k table and explicit bars in the chart. Kept moderate so
-# the whole column — cardinality table + top-k table + donut — fits on ONE
+# the whole column — cardinality table + top-k table + bar chart — fits on ONE
 # page/slide with the chart next to its tables; the table note still reports
 # "top N of M" so nothing is silently hidden. For id-like columns (≈100%
 # distinct) the top-k table is dropped entirely (it would be a list of unique
-# values — pure noise), which also frees the room the donut needs (see build).
+# values — pure noise), which also frees the room the chart needs (see build).
 TOP_TABLE_ROWS = 8
-PIE_TOP_K = 6
+CHART_TOP_K = 6
 # Truncate very long category labels in tables (the renderer also wraps). Kept
 # tight so a column with long id-like values (names, tickets) still fits its page.
 LABEL_MAX = 28
@@ -208,26 +223,74 @@ def _fallback_cardinality(cat: dict, n_rows) -> dict:
     }
 
 
-def _pie_make(top, n_distinct, title, n_rows):
+def _llm_index(profile: dict, ctx: dict) -> dict:
-    """Return a zero-arg callable that builds the donut figure lazily."""
+    """Map column name -> its LLM dictionary entry (description/unit/...).
+
+    Reads the ``llm.dictionary`` list that ``eda_llm_insights`` stored in the
+    profile (``profile['llm']``; falls back to ``ctx['llm']``). Returns an empty
+    dict when ``run_llm`` did not run, so the caller degrades cleanly. Fully
+    defensive: never raises on malformed input.
+    """
+    llm = profile.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        llm = ctx.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        return {}
+    entries = llm.get("dictionary")
+    if not isinstance(entries, (list, tuple)):
+        return {}
+    index: dict = {}
+    for e in entries:
+        if not isinstance(e, dict):
+            continue
+        col = e.get("column")
+        if col is None:
+            continue
+        index[model._safe_str(col)] = e
+    return index
+
+
+def _llm_desc_unit_block(name: str, llm_index: dict):
+    """Markdown block with the LLM business description + unit of a column, or
+    None when no LLM entry matches the column (clean fallback without LLM)."""
+    entry = llm_index.get(model._safe_str(name))
+    if not isinstance(entry, dict):
+        return None
+    raw_desc = entry.get("description") or entry.get("business_meaning")
+    desc = " ".join(model._safe_str(raw_desc).split()) if raw_desc else ""
+    raw_unit = entry.get("unit")
+    unit = " ".join(model._safe_str(raw_unit).split()) if raw_unit else ""
+    parts = []
+    if desc:
+        parts.append(f"**Descripción:** {desc}")
+    if unit:
+        parts.append(f"**Unidad:** {unit}")
+    if not parts:
+        return None
+    return model.Markdown(text=" · ".join(parts))
+
+
+def _bar_make(top, n_distinct, title, n_rows):
+    """Return a zero-arg callable that builds the bar figure lazily."""
 
     def make():
         try:
-            from datascience.categorical_top_pie_figure import (
+            from datascience.categorical_top_bar_figure import (
-                categorical_top_pie_figure,
+                categorical_top_bar_figure,
             )
 
-            return categorical_top_pie_figure(
+            return categorical_top_bar_figure(
                 top=top, n_distinct=n_distinct or 0, title=title,
-                top_k=PIE_TOP_K, n_rows=n_rows)
+                top_k=CHART_TOP_K, n_rows=n_rows)
         except Exception:  # noqa: BLE001 — minimal local fallback figure.
-            return _fallback_pie(top, title)
+            return _fallback_bar(top, title)
 
     return make
 
 
-def _fallback_pie(top, title):
+def _fallback_bar(top, title):
-    """Minimal donut figure used only if the registry function is unavailable."""
+    """Minimal horizontal-bar figure used only if the registry function is
+    unavailable. Largest category on top, the rest folded into "Otros"."""
     import matplotlib
 
     matplotlib.use("Agg")
@@ -238,8 +301,8 @@ def _fallback_pie(top, title):
     items = [t for t in (top or [])
              if isinstance(t, dict) and isinstance(t.get("count"), (int, float))]
     items = sorted(items, key=lambda t: t.get("count") or 0, reverse=True)
-    head = items[:PIE_TOP_K]
+    head = items[:CHART_TOP_K]
-    rest = items[PIE_TOP_K:]
+    rest = items[CHART_TOP_K:]
     labels = [_truncate(t.get("value"), 20) for t in head]
     sizes = [float(t.get("count") or 0) for t in head]
     if rest:
@@ -249,10 +312,13 @@ def _fallback_pie(top, title):
         ax.text(0.5, 0.5, "sin datos categóricos", ha="center", va="center")
         ax.axis("off")
         return fig
-    ax.pie(sizes, labels=None, wedgeprops={"width": 0.42},
+    # barh draws bottom-up, so reverse to put the largest category on top.
-           autopct=lambda p: f"{p:.0f}%" if p >= 4 else "")
+    y_pos = range(len(labels))
-    ax.legend(labels, loc="center left", bbox_to_anchor=(1.0, 0.5),
+    ax.barh(list(y_pos), list(reversed(sizes)), color="#4C72B0",
-              fontsize=7, frameon=False)
+            edgecolor="white")
+    ax.set_yticks(list(y_pos))
+    ax.set_yticklabels(list(reversed(labels)), fontsize=7)
+    ax.set_xlabel("conteo", fontsize=8)
     ax.set_title(_truncate(title, 40))
     fig.tight_layout()
     return fig
@@ -373,22 +439,17 @@ def _topk_table(cat: dict):
                            note=note)
 
 
-def _intro_blocks(n_rows, mark_term: bool = False):
+def _intro_blocks(mark_term: bool = False):
-    total = _fmt_int(n_rows)
+    # The full explanation of entropy AND of how each categorical page is laid out
-    # Mark the first appearance of the term as a clickable glossary jump when the
+    # lives in the GLOSARIO chapter; the chapter body keeps only the minimal
-    # term was registered (mark_term). The full definition of entropy lives in the
+    # clickable terms — no descriptive prose — to avoid duplicating the glossary.
-    # GLOSARIO chapter, so the intro only names the clickable term here instead of
+    # The dataset row total is not repeated here: each column's cardinality table
-    # repeating the long explanation (avoids the redundancy with the glossary).
+    # already carries "Total filas (dataset)".
     entropia = ("[[term:entropia]]entropía[[/term]]" if mark_term
                 else "entropía")
-    text = (
+    pagina = ("[[term:pagina_categorica]]cómo se organiza cada página[[/term]]"
-        f"Cada columna categórica ocupa su propia página: sus métricas de "
+              if mark_term else "cómo se organiza cada página")
-        f"cardinalidad —incluida la {entropia}—, una nota que señala cardinalidad "
+    text = f"Términos: {entropia} · {pagina}."
-        "problemática, la tabla de las categorías más frecuentes y un gráfico de "
-        "tarta (donut) de las más comunes, todo junto."
-    )
-    if n_rows is not None:
-        text += f" El dataset tiene {total} filas en total como referencia."
     return [
         model.Heading(text="Entropía y cardinalidad", level=2),
         model.Markdown(text=text),
@@ -406,15 +467,22 @@ def build_cat_distr(profile: dict, ctx: dict):
         return None
 
     n_rows = profile.get("n_rows")
-    # Register "entropía" in the shared glossary collector (if present) and mark
+    # Register "entropía" and the "how each categorical page is laid out" term in
-    # its first appearance clickable. End-to-end glossary example (mejora 6).
+    # the shared glossary collector (if present) and mark their first appearance
+    # clickable. End-to-end glossary example (mejora 6).
     glossary = ctx.get("glossary")
     mark_term = False
     if isinstance(glossary, model.GlossaryCollector):
         glossary.add(_TERM_ENTROPIA_KEY, _TERM_ENTROPIA_LABEL,
                      _TERM_ENTROPIA_DEF)
+        glossary.add(_TERM_PAGINA_KEY, _TERM_PAGINA_LABEL)
         mark_term = True
-    blocks = list(_intro_blocks(n_rows, mark_term=mark_term))
+    blocks = list(_intro_blocks(mark_term=mark_term))
+
+    # Business description + unit per column come from the LLM dictionary
+    # (profile['llm']['dictionary'], matched by column name); absent without
+    # run_llm, in which case the per-column description block is simply omitted.
+    llm_index = _llm_index(profile, ctx)
 
     rendered = cat_cols[:MAX_COLS]
     for idx, col in enumerate(rendered):
@@ -422,31 +490,36 @@ def build_cat_distr(profile: dict, ctx: dict):
         cat = col.get("categorical") or {}
         card = _normalize_card(_cardinality(cat, n_rows))
 
-        # One Group per categorical column: heading + cardinality table + flag
+        # One Group per categorical column: heading + (optional) LLM description +
-        # note + top-k table + donut figure are kept together and the renderer
+        # cardinality table + flag note + top-k table + bar figure are kept
-        # starts each on a fresh page/slide (page_break_before) so every column
+        # together and the renderer starts each on a fresh page/slide
-        # gets its own page with its chart next to its tables. The first column
+        # (page_break_before) so every column gets its own page with its chart next
-        # may share the intro's page (no forced break) to avoid a near-empty page.
+        # to its tables. The first column may share the intro's page (no forced
-        col_blocks = [
+        # break) to avoid a near-empty page.
-            model.Heading(text=str(name), level=2),
+        col_blocks = [model.Heading(text=str(name), level=2)]
-            _cardinality_block(card),
+        desc_block = _llm_desc_unit_block(name, llm_index)
-        ]
+        if desc_block is not None:
+            col_blocks.append(desc_block)
+        col_blocks.append(_cardinality_block(card))
         note = _flag_note(card)
         if note is not None:
             col_blocks.append(note)
         # For id-like columns (≈100% distinct) the top-k is a list of unique
         # values — pure noise; skip it (the flag note already explains why) and
-        # let the donut take that room so the whole column fits one page/slide.
+        # let the bar chart take that room so the whole column fits one page/slide.
         if not card.get("id_like"):
             topk = _topk_table(cat)
             if topk is not None:
                 col_blocks.append(topk)
         col_blocks.append(model.Figure(
-            make=_pie_make(cat.get("top") or [], card.get("n_distinct"),
+            make=_bar_make(cat.get("top") or [], card.get("n_distinct"),
                            str(name), n_rows),
             caption=(f"Categorías más comunes de «{_truncate(name, 32)}» "
-                     "(donut: top-k + «Otros»)")))
+                     "(barras: top-k + «Otros»)")))
-        blocks.append(model.Group(blocks=col_blocks,
+        # layout="side_by_side": in PPTX the cardinality table goes to the LEFT and
+        # the bar chart to the RIGHT of the same slide; the PDF renderer stacks it
+        # (the A5 mobile page is too narrow for two readable columns).
+        blocks.append(model.Group(blocks=col_blocks, layout="side_by_side",
                                   page_break_before=(idx > 0)))
 
     if len(cat_cols) > len(rendered):

python/functions/datascience/automatic_eda/chapters/cat_distr_test.py

@@ -2,12 +2,14 @@
 
 Self-contained: builds synthetic TableProfiles (no DuckDB) so the suite is fast
 and deterministic. Verifies that ``build_cat_distr`` emits the blocks the user
-asked for (distinct/total/%-distinct/unique metrics, top-k table and a donut
+asked for (distinct/total/%-distinct/unique metrics, top-k table and a bar
 figure), that EACH categorical column is wrapped in its own keep-together
-``Group`` that starts on a fresh page/slide (one column per page, chart next to
+``Group`` laid out ``side_by_side`` (PPTX: table left / bars right) that starts on
-its tables), that the long entropy explanation is NOT repeated inline (it lives
+a fresh page/slide (one column per page, chart next to its tables), that the LLM
-in the glossary — only the clickable term is kept), that the chapter renders
+business description + unit are shown per column when the profile carries an LLM
-inside the full document to both PDF and PPTX showing that content, that a
+block, that the long entropy / page-layout explanations are NOT repeated inline
+(they live in the glossary — only the clickable terms are kept), that the chapter
+renders inside the full document to both PDF and PPTX showing that content, that a
 profile with no categorical columns yields ``None`` without raising, and that
 long labels / many columns are never cut in either output.
 """
@@ -116,6 +118,10 @@ def test_golden_build_cat_distr_emite_bloques_pedidos():
     assert "log2" not in md.text          # redundant explanation removed.
     assert "máxima diversidad" not in md.text
 
+    # The donut/pie is gone: the intro no longer mentions tarta/donut (the chart
+    # is now a bar chart; the long page-layout explanation moved to the glossary).
+    assert "donut" not in md.text and "tarta" not in md.text
+
     # Per-column blocks are wrapped in keep-together Groups: flatten to inspect.
     flat = _flatten(ch.blocks)
     kv = next(b for b in flat if isinstance(b, KVTable))
@@ -128,11 +134,13 @@ def test_golden_build_cat_distr_emite_bloques_pedidos():
     assert any("Entropía" in lbl for lbl in labels)
     assert "únicos" in values and "%" in values
     assert "bits" in values and "norm" in values   # entropy + max + normalized.
-    # Top-k table + pie figure.
+    # Top-k table + bar figure.
     dt = next(b for b in flat if isinstance(b, DataTable))
     assert dt.header == ["Valor", "Conteo", "%"]
     assert any("neumaticos" in str(cell) for row in dt.rows for cell in row)
     assert any(isinstance(b, Figure) for b in flat)
+    # Each per-column Group is laid out side_by_side (table left / bars right).
+    assert all(g.layout == "side_by_side" for g in _column_groups(ch))
     # id-like column flagged with a Note that also explains the top-k is dropped.
     idnote = next((b for b in flat
                    if isinstance(b, Note) and "identificador" in b.text), None)
@@ -140,9 +148,9 @@ def test_golden_build_cat_distr_emite_bloques_pedidos():
     assert "No se lista el top" in idnote.text
 
 
-def test_golden_idlike_omite_topk_y_conserva_donut():
+def test_golden_idlike_omite_topk_y_conserva_grafico():
     # The id-like column (uuid, 100% distinct) must NOT carry a top-k DataTable
-    # (it would be a list of unique values), but must still keep its donut Figure
+    # (it would be a list of unique values), but must still keep its bar Figure
     # and its cardinality table so it stays a full per-column page.
     ch = build_cat_distr(_profile(), {})
     groups = _column_groups(ch)
@@ -151,7 +159,7 @@ def test_golden_idlike_omite_topk_y_conserva_donut():
     kinds = [b.kind for b in uuid_group.blocks]
     assert "data_table" not in kinds      # top-k of unique values dropped.
     assert "kv_table" in kinds            # cardinality kept.
-    assert "figure" in kinds              # donut kept (chart per column).
+    assert "figure" in kinds              # bar chart kept (chart per column).
     # A non-id-like column keeps its top-k table.
     cat_group = next(g for g in groups
                      if any(getattr(b, "text", "") == "categoria"
@@ -205,7 +213,7 @@ def test_golden_render_pdf_una_pagina_por_columna():
         assert "Entrop" in txt
         assert "distintos" in txt
         assert "categoria" in txt and "neumaticos" in txt
-        assert "donut" in txt           # figure caption rendered as text.
+        assert "barras" in txt          # bar-chart caption rendered as text (PDF).
         assert "identificador" in txt   # id-like note rendered.
 
 
@@ -258,9 +266,11 @@ def _profile_high_card() -> dict:
 
 
 def test_golden_pptx_una_slide_por_columna_con_su_grafico():
-    """Each categorical column occupies EXACTLY ONE cat_distr slide that carries
+    """Cada columna categórica ocupa EXACTAMENTE UN slide cat_distr que lleva su
-    BOTH its cardinality table and its donut figure (picture) — i.e. the chart is
+    gráfico (picture) en la misma slide — el chart nunca se separa de su columna,
-    never separated from its table, even for a high-cardinality column."""
+    ni siquiera para una columna de alta cardinalidad. Con layout side_by_side la
+    tabla se rasteriza a imagen, así que la comprobación se hace por presencia de
+    picture (no por el texto de la tabla)."""
     from pptx.enum.shapes import MSO_SHAPE_TYPE
 
     prof = _profile_high_card()
@@ -272,7 +282,7 @@ def test_golden_pptx_una_slide_por_columna_con_su_grafico():
         prs = Presentation(out)
 
         # Per column: the cat_distr slides whose text mentions it, and whether the
-        # owning slide also has the donut caption + an actual picture shape.
+        # owning slide also carries an actual picture shape (its chart).
         slides_with_col = {n: [] for n in cat_names}
         owner_has_chart = {n: False for n in cat_names}
         for i, sl in enumerate(prs.slides):
@@ -288,15 +298,106 @@ def test_golden_pptx_una_slide_por_columna_con_su_grafico():
             for n in cat_names:
                 if n in txt:
                     slides_with_col[n].append(i)
-                    has_table = "Cardinalidad" in txt or "distintos" in txt
+                    if has_pic:
-                    if has_pic and "donut" in txt and has_table:
                         owner_has_chart[n] = True
 
         for n in cat_names:
             # Exactly one slide carries the column (not split across slides).
             assert len(slides_with_col[n]) == 1, (n, slides_with_col[n])
-            # That single slide also holds its table AND its donut picture.
+            # That single slide also holds its chart picture.
-            assert owner_has_chart[n], (n, "tabla y donut no están en el mismo slide")
+            assert owner_has_chart[n], (n, "el gráfico no está en el slide de la columna")
+
+
+def test_golden_pptx_columna_side_by_side_tabla_izq_barra_der():
+    """Con layout side_by_side, una columna categórica coloca su tabla de
+    cardinalidad (imagen) en la mitad izquierda y su gráfico de barras (imagen) en
+    la mitad derecha de la MISMA slide. Verifica que al menos una columna queda en
+    dos columnas (tabla-izq / barras-der), evidencia del side_by_side en PPTX."""
+    from pptx.enum.shapes import MSO_SHAPE_TYPE
+    from pptx.util import Inches
+
+    with tempfile.TemporaryDirectory() as d:
+        out = os.path.join(d, "eda.pptx")
+        render_automatic_eda_pptx(_profile(), out, {"title": "EDA"})
+        prs = Presentation(out)
+        centre = int(Inches(13.333 / 2.0))   # half of the 16:9 slide width.
+        two_col_slides = 0
+        for sl in prs.slides:
+            texts, lefts = [], []
+            for sh in sl.shapes:
+                if sh.has_text_frame:
+                    texts.append(sh.text_frame.text)
+                if (sh.shape_type == MSO_SHAPE_TYPE.PICTURE
+                        and sh.left is not None):
+                    lefts.append(sh.left)
+            txt = re.sub(r"\s+", " ", " ".join(texts))
+            if "Distribuciones categ" not in txt:
+                continue
+            # One picture starts in the left half, another in the right half.
+            if len(lefts) >= 2 and min(lefts) < centre and max(lefts) > centre:
+                two_col_slides += 1
+        assert two_col_slides >= 1, (
+            "ninguna columna quedó con tabla-izq / barras-der (side_by_side)")
+
+
+def _profile_with_llm() -> dict:
+    """The base profile plus an ``llm`` block (as eda_llm_insights would store it
+    with run_llm=True): a data dictionary with description/unit per column."""
+    prof = _profile()
+    prof["llm"] = {
+        "dictionary": [
+            {"column": "categoria",
+             "description": "Familia de producto del recambio",
+             "business_meaning": "Agrupa el catálogo por tipo de pieza",
+             "unit": "categoría"},
+            {"column": "uuid",
+             "description": "Identificador único de registro",
+             "unit": ""},
+        ],
+    }
+    return prof
+
+
+def test_llm_descripcion_y_unidad_por_columna():
+    # With an LLM dictionary, each categorical column whose name matches shows its
+    # business description and unit in a per-column markdown block.
+    ch = build_cat_distr(_profile_with_llm(), {})
+    groups = _column_groups(ch)
+    cat_group = next(g for g in groups
+                     if any(getattr(b, "text", "") == "categoria"
+                            for b in g.blocks))
+    md = " ".join(b.text for b in cat_group.blocks
+                  if getattr(b, "kind", "") == "markdown")
+    assert "Descripción" in md and "Familia de producto" in md
+    assert "Unidad" in md and "categoría" in md
+
+
+def test_edge_sin_llm_no_anade_descripcion():
+    # Without an LLM block the per-column description markdown is simply omitted;
+    # the column still renders its cardinality table and bar figure.
+    ch = build_cat_distr(_profile(), {})
+    for g in _column_groups(ch):
+        mds = [b.text for b in g.blocks if getattr(b, "kind", "") == "markdown"]
+        assert not any("Descripción" in t for t in mds)
+
+
+def test_pagina_categorica_clicable_y_definicion_en_glosario():
+    # The "how each categorical page is laid out" term is registered + marked
+    # clickable in the intro, and its full definition lands in the glossary
+    # chapter (canonical baseline catalog), not inline.
+    from datascience.automatic_eda.chapters.glosario import build_glosario
+
+    gc = GlossaryCollector()
+    ch = build_cat_distr(_profile(), {"glossary": gc})
+    md = next(b for b in ch.blocks if isinstance(b, Markdown))
+    assert "[[term:pagina_categorica]]" in md.text
+    assert gc.has("pagina_categorica")
+    glos = build_glosario(_profile(), {"glossary": gc})
+    entry = next(b for b in glos.blocks
+                 if getattr(b, "kind", "") == "glossary_entry"
+                 and b.key == "pagina_categorica")
+    assert "barras" in entry.definition
+    assert "identificador" in entry.definition
 
 
 def test_edge_sin_categoricas_devuelve_none():

python/functions/datascience/automatic_eda/chapters/correlacion.py

@@ -31,7 +31,7 @@ import math
 
 from .. import model
 
-CHAPTER_VERSION = "1.0.0"
+CHAPTER_VERSION = "1.1.0"
 CHAPTER_ID = "correlacion"
 CHAPTER_TITLE = "Correlación"
 
@@ -47,6 +47,13 @@ _MAX_MATRIX_LABELS = 16
 # How many pairs to show in each of the top-positive / top-negative tables.
 _TOP_N = 10
 
+# How many of the strongest numeric-numeric pairs to draw as scatter plots on
+# each sign (positive / negative). A scatter per pair carries a fitted line/curve
+# and a relationship-type label; keeping the count small keeps the chapter
+# readable on a phone / a slide. Only signed (Pearson/Spearman) pairs qualify —
+# Cramér's V / correlation ratio pairs are not numeric-numeric, so no scatter.
+_SCATTER_TOP_N = 3
+
 # Glossary terms this chapter explains. Each is registered in the shared
 # collector (ctx['glossary']) and marked clickable on its first appearance in the
 # body — the canonical two-step pattern (see ``cat_distr`` for the reference
@@ -314,6 +321,139 @@ def _fdr_text(corr: dict, mark_term: bool = False) -> str | None:
     return " ".join(parts)
 
 
+def _is_seq(values) -> bool:
+    """True for a non-empty list/tuple of values (a raw numeric column)."""
+    return isinstance(values, (list, tuple)) and len(values) > 0
+
+
+def _select_scatter_pairs(pairs: list, top_n: int = _SCATTER_TOP_N):
+    """Pick the strongest numeric-numeric pairs to draw as scatters.
+
+    Only signed (Pearson/Spearman) pairs are numeric-numeric and thus eligible
+    for a scatter with a fitted curve. Returns up to ``top_n`` of the strongest
+    positive pairs followed by up to ``top_n`` of the strongest negative ones,
+    each ranked by magnitude. Mixed-type metrics (Cramér's V, correlation ratio,
+    mutual information) are excluded — they have no x/y scatter interpretation.
+    """
+    positive = []
+    negative = []
+    for pair in pairs:
+        if not isinstance(pair, dict) or not _is_signed(pair):
+            continue
+        value = pair.get("value")
+        if not _is_num(value):
+            continue
+        if value > 0:
+            positive.append(pair)
+        elif value < 0:
+            negative.append(pair)
+    positive.sort(key=lambda p: abs(float(p.get("value", 0.0))), reverse=True)
+    negative.sort(key=lambda p: abs(float(p.get("value", 0.0))), reverse=True)
+    return positive[:top_n] + negative[:top_n]
+
+
+def _classification_note(a: str, b: str, cls: dict) -> str:
+    """Human-readable sentence describing the relationship of a pair.
+
+    Plain text (not baked into the figure image) so the type label is selectable
+    in the PDF / extractable by pdftotext, and sits right next to its scatter
+    inside the keep-together Group.
+    """
+    tipo = model._safe_str(cls.get("tipo")) or "sin forma clara"
+    bits = []
+    pearson = cls.get("pearson")
+    spearman = cls.get("spearman")
+    r2_lin = cls.get("r2_linear")
+    r2_poly = None
+    for key in ("r2_poly2", "r2_poly3"):
+        v = cls.get(key)
+        if _is_num(v) and (r2_poly is None or float(v) > r2_poly):
+            r2_poly = float(v)
+    if _is_num(pearson):
+        bits.append(f"Pearson r={float(pearson):+.2f}")
+    if _is_num(spearman):
+        bits.append(f"Spearman ρ={float(spearman):+.2f}")
+    if _is_num(r2_lin):
+        bits.append(f"R² lineal={float(r2_lin):.2f}")
+    if r2_poly is not None:
+        bits.append(f"R² polinómico={r2_poly:.2f}")
+    metrics = "; ".join(bits)
+    text = (f"Relación **{tipo}** entre «{a}» y «{b}»."
+            + (f" {metrics}." if metrics else ""))
+    return text
+
+
+def _scatter_blocks(pairs: list, raw_numeric):
+    """Build keep-together scatter Groups for the strongest num-num pairs.
+
+    Returns a list of blocks (a Heading plus one Group per pair), or an empty
+    list when there is no raw numeric data (e.g. the lite profile drops
+    ``ctx['raw_numeric']`` to skip live recomputation) or the relationship
+    helpers are unavailable. Never raises: any failure degrades to no scatters,
+    leaving the matrix + tables intact.
+    """
+    if not isinstance(raw_numeric, dict) or not raw_numeric:
+        return []
+    selected = _select_scatter_pairs(pairs)
+    if not selected:
+        return []
+
+    # The relationship helpers live in the datascience package. Import lazily so
+    # the chapter still builds (matrix + tables) when they are absent.
+    try:
+        from datascience.classify_relationship_type import (
+            classify_relationship_type,
+        )
+        from datascience.relationship_scatter_figure import (
+            relationship_scatter_figure,
+        )
+    except Exception:  # noqa: BLE001 — degrade, never break the chapter.
+        return []
+
+    groups = []
+    for pair in selected:
+        a = pair.get("a")
+        b = pair.get("b")
+        xs = raw_numeric.get(a)
+        ys = raw_numeric.get(b)
+        # Edge: a selected pair has no raw column (aggregated profile, renamed
+        # column, …) — skip just that pair, keep the rest.
+        if not _is_seq(xs) or not _is_seq(ys):
+            continue
+        try:
+            cls = classify_relationship_type(list(xs), list(ys)) or {}
+        except Exception:  # noqa: BLE001
+            continue
+        a_lbl = model._safe_str(a)
+        b_lbl = model._safe_str(b)
+
+        def _make(xs=xs, ys=ys, a_lbl=a_lbl, b_lbl=b_lbl, cls=cls):
+            return relationship_scatter_figure(
+                list(xs), list(ys), x_label=a_lbl, y_label=b_lbl,
+                classification=cls)
+
+        groups.append(model.Group(blocks=[
+            model.Heading(text=f"{a_lbl} ↔ {b_lbl}", level=2),
+            model.Figure(
+                make=_make,
+                caption=(f"Dispersión de «{a_lbl}» frente a «{b_lbl}» con la "
+                         "curva de ajuste del mejor modelo.")),
+            model.Markdown(text=_classification_note(a_lbl, b_lbl, cls)),
+        ]))
+
+    if not groups:
+        return []
+    intro = model.Markdown(text=(
+        "Para los pares numéricos más fuertes (positivos y negativos) se dibuja "
+        "la nube de puntos con su ajuste y se clasifica el **tipo de relación**: "
+        "**lineal** (una recta basta), **polinómica** (curva de grado 2/3 que "
+        "mejora claramente el ajuste lineal), **monótona no-lineal** (crece o "
+        "decrece siempre pero no en línea recta; Spearman ≫ Pearson) o "
+        "**débil/sin forma**."))
+    return [model.Heading(text="Relaciones más fuertes (scatter)", level=2),
+            intro] + groups
+
+
 def build_correlacion(profile: dict, ctx: dict):
     """Build the Correlation Chapter, or None if there are no pairs to show.
 
@@ -392,6 +532,18 @@ def build_correlacion(profile: dict, ctx: dict):
             "No se han hallado correlaciones negativas significativas entre "
             "columnas numéricas.")))
 
+    # 2.5) Scatter plots of the strongest numeric-numeric pairs, each with its
+    # fitted curve and a relationship-type label (lineal / polinómica / monótona
+    # / débil). Needs the raw numeric sample (ctx['raw_numeric'], row-aligned);
+    # when it is absent (aggregated/lite profile) the scatters are simply omitted
+    # and the matrix + tables above stand on their own.
+    raw_numeric = None
+    if isinstance(ctx, dict):
+        raw_numeric = ctx.get("raw_numeric") or profile.get("raw_numeric")
+    else:
+        raw_numeric = profile.get("raw_numeric")
+    blocks.extend(_scatter_blocks(pairs, raw_numeric))
+
     # 3) Spuriousness caveat for level-based correlations (Granger–Newbold).
     caveat = corr.get("levels_caveat")
     if isinstance(caveat, str) and caveat.strip():

python/functions/datascience/automatic_eda/chapters/correlacion_test.py

@@ -175,6 +175,105 @@ def test_anticorte_matriz_ancha_y_etiquetas_largas_no_se_cortan():
         assert "azufre" in _pdf_text(pdf)
 
 
+def _raw_numeric_for_profile(n: int = 80) -> dict:
+    """Row-aligned raw numeric sample matching the signed pairs of _profile().
+
+    Builds columns with a clear, deterministic shape so the relationship-type
+    classifier has something unambiguous to label:
+      - density vs alcohol: strong negative linear (the top-negative pair).
+      - alcohol vs quality: positive linear.
+      - ph, fixed_acidity, sulphates: filler columns for the remaining pairs.
+    """
+    import math as _m
+
+    alcohol = [8.0 + 0.05 * i for i in range(n)]
+    density = [1.0 - 0.002 * a for a in alcohol]           # neg linear vs alcohol
+    quality = [3.0 + 0.4 * a + (0.1 if i % 2 else -0.1)    # pos linear vs alcohol
+               for i, a in enumerate(alcohol)]
+    ph = [3.0 + 0.3 * _m.sin(i / 5.0) for i in range(n)]
+    fixed_acidity = [7.0 - 0.5 * p for p in ph]            # neg linear vs ph
+    sulphates = [0.5 + 0.01 * (i % 7) for i in range(n)]
+    return {
+        "alcohol": alcohol, "density": density, "quality": quality,
+        "ph": ph, "fixed_acidity": fixed_acidity, "sulphates": sulphates,
+    }
+
+
+def test_golden_scatters_de_pares_num_num_con_tipo_de_relacion():
+    """Con ctx['raw_numeric'], el capítulo añade scatters (Figure dentro de Group)
+    de los pares num-num más fuertes, cada uno con su etiqueta de tipo en texto."""
+    from datascience.automatic_eda.model import Group
+
+    ctx = {"raw_numeric": _raw_numeric_for_profile()}
+    ch = build_correlacion(_profile(), ctx)
+    assert ch is not None
+    groups = [b for b in ch.blocks if isinstance(b, Group)]
+    assert groups, "debe emitir al menos un Group con scatter"
+    # Cada Group lleva su figura (lazy) y una nota de texto con el tipo.
+    for g in groups:
+        gkinds = [b.kind for b in g.blocks]
+        assert "figure" in gkinds and "markdown" in gkinds
+    # La sección y la etiqueta de tipo aparecen como texto plano (extraíble).
+    headings = " ".join(b.text for b in ch.blocks if b.kind == "heading")
+    assert "Relaciones más fuertes" in headings
+    body = " ".join(b.text for g in groups for b in g.blocks
+                    if b.kind == "markdown")
+    assert any(t in body for t in
+               ("lineal", "polinómica", "monótona", "sin forma"))
+    # El par num-num más fuerte (density ↔ alcohol) tiene scatter; el par cat-cat
+    # (region ↔ type) NO — no es numérico.
+    assert "density" in body or "alcohol" in body
+    assert "region" not in body and "type" not in body
+
+
+def test_golden_pdf_muestra_scatters_con_etiqueta_de_tipo():
+    """En el PDF, el capítulo Correlación incluye los scatters y su etiqueta de
+    tipo en texto seleccionable (pdftotext la encuentra)."""
+    prof = _profile()
+    ctx = {"raw_numeric": _raw_numeric_for_profile()}
+    with tempfile.TemporaryDirectory() as d:
+        pdf = os.path.join(d, "corr_scatter.pdf")
+        rp = render_automatic_eda_pdf(prof, pdf, {"title": "EDA — wine",
+                                                  "ctx": ctx})
+        assert rp["path"] == pdf and rp["n_pages"] >= 1
+        txt = _pdf_text(pdf)
+        assert "Relaciones" in txt and "scatter" in txt.lower()
+        # Alguna etiqueta de tipo de relación, en texto.
+        assert any(t in txt for t in
+                   ("lineal", "polin", "monóton", "monoton", "sin forma"))
+
+
+def test_edge_sin_raw_numeric_omite_scatters_sin_lanzar():
+    """profile lite / ctx None: sin raw_numeric el capítulo omite los scatters
+    pero sigue emitiendo matriz + tablas (no lanza)."""
+    from datascience.automatic_eda.model import Group
+
+    for ctx in (None, {}, {"raw_numeric": None}, {"raw_numeric": {}}):
+        ch = build_correlacion(_profile(), ctx)
+        assert ch is not None
+        assert not [b for b in ch.blocks if isinstance(b, Group)]
+        # La matriz y al menos una tabla top siguen presentes.
+        assert any(b.kind == "figure" for b in ch.blocks)
+        assert any(b.kind == "data_table" for b in ch.blocks)
+
+
+def test_edge_par_sin_columna_cruda_se_omite_sin_lanzar():
+    """Si un par seleccionado no tiene su columna en raw_numeric, se omite ese
+    par (no lanza); los demás scatters se construyen igual."""
+    from datascience.automatic_eda.model import Group
+
+    raw = _raw_numeric_for_profile()
+    raw.pop("density", None)   # rompe el par density ↔ alcohol
+    ch = build_correlacion(_profile(), {"raw_numeric": raw})
+    assert ch is not None
+    groups = [b for b in ch.blocks if isinstance(b, Group)]
+    body = " ".join(b.text for g in groups for b in g.blocks
+                    if b.kind == "markdown")
+    # density desaparece de los scatters; otros pares (p.ej. ph↔fixed_acidity,
+    # alcohol↔quality) pueden seguir presentes sin error.
+    assert "density" not in body
+
+
 def test_glosario_engancha_metodos_y_fdr():
     """Mejora 4b: los métodos de correlación (Pearson, Spearman, Cramér's V,
     razón de correlación) y la corrección por comparaciones múltiples (FDR) se

python/functions/datascience/automatic_eda/chapters/glosario.py

@@ -17,10 +17,63 @@ from __future__ import annotations
 
 from .. import model
 
-CHAPTER_VERSION = "1.0.0"
+CHAPTER_VERSION = "1.1.0"
 CHAPTER_ID = "glosario"
 CHAPTER_TITLE = "Glosario"
 
+# Canonical definitions for cross-cutting terms — the "how to read it" entries
+# that do not belong to a single chapter. A chapter only needs to *register* the
+# term (``ctx['glossary'].add(key, label)``) and mark its in-text appearance with
+# ``[[term:key]]…[[/term]]``; this chapter supplies the full definition here when
+# the collector carries the term without one. Keeping the prose in a single place
+# avoids repeating a long paragraph inline in every chapter that names the term
+# (the explanation moved out of the NUM DISTR and CAT DISTR intros lives here).
+_BASELINE_TERMS = {
+    "histograma_boxplot": {
+        "label": "Cómo leer el histograma y el boxplot",
+        "definition": (
+            "Para cada columna numérica se muestra su histograma con tres líneas "
+            "de referencia: la media (línea roja discontinua), la mediana (línea "
+            "verde continua) y la banda ±1σ (zona sombreada que cubre una "
+            "desviación estándar a cada lado de la media). Debajo, alineado al "
+            "mismo eje horizontal, un boxplot de Tukey: la caja abarca del primer "
+            "al tercer cuartil (P25–P75), la línea interior es la mediana y los "
+            "bigotes llegan hasta 1,5·IQR; los puntos rojos señalan que hay "
+            "valores más allá de las vallas (posibles atípicos). Comparar la media "
+            "con la mediana revela la asimetría: si la media supera a la mediana la "
+            "cola larga cae hacia los valores altos (asimetría a la derecha), y al "
+            "revés hacia los bajos."),
+    },
+    "pagina_categorica": {
+        "label": "Cómo se organiza cada página categórica",
+        "definition": (
+            "Cada columna categórica ocupa su propia página: muestra sus métricas "
+            "de cardinalidad —incluida la entropía—, una nota que señala "
+            "cardinalidad problemática (columnas que se comportan como "
+            "identificador, con casi todos los valores distintos, o dominadas por "
+            "una sola categoría), la tabla de las categorías más frecuentes (top-k, "
+            "con su conteo y porcentaje) y un gráfico de barras de las categorías "
+            "más comunes (top-k más una barra «Otros» que agrupa la cola). El total "
+            "de filas del dataset se usa como referencia para interpretar los "
+            "conteos."),
+    },
+}
+
+
+def _resolve_term(term: dict) -> tuple:
+    """Return (label, definition) for a collected term, completing a missing
+    definition (and, if absent, the label) from the canonical baseline catalog."""
+    key = model._safe_str(term.get("key"))
+    label = model._safe_str(term.get("label"))
+    definition = model._safe_str(term.get("definition"))
+    base = _BASELINE_TERMS.get(key)
+    if base:
+        if not definition.strip():
+            definition = model._safe_str(base.get("definition"))
+        if not label.strip() or label == key:
+            label = model._safe_str(base.get("label")) or label
+    return label, definition
+
 
 def build_glosario(profile: dict, ctx: dict):
     """Build the glossary Chapter from the shared collector, or None if empty."""
@@ -36,12 +89,14 @@ def build_glosario(profile: dict, ctx: dict):
             "Cada término va resaltado en el texto y, al pulsarlo, salta a su "
             "definición en esta sección.")),
     ]
-    # One clickable destination per term, alphabetically by visible label.
+    # One clickable destination per term, alphabetically by visible label. A term
+    # registered without a definition is completed from the canonical baseline.
     for term in glossary.terms(by="label"):
+        label, definition = _resolve_term(term)
         blocks.append(model.GlossaryEntry(
             key=model._safe_str(term.get("key")),
-            label=model._safe_str(term.get("label")),
+            label=label,
-            definition=model._safe_str(term.get("definition"))))
+            definition=definition))
 
     return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
                          version=CHAPTER_VERSION, blocks=blocks)

python/functions/datascience/automatic_eda/chapters/missingness.py

@@ -0,0 +1,594 @@
+"""Missingness chapter (MISSINGNESS) — patterns of missing data.
+
+Complements the CALIDAD chapter: where CALIDAD reports *how much* is missing per
+column (the null percentage that lowers the completeness score), this chapter
+reports the **pattern** of the missing data — whether columns tend to be missing
+*together* (co-occurrence of absences) or independently. That distinction is what
+separates data that is missing completely at random ([[term:mcar]]MCAR[[/term]])
+from data missing as a function of another variable ([[term:mar]]MAR[[/term]]),
+which is the key question to settle before imputing or modelling.
+
+The chapter activates only when the table actually has missing data (at least one
+column with a null in the aggregated profile); otherwise it returns ``None`` and
+disappears from the document.
+
+Sections, in order:
+
+1. **Resumen global** — % of missing cells in the dataset, number of columns with
+   nulls, and complete rows (no missing) vs incomplete rows (≥1 missing).
+2. **Ranking por columna** — columns sorted by their null percentage, with a
+   horizontal bar figure.
+3. **Co-ocurrencia de ausencias** — the correlation of the binary is-null masks
+   between columns (which columns tend to be missing together): a heatmap plus a
+   table of the top column pairs that co-miss.
+4. **Patrones de fila** — the most frequent "which columns are missing together"
+   row patterns, in the style of missingno's pattern matrix.
+5. **Lectura MCAR/MAR** — an interpretive, *exploratory* note (not a confirmatory
+   test such as Little's) reading the absence correlations as a hint of MCAR
+   (independent absences) vs MAR (co-occurring absences).
+
+The aggregate per-column null counts come from the ``eda`` group ``TableProfile``
+(``columns[i]['null_count'] / 'null_pct'`` and the table-level ``null_cell_pct``).
+The per-row is-null mask needed for co-occurrence is built from raw data: a single
+DuckDB push-down over ``ctx['db_path'] / ctx['table']`` (same pattern as the
+AGREGACION chapter) covering ALL columns, with a fallback to the numeric-only
+``ctx['raw_numeric']`` when no database is reachable. All the heavy lifting is
+delegated to pure registry functions (``missingness_overview``,
+``missingness_correlation``, ``missingness_row_patterns``) and two figure helpers
+(``missingness_rank_bar_figure``, ``missingness_corr_heatmap_figure``); every one
+is imported lazily and degrades to an honest note so this chapter never raises.
+
+Contract: build_<id>(profile, ctx) -> Chapter | None ; CHAPTER_VERSION = "x.y.z".
+"""
+
+from __future__ import annotations
+
+from .. import model
+
+CHAPTER_VERSION = "1.0.0"
+CHAPTER_ID = "missingness"
+CHAPTER_TITLE = "Datos faltantes"
+
+# Sample cap for the per-row is-null mask push-down. Co-occurrence and row
+# patterns are computed on this sample; the global % of missing cells and the
+# per-column ranking come from the (exact) aggregated profile instead.
+MASK_SAMPLE = 5000
+# Thresholds for the MCAR/MAR heuristic note. A pair counts as a *strong*
+# co-occurrence when the absence correlation alone is high; as a *partial*
+# co-occurrence when the absences overlap materially (high Jaccard) even if the
+# Pearson correlation is modest — the usual case when one column is missing far
+# more often than the other (e.g. Cabin 77% vs Age 20% in Titanic), which dilutes
+# the correlation while the rows still co-miss in absolute terms.
+_CORR_STRONG = 0.30
+_JACCARD_NOTABLE = 0.20
+# Rows shown in the top-pairs and row-patterns tables (bounded, never silently
+# truncated: the table note reports the full count).
+_TOP_PAIRS = 12
+_TOP_PATTERNS = 12
+# Truncate long column names in tables (the renderer also wraps).
+_LABEL_MAX = 28
+
+# Glossary terms this chapter explains (contract §11.1). Registered in the shared
+# collector and marked clickable on their first appearance.
+_TERMS = {
+    "missingness": (
+        "Patrón de datos faltantes (missingness)",
+        "El patrón con el que faltan los datos: cuánto falta, en qué columnas y "
+        "si las ausencias de unas columnas coinciden (co-ocurren) con las de "
+        "otras. Analizarlo —no solo contar nulos— distingue datos que faltan al "
+        "azar (MCAR) de los que faltan en función de otra variable (MAR), lo que "
+        "decide cómo imputar o si descartar filas sin sesgar el análisis.",
+    ),
+    "mcar": (
+        "MCAR (Missing Completely At Random)",
+        "Los valores faltan de forma independiente de cualquier dato, observado o "
+        "no: las ausencias de unas columnas no se relacionan entre sí ni con los "
+        "valores. Es el caso más benigno —descartar filas o imputar la media no "
+        "introduce sesgo—, pero rara vez se cumple del todo en datos reales.",
+    ),
+    "mar": (
+        "MAR (Missing At Random)",
+        "La probabilidad de que un valor falte depende de OTRAS variables "
+        "observadas (p. ej. una medición que falta más en cierto grupo). Las "
+        "ausencias co-ocurren entre columnas o se relacionan con los valores de "
+        "otras; imputar exige condicionar en esas variables para no sesgar. La "
+        "co-ocurrencia fuerte de ausencias es un indicio (exploratorio) de MAR.",
+    ),
+}
+
+
+# --------------------------------------------------------------------------- #
+# Small defensive formatters (own copy: the chapter never imports siblings).
+# --------------------------------------------------------------------------- #
+def _fmt_int(value) -> str:
+    if value is None:
+        return "—"
+    try:
+        return f"{int(round(float(value))):,}".replace(",", ".")
+    except (TypeError, ValueError):
+        return model._safe_str(value)
+
+
+def _fmt_pct(value, decimals: int = 1) -> str:
+    """Format an already-0-100 value as a percentage. None -> placeholder."""
+    if value is None:
+        return "—"
+    try:
+        return f"{float(value):.{decimals}f}%"
+    except (TypeError, ValueError):
+        return model._safe_str(value)
+
+
+def _fmt_num(value, decimals: int = 3) -> str:
+    if value is None:
+        return "—"
+    try:
+        f = float(value)
+    except (TypeError, ValueError):
+        return model._safe_str(value)
+    if f != f:  # NaN
+        return "—"
+    text = f"{f:.{decimals}f}".rstrip("0").rstrip(".")
+    return text if text else "0"
+
+
+def _truncate(text, limit: int = _LABEL_MAX) -> str:
+    s = model._safe_str(text)
+    if len(s) <= limit:
+        return s
+    return s[: max(1, limit - 1)].rstrip() + "…"
+
+
+def _term(key: str, label: str, mark: bool) -> str:
+    if mark:
+        return f"[[term:{key}]]**{label}**[[/term]]"
+    return f"**{label}**"
+
+
+# --------------------------------------------------------------------------- #
+# Profile reads (exact, all rows).
+# --------------------------------------------------------------------------- #
+def _null_count_of(col: dict):
+    """Best-effort null count of a column: ``null_count`` or null_pct*n_rows."""
+    nc = col.get("null_count")
+    if isinstance(nc, (int, float)) and not isinstance(nc, bool):
+        return int(nc)
+    np_ = col.get("null_pct")
+    nr = col.get("n_rows")
+    if isinstance(np_, (int, float)) and isinstance(nr, (int, float)):
+        return int(round(float(np_) * float(nr)))
+    return 0
+
+
+def _columns_with_nulls(profile: dict):
+    """Return ``[(name, null_count, null_pct_0_100)]`` for columns with nulls,
+    sorted by null percentage descending. Reads the aggregated profile (exact)."""
+    cols = profile.get("columns") or []
+    out = []
+    for c in cols:
+        if not isinstance(c, dict):
+            continue
+        nc = _null_count_of(c)
+        if nc <= 0:
+            continue
+        np_ = c.get("null_pct")
+        nr = c.get("n_rows") or profile.get("n_rows")
+        if isinstance(np_, (int, float)) and not isinstance(np_, bool):
+            pct = float(np_) * 100.0 if np_ <= 1.0 else float(np_)
+        elif nr:
+            pct = nc / float(nr) * 100.0
+        else:
+            pct = None
+        out.append((c.get("name") or "(col)", nc, pct))
+    out.sort(key=lambda t: (t[2] if t[2] is not None else -1.0), reverse=True)
+    return out
+
+
+def _global_missing_pct(profile: dict):
+    """Table-level % of missing cells (0-100), exact, from the profile."""
+    v = profile.get("null_cell_pct")
+    if isinstance(v, (int, float)) and not isinstance(v, bool):
+        return float(v) * 100.0 if v <= 1.0 else float(v)
+    return None
+
+
+# --------------------------------------------------------------------------- #
+# Per-row is-null mask (sample): DuckDB push-down, fallback to raw_numeric.
+# --------------------------------------------------------------------------- #
+def _build_query_fn(ctx: dict):
+    """Return ``(query_fn, table)`` for a DuckDB-backed ctx, or ``(None, None)``.
+
+    Mirrors build_eda_render_ctx: a read-only closure over the registry wrapper.
+    Only DuckDB is supported here; any other backend degrades to raw_numeric."""
+    db_path = ctx.get("db_path")
+    table = ctx.get("table")
+    if not db_path or not table:
+        return None, None
+    try:
+        from infra import duckdb_query_readonly
+    except Exception:  # noqa: BLE001 — wrapper unavailable -> degrade.
+        return None, None
+
+    def query_fn(sql):
+        return duckdb_query_readonly(db_path, sql)
+
+    return query_fn, table
+
+
+def _null_mask(profile: dict, ctx: dict):
+    """Build the per-row is-null mask ``{col: [0/1, ...]}``.
+
+    Tries a single DuckDB push-down over ALL columns first (so categorical
+    columns like Cabin are covered, not only numeric ones); falls back to the
+    numeric-only ``ctx['raw_numeric']`` (None -> missing); returns ``(None, 0,
+    None)`` when neither is reachable. Never raises.
+    Returns ``(mask, n_sampled, source)`` with source in {"db","raw_numeric"}.
+    """
+    cols = profile.get("columns") or []
+    names = [c.get("name") for c in cols
+             if isinstance(c, dict) and c.get("name")]
+    # 1) DuckDB push-down over every column (covers categoricals too).
+    query_fn, table = _build_query_fn(ctx)
+    if query_fn is not None and names:
+        try:
+            from datascience.extract_null_mask import extract_null_mask
+
+            res = extract_null_mask(query_fn, table, names, max_rows=MASK_SAMPLE)
+            if isinstance(res, dict) and res.get("status") == "ok":
+                mask = res.get("mask") or {}
+                if mask:
+                    return mask, int(res.get("n") or 0), "db"
+        except Exception:  # noqa: BLE001 — degrade to raw_numeric.
+            pass
+    # 2) Fallback: numeric-only mask derived from raw_numeric (None -> missing).
+    rn = ctx.get("raw_numeric")
+    if isinstance(rn, dict) and rn:
+        mask = {}
+        for col, vals in rn.items():
+            if isinstance(vals, (list, tuple)):
+                mask[col] = [1 if v is None else 0 for v in vals]
+        if mask:
+            n = max((len(v) for v in mask.values()), default=0)
+            return mask, n, "raw_numeric"
+    return None, 0, None
+
+
+# --------------------------------------------------------------------------- #
+# Lazy registry delegations (each degrades to None on any failure).
+# --------------------------------------------------------------------------- #
+def _overview(mask: dict):
+    try:
+        from datascience.missingness_overview import missingness_overview
+
+        out = missingness_overview(mask)
+        return out if isinstance(out, dict) else None
+    except Exception:  # noqa: BLE001
+        return None
+
+
+def _correlation(mask: dict, top_k: int):
+    try:
+        from datascience.missingness_correlation import missingness_correlation
+
+        out = missingness_correlation(mask, top_k=top_k)
+        return out if isinstance(out, dict) else None
+    except Exception:  # noqa: BLE001
+        return None
+
+
+def _row_patterns(mask: dict, top_n: int):
+    try:
+        from datascience.missingness_row_patterns import missingness_row_patterns
+
+        out = missingness_row_patterns(mask, top_n=top_n)
+        return out if isinstance(out, dict) else None
+    except Exception:  # noqa: BLE001
+        return None
+
+
+def _rank_bar_make(names, pcts, title):
+    def make():
+        try:
+            from datascience.missingness_rank_bar_figure import (
+                missingness_rank_bar_figure,
+            )
+
+            return missingness_rank_bar_figure(names, pcts, title=title)
+        except Exception:  # noqa: BLE001 — minimal fallback figure.
+            return _fallback_fig("ranking de nulos no disponible")
+
+    return make
+
+
+def _heatmap_make(matrix, labels, title):
+    def make():
+        try:
+            from datascience.missingness_corr_heatmap_figure import (
+                missingness_corr_heatmap_figure,
+            )
+
+            return missingness_corr_heatmap_figure(matrix, labels, title=title)
+        except Exception:  # noqa: BLE001 — minimal fallback figure.
+            return _fallback_fig("heatmap de co-ocurrencia no disponible")
+
+    return make
+
+
+def _fallback_fig(message: str):
+    import matplotlib
+
+    matplotlib.use("Agg")
+    from matplotlib.figure import Figure
+
+    fig = Figure(figsize=(5.0, 2.2))
+    ax = fig.add_subplot(111)
+    ax.text(0.5, 0.5, message, ha="center", va="center")
+    ax.axis("off")
+    return fig
+
+
+# --------------------------------------------------------------------------- #
+# Block builders.
+# --------------------------------------------------------------------------- #
+def _summary_block(profile: dict, with_nulls: list, overview, sampled, n_total):
+    rows = []
+    gpct = _global_missing_pct(profile)
+    rows.append(("Celdas faltantes (global)", _fmt_pct(gpct)))
+    rows.append(("Columnas con faltantes", str(len(with_nulls))))
+    all_null = profile.get("all_null_cols")
+    if isinstance(all_null, (list, tuple)) and all_null:
+        rows.append(("Columnas 100% faltantes", str(len(all_null))))
+    if isinstance(overview, dict):
+        cr = overview.get("complete_rows")
+        ir = overview.get("incomplete_rows")
+        suffix = ""
+        if (isinstance(sampled, int) and isinstance(n_total, (int, float))
+                and sampled and n_total and sampled < n_total):
+            suffix = f" (sobre muestra de {_fmt_int(sampled)} filas)"
+        if cr is not None:
+            rows.append(("Filas completas (sin faltantes)",
+                         f"{_fmt_int(cr)} ({_fmt_pct(overview.get('complete_pct'))})"
+                         + suffix))
+        if ir is not None:
+            rows.append(("Filas con ≥1 faltante",
+                         f"{_fmt_int(ir)} "
+                         f"({_fmt_pct(overview.get('incomplete_pct'))})" + suffix))
+    return model.KVTable(rows=rows, title="Resumen de datos faltantes")
+
+
+def _ranking_block(with_nulls: list):
+    header = ["Columna", "Faltantes", "% faltante"]
+    rows = [[_truncate(n), _fmt_int(c), _fmt_pct(p)] for (n, c, p) in with_nulls]
+    if not rows:
+        return None
+    return model.DataTable(
+        header=header, rows=rows, title="Faltantes por columna",
+        note="ordenado de más a menos faltante")
+
+
+def _ranking_figure(with_nulls: list):
+    names = [n for (n, _, p) in with_nulls if p is not None]
+    pcts = [p for (_, _, p) in with_nulls if p is not None]
+    if not names:
+        return None
+    return model.Figure(
+        make=_rank_bar_make(names, pcts, "% de valores faltantes por columna"),
+        caption="Porcentaje de valores faltantes por columna (barras).")
+
+
+def _pairs_block(corr: dict):
+    """Top column pairs whose absences co-occur, as a table, or None."""
+    pairs = (corr or {}).get("pairs") or []
+    header = ["Columna A", "Columna B", "Corr. ausencia", "Co-faltan", "Jaccard"]
+    rows = []
+    for p in pairs[:_TOP_PAIRS]:
+        if not isinstance(p, dict):
+            continue
+        rows.append([
+            _truncate(p.get("a")),
+            _truncate(p.get("b")),
+            _fmt_num(p.get("corr")),
+            _fmt_int(p.get("co_missing")),
+            _fmt_num(p.get("jaccard")),
+        ])
+    if not rows:
+        return None
+    shown = len(rows)
+    total = len(pairs)
+    note = ("correlación de las máscaras is-null entre columnas; "
+            "«Co-faltan» = nº de filas en que ambas faltan a la vez")
+    if total > shown:
+        note += f" — top {shown} de {total} pares"
+    return model.DataTable(header=header, rows=rows,
+                           title="Pares de columnas que co-faltan", note=note)
+
+
+def _heatmap_block(corr: dict):
+    cols = (corr or {}).get("columns") or []
+    matrix = (corr or {}).get("matrix") or []
+    if len(cols) < 2 or not matrix:
+        return None
+    labels = [_truncate(c, 16) for c in cols]
+    return model.Figure(
+        make=_heatmap_make(matrix, labels, "Co-ocurrencia de ausencias"),
+        caption=("Correlación de las ausencias entre columnas (azul = faltan "
+                 "juntas; rojo = cuando una falta la otra tiende a estar)."))
+
+
+def _patterns_block(patterns_res: dict):
+    patterns = (patterns_res or {}).get("patterns") or []
+    header = ["Columnas que faltan juntas", "Filas", "%"]
+    rows = []
+    for p in patterns[:_TOP_PATTERNS]:
+        if not isinstance(p, dict):
+            continue
+        cols = p.get("missing_cols") or []
+        if cols:
+            label = ", ".join(_truncate(c, 18) for c in cols)
+        else:
+            label = "(fila completa — sin faltantes)"
+        rows.append([label, _fmt_int(p.get("n_rows")), _fmt_pct(p.get("pct"))])
+    if not rows:
+        return None
+    total = (patterns_res or {}).get("n_patterns")
+    shown = len(rows)
+    note = "cada fila es un patrón de «qué columnas faltan juntas»"
+    if isinstance(total, int) and total > shown:
+        note += f" — top {shown} de {total} patrones distintos"
+    return model.DataTable(header=header, rows=rows,
+                           title="Patrones de fila más comunes", note=note)
+
+
+def _mcar_mar_note(corr: dict, mark: bool):
+    """Interpretive, exploratory MCAR/MAR note from the absence correlations.
+
+    Reads the absence correlations at two levels so the verdict never contradicts
+    the visible evidence: a *strong* correlation flags a clear non-random (MAR)
+    pattern; a *partial* overlap (many rows co-miss — high Jaccard — even if the
+    correlation is diluted by one column being missing far more often) flags a
+    localized possible-MAR and cites the concrete co-missing pair; only when
+    neither holds does it read the absences as compatible with MCAR."""
+
+    def _pairs_with(attr_ok):
+        out = []
+        for p in (corr or {}).get("pairs") or []:
+            if isinstance(p, dict) and attr_ok(p):
+                out.append(p)
+        return out
+
+    def _cf(v):
+        try:
+            return float(v)
+        except (TypeError, ValueError):
+            return 0.0
+
+    strong = _pairs_with(lambda p: abs(_cf(p.get("corr"))) >= _CORR_STRONG)
+    partial = _pairs_with(
+        lambda p: _cf(p.get("corr")) > 0 and _cf(p.get("jaccard")) >= _JACCARD_NOTABLE)
+    mcar = _term("mcar", "MCAR", mark)
+    mar = _term("mar", "MAR", mark)
+    head = (
+        "**Lectura exploratoria MCAR/MAR.** Esta es una heurística basada en la "
+        "correlación de las ausencias entre columnas, NO un test confirmatorio "
+        "(como el de Little); orienta, no demuestra. ")
+    if strong:
+        top = strong[0]
+        ev = (f"«{model._safe_str(top.get('a'))}» y "
+              f"«{model._safe_str(top.get('b'))}» "
+              f"(corr {_fmt_num(top.get('corr'))})")
+        body = (
+            f"Hay ausencias que co-ocurren con fuerza —{ev}—: las columnas no "
+            f"faltan de forma independiente, lo que es un indicio de un patrón no "
+            f"aleatorio ({mar}). Antes de imputar o descartar filas conviene "
+            f"comprobar si la ausencia depende de otra variable observada; en ese "
+            f"caso la imputación debería condicionar en ella para no sesgar.")
+    elif partial:
+        top = max(partial, key=lambda p: _cf(p.get("jaccard")))
+        ev = (f"«{model._safe_str(top.get('a'))}» y "
+              f"«{model._safe_str(top.get('b'))}» faltan a la vez en "
+              f"{_fmt_int(top.get('co_missing'))} filas "
+              f"(Jaccard {_fmt_num(top.get('jaccard'))})")
+        body = (
+            f"Hay co-ocurrencia parcial de ausencias —{ev}—: algunas columnas "
+            f"tienden a faltar juntas aunque la correlación global sea modesta "
+            f"(habitual cuando una columna falta mucho más que la otra). Es un "
+            f"indicio de un posible patrón localizado no aleatorio ({mar}); "
+            f"conviene revisar si esa ausencia depende de otra variable observada "
+            f"antes de imputar, en lugar de asumir que faltan al azar.")
+    else:
+        body = (
+            f"Las ausencias entre columnas no muestran correlación ni solape "
+            f"relevante: parecen independientes, lo que es compatible con que "
+            f"falten al azar ({mcar}). Aun así, la ausencia podría depender de "
+            f"variables no observadas (la heurística no lo descarta).")
+    return model.Markdown(text=head + body)
+
+
+def _intro_block(mark: bool, source):
+    missingness = _term("missingness", "missingness", mark)
+    text = (
+        f"Este capítulo analiza el {missingness} de la tabla: no solo cuánto "
+        "falta (eso lo cubre la calidad), sino DÓNDE falta y si las columnas "
+        "faltan juntas. La co-ocurrencia de ausencias se calcula sobre la matriz "
+        "binaria «is-null» por fila.")
+    if source == "raw_numeric":
+        text += (" Nota: no se pudo leer la tabla cruda completa, así que la "
+                 "co-ocurrencia se limita a las columnas numéricas disponibles.")
+    return model.Markdown(text=text)
+
+
+# --------------------------------------------------------------------------- #
+# Entry point.
+# --------------------------------------------------------------------------- #
+def build_missingness(profile: dict, ctx: dict):
+    """Build the missingness Chapter, or None if the table has no missing data."""
+    if not isinstance(profile, dict):
+        profile = {}
+    ctx = ctx or {}
+
+    with_nulls = _columns_with_nulls(profile)
+    if not with_nulls:
+        return None  # no missing data anywhere -> chapter does not apply.
+
+    # Register glossary terms (if a collector is present) and mark them clickable.
+    glossary = ctx.get("glossary")
+    mark = False
+    if isinstance(glossary, model.GlossaryCollector):
+        for key, (label, definition) in _TERMS.items():
+            glossary.add(key, label, definition)
+        mark = True
+
+    # Per-row is-null mask (sample) for co-occurrence and row patterns.
+    mask, sampled, source = _null_mask(profile, ctx)
+    overview = _overview(mask) if mask else None
+    n_total = profile.get("n_rows")
+
+    blocks = [
+        model.Heading(text="Cuánto y dónde faltan datos", level=2),
+        _intro_block(mark, source),
+        _summary_block(profile, with_nulls, overview, sampled, n_total),
+        model.Heading(text="Faltantes por columna", level=2),
+    ]
+    ranking = _ranking_block(with_nulls)
+    if ranking is not None:
+        blocks.append(ranking)
+    rank_fig = _ranking_figure(with_nulls)
+    if rank_fig is not None:
+        blocks.append(rank_fig)
+
+    # Co-occurrence + row patterns need the per-row mask. Without it, say so.
+    if not mask:
+        blocks.append(model.Note(
+            "No se pudo construir la matriz «is-null» por fila (sin acceso a los "
+            "datos crudos), así que no se analiza la co-ocurrencia de ausencias "
+            "ni los patrones de fila en este informe."))
+        return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
+                             version=CHAPTER_VERSION, blocks=blocks)
+
+    corr = _correlation(mask, _TOP_PAIRS) or {}
+    co_blocks = [model.Heading(text="Co-ocurrencia de ausencias", level=2)]
+    heatmap = _heatmap_block(corr)
+    if heatmap is not None:
+        co_blocks.append(heatmap)
+    pairs = _pairs_block(corr)
+    if pairs is not None:
+        co_blocks.append(pairs)
+    if heatmap is None and pairs is None:
+        co_blocks.append(model.Note(
+            "Ninguna pareja de columnas comparte ausencias con variación "
+            "suficiente para correlacionarlas (p. ej. una sola columna con "
+            "faltantes), así que no hay co-ocurrencia que mostrar."))
+    # Keep the co-occurrence heading next to its heatmap and table.
+    blocks.append(model.Group(blocks=co_blocks))
+
+    patterns_res = _row_patterns(mask, _TOP_PATTERNS) or {}
+    patterns = _patterns_block(patterns_res)
+    if patterns is not None:
+        blocks.append(model.Heading(text="Patrones de fila", level=2))
+        blocks.append(patterns)
+
+    blocks.append(model.Heading(text="Lectura MCAR / MAR", level=2))
+    blocks.append(_mcar_mar_note(corr, mark))
+
+    return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
+                         version=CHAPTER_VERSION, blocks=blocks)

python/functions/datascience/automatic_eda/chapters/missingness_test.py

@@ -0,0 +1,162 @@
+"""Tests for the MISSINGNESS chapter.
+
+Covers the Definition of Done for this chapter:
+  * Activates (non-None Chapter with the expected sections) when the profile has
+    missing data, building the co-occurrence from the per-row is-null mask.
+  * Returns None when the table has no missing data at all (edge case).
+  * Registers the MCAR/MAR/missingness glossary terms.
+  * The DuckDB push-down path covers categorical columns (not only numeric),
+    so a categorical column that co-misses with a numeric one is detected.
+"""
+
+import os
+import sys
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_FUNCTIONS = os.path.abspath(os.path.join(_HERE, "..", "..", ".."))  # python/functions
+if _FUNCTIONS not in sys.path:
+    sys.path.insert(0, _FUNCTIONS)
+
+from datascience.automatic_eda import model  # noqa: E402
+from datascience.automatic_eda.chapters.missingness import (  # noqa: E402
+    build_missingness,
+)
+
+
+def _titles(chapter):
+    """Collect heading texts and table/figure titles for assertions."""
+    out = []
+    for b in chapter.blocks:
+        kind = getattr(b, "kind", None)
+        if kind == "heading":
+            out.append(("heading", getattr(b, "text", "")))
+        elif kind in ("data_table", "kv_table"):
+            out.append((kind, getattr(b, "title", "")))
+        elif kind == "group":
+            for inner in getattr(b, "blocks", []):
+                ik = getattr(inner, "kind", None)
+                if ik == "heading":
+                    out.append(("heading", getattr(inner, "text", "")))
+                elif ik in ("data_table", "kv_table"):
+                    out.append((ik, getattr(inner, "title", "")))
+                elif ik == "figure":
+                    out.append(("figure", getattr(inner, "caption", "")))
+        elif kind == "figure":
+            out.append(("figure", getattr(b, "caption", "")))
+    return out
+
+
+def _all_text(chapter):
+    parts = []
+    def walk(blocks):
+        for b in blocks:
+            for attr in ("text", "title", "note", "caption"):
+                v = getattr(b, attr, None)
+                if v:
+                    parts.append(str(v))
+            if getattr(b, "kind", None) == "group":
+                walk(getattr(b, "blocks", []))
+    walk(chapter.blocks)
+    return "\n".join(parts)
+
+
+def test_returns_none_when_no_missing_data():
+    profile = {
+        "n_rows": 4,
+        "null_cell_pct": 0.0,
+        "columns": [
+            {"name": "a", "null_count": 0, "null_pct": 0.0, "n_rows": 4},
+            {"name": "b", "null_count": 0, "null_pct": 0.0, "n_rows": 4},
+        ],
+    }
+    assert build_missingness(profile, {}) is None
+
+
+def test_activates_with_cooccurrence_via_raw_numeric():
+    # a and b are missing in EXACTLY the same rows (0,1,2) -> perfect absence
+    # correlation. c has no nulls. No db_path -> the chapter falls back to the
+    # numeric raw_numeric mask.
+    profile = {
+        "n_rows": 6,
+        "null_cell_pct": (0.5 + 0.5 + 0.0) / 3.0,
+        "columns": [
+            {"name": "a", "null_count": 3, "null_pct": 0.5, "n_rows": 6},
+            {"name": "b", "null_count": 3, "null_pct": 0.5, "n_rows": 6},
+            {"name": "c", "null_count": 0, "null_pct": 0.0, "n_rows": 6},
+        ],
+    }
+    glossary = model.GlossaryCollector()
+    ctx = {
+        "raw_numeric": {
+            "a": [None, None, None, 1.0, 2.0, 3.0],
+            "b": [None, None, None, 4.0, 5.0, 6.0],
+        },
+        "glossary": glossary,
+    }
+    ch = build_missingness(profile, ctx)
+    assert ch is not None
+    assert ch.id == "missingness"
+    assert ch.blocks
+
+    titles = _titles(ch)
+    headings = {t for (k, t) in titles if k == "heading"}
+    # Core sections present.
+    assert any("Cuánto y dónde" in h for h in headings)
+    assert any("Faltantes por columna" in h for h in headings)
+    assert any("Co-ocurrencia" in h for h in headings)
+    assert any("MCAR" in h for h in headings)
+    # A summary KVTable, a ranking DataTable, a co-occurrence figure and the
+    # pairs table all exist.
+    kinds = {k for (k, _) in titles}
+    assert "kv_table" in kinds
+    assert "data_table" in kinds
+    assert "figure" in kinds
+
+    # Glossary terms registered.
+    keys = {t["key"] for t in glossary.terms()}
+    assert {"missingness", "mcar", "mar"} <= keys
+
+    # The MCAR/MAR note reads the co-occurrence; with a perfect overlap it must
+    # flag the non-random (MAR) reading.
+    text = _all_text(ch)
+    assert "MAR" in text
+
+
+def test_db_pushdown_covers_categorical_column(tmp_path):
+    """The is-null mask push-down must cover a categorical column, so a
+    categorical that co-misses with a numeric one shows up in the pairs."""
+    import duckdb
+
+    db = str(tmp_path / "miss.duckdb")
+    con = duckdb.connect(db)
+    con.execute("CREATE TABLE t (num1 DOUBLE, num2 DOUBLE, cat VARCHAR)")
+    # num1 and cat are NULL together in the first 4 of 10 rows; num2 never null.
+    rows = []
+    for i in range(10):
+        if i < 4:
+            rows.append((None, float(i), None))
+        else:
+            rows.append((float(i), float(i), f"c{i}"))
+    con.executemany("INSERT INTO t VALUES (?,?,?)", rows)
+    con.close()
+
+    profile = {
+        "n_rows": 10,
+        "null_cell_pct": (0.4 + 0.0 + 0.4) / 3.0,
+        "columns": [
+            {"name": "num1", "null_count": 4, "null_pct": 0.4, "n_rows": 10},
+            {"name": "num2", "null_count": 0, "null_pct": 0.0, "n_rows": 10},
+            {"name": "cat", "null_count": 4, "null_pct": 0.4, "n_rows": 10},
+        ],
+    }
+    ctx = {"db_path": db, "table": "t", "glossary": model.GlossaryCollector()}
+    ch = build_missingness(profile, ctx)
+    assert ch is not None
+
+    # The pairs table must mention both num1 and cat (they co-miss perfectly),
+    # which is only possible if the mask covered the categorical column.
+    text = _all_text(ch)
+    assert "num1" in text and "cat" in text
+    # Co-occurrence section + a pairs data table exist.
+    titles = _titles(ch)
+    assert any("co-faltan" in (t or "").lower() for (k, t) in titles)

python/functions/datascience/automatic_eda/chapters/num_distr.py

@@ -35,10 +35,21 @@ try:
 except Exception:  # noqa: BLE001 — keep the chapter importable no matter what.
     build_boxplot_stats = None  # type: ignore[assignment]
 
-CHAPTER_VERSION = "1.2.0"
+CHAPTER_VERSION = "1.3.0"
 CHAPTER_ID = "num_distr"
 CHAPTER_TITLE = "Distribuciones numéricas"
 
+# Glossary term this chapter explains. The long "how to read the histogram and
+# the boxplot" paragraph used to live inline in the intro; it now lives in the
+# GLOSARIO chapter (canonical definition in ``glosario._BASELINE_TERMS``) and the
+# intro only names the clickable term — one click jumps to the full explanation,
+# so the information is relocated, not lost (mejora glosario).
+_TERM_HISTOBOX_KEY = "histograma_boxplot"
+_TERM_HISTOBOX_LABEL = "Cómo leer el histograma y el boxplot"
+
+# Key under which eda_llm_insights stores its interpretive block in the profile.
+LLM_KEY = "llm"
+
 # Plain-Spanish gloss for every label ``detect_distribution_type`` can emit, so a
 # non-expert reader understands the shape and the suggested next step (MUST-4.3).
 _DIST_GLOSS = {
@@ -99,6 +110,53 @@ def _numeric_columns(profile: dict) -> list:
     return out
 
 
+def _llm_index(profile: dict, ctx: dict) -> dict:
+    """Map column name -> its LLM dictionary entry (description/unit/...).
+
+    Reads the ``llm.dictionary`` list that ``eda_llm_insights`` stored in the
+    profile (``profile['llm']``; falls back to ``ctx['llm']``). Returns an empty
+    dict when ``run_llm`` did not run, so the caller degrades cleanly. Fully
+    defensive: never raises on malformed input.
+    """
+    llm = profile.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        llm = ctx.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        return {}
+    entries = llm.get("dictionary")
+    if not isinstance(entries, (list, tuple)):
+        return {}
+    index: dict = {}
+    for e in entries:
+        if not isinstance(e, dict):
+            continue
+        col = e.get("column")
+        if col is None:
+            continue
+        index[model._safe_str(col)] = e
+    return index
+
+
+def _llm_desc_unit_block(name: str, llm_index: dict):
+    """Markdown block with the LLM business description + unit of a column, or
+    None when no LLM entry matches the column (clean fallback without LLM)."""
+    entry = llm_index.get(model._safe_str(name))
+    if not isinstance(entry, dict):
+        return None
+    raw_desc = entry.get("description") or entry.get("business_meaning")
+    desc = " ".join(model._safe_str(raw_desc).split()) if raw_desc else ""
+    raw_unit = entry.get("unit")
+    unit = " ".join(model._safe_str(raw_unit).split()) if raw_unit else ""
+    parts = []
+    if desc:
+        parts.append(f"**Descripción:** {desc}")
+    if unit:
+        parts.append(f"**Unidad:** {unit}")
+    if not parts:
+        return None
+    return model.Markdown(text=" · ".join(parts))
+
+
 def _make_hist_box(name: str, numeric: dict, box: dict):
     """Build the histogram (with mean/median/±σ lines) + boxplot figure.
 
@@ -271,15 +329,26 @@ def build_num_distr(profile: dict, ctx: dict):
     if not numerics:
         return None  # chapter does not apply to a dataset with no numerics.
 
+    # Register the "how to read the histogram and boxplot" term in the shared
+    # glossary collector (if present) and mark its first appearance clickable. The
+    # full explanation (colour code, 1,5·IQR rule, asymmetry reading) lives in the
+    # GLOSARIO chapter instead of inline here: the intro only names the term.
+    glossary = ctx.get("glossary")
+    mark_term = False
+    if isinstance(glossary, model.GlossaryCollector):
+        glossary.add(_TERM_HISTOBOX_KEY, _TERM_HISTOBOX_LABEL)
+        mark_term = True
+    como_leer = ("[[term:histograma_boxplot]]cómo leer estos gráficos[[/term]]"
+                 if mark_term else "cómo leer estos gráficos")
     intro = (
-        "Para cada columna numérica se muestra su **histograma** con tres líneas "
+        "Cada columna numérica muestra su **histograma** (con la **media**, la "
-        "de referencia: la **media** (línea roja discontinua), la **mediana** "
+        "**mediana** y la banda **±1σ**) y, debajo y al mismo eje, su **boxplot "
-        "(línea verde continua) y la banda **±1σ** (zona sombreada). Debajo, "
+        f"de Tukey** — {como_leer}.")
-        "alineado al mismo eje, un **boxplot de Tukey**: la caja abarca del "
+
-        "primer al tercer cuartil (P25–P75), la línea interior es la mediana y "
+    # Business description + unit per column come from the LLM dictionary
-        "los bigotes llegan hasta 1,5·IQR; los puntos rojos señalan que hay "
+    # (profile['llm']['dictionary'], matched by column name); absent without
-        "valores más allá de las vallas. Comparar media y mediana revela la "
+    # run_llm, in which case the per-column description block is simply omitted.
-        "asimetría de la distribución.")
+    llm_index = _llm_index(profile, ctx)
 
     blocks = [
         model.Heading(text=CHAPTER_TITLE, level=1),
@@ -293,17 +362,20 @@ def build_num_distr(profile: dict, ctx: dict):
                 box = build_boxplot_stats(numeric) or {}
             except Exception:  # noqa: BLE001 — degrade, never raise.
                 box = {}
-        # Keep the column heading, its figure and its stats note together on the
+        # Keep the column heading, its (optional) LLM description, its figure and
-        # same page/slide (mejora 3 — keep-together): the renderers measure the
+        # its stats note together on the same page/slide (mejora 3 —
-        # whole Group and move it whole when it would not fit.
+        # keep-together): the renderers measure the whole Group and move it whole
-        blocks.append(model.Group(blocks=[
+        # when it would not fit.
-            model.Heading(text=str(name), level=2),
+        col_blocks = [model.Heading(text=str(name), level=2)]
-            model.Figure(
+        desc_block = _llm_desc_unit_block(name, llm_index)
-                make=_figure_maker(name, numeric, box),
+        if desc_block is not None:
-                caption=f"Distribución de «{name}» — histograma "
+            col_blocks.append(desc_block)
-                        f"(media/mediana/±σ) y boxplot."),
+        col_blocks.append(model.Figure(
-            model.Markdown(text=_stats_note(name, numeric, box)),
+            make=_figure_maker(name, numeric, box),
-        ]))
+            caption=f"Distribución de «{name}» — histograma "
+                    f"(media/mediana/±σ) y boxplot."))
+        col_blocks.append(model.Markdown(text=_stats_note(name, numeric, box)))
+        blocks.append(model.Group(blocks=col_blocks))
 
     return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
                          version=CHAPTER_VERSION, blocks=blocks)

python/functions/datascience/automatic_eda/chapters/num_distr_test.py

@@ -101,7 +101,7 @@ def test_golden_chapter_estructura_y_bloques():
 
 
 def test_golden_media_mediana_sigma_y_boxplot_presentes():
-    # The intro documents the three reference lines and the Tukey boxplot; the
+    # The short intro names the three reference lines and the Tukey boxplot; the
     # per-column note carries the actual mean/median/σ numbers and the shape.
     ch = build_num_distr(_profile(n_numeric=1, extra_categorical=False), {})
     md_texts = " ".join(b.text for b in _flatten(ch.blocks)
@@ -110,10 +110,58 @@ def test_golden_media_mediana_sigma_y_boxplot_presentes():
     assert "±1σ" in md_texts or "σ" in md_texts
     assert "boxplot" in md_texts.lower()
     assert "Tukey" in md_texts
+    # The long "how to read it" explanation moved to the glossary: the colour-code
+    # / 1,5·IQR walkthrough is no longer inline in the chapter body.
+    assert "1,5·IQR" not in md_texts
+    assert "línea roja" not in md_texts
     # distribution_type gloss surfaced for the column (right-skewed preset).
     assert _DIST_GLOSS["right-skewed"].split(";")[0][:20] in md_texts
 
 
+def test_glosario_histograma_boxplot_clicable_y_definicion():
+    # With a glossary collector the intro marks the clickable term and the FULL
+    # explanation (the long paragraph removed from the body) lands in the glossary.
+    from datascience.automatic_eda.chapters.glosario import build_glosario
+
+    gc = model.GlossaryCollector()
+    prof = _profile(n_numeric=1, extra_categorical=False)
+    ch = build_num_distr(prof, {"glossary": gc})
+    intro = next(b for b in ch.blocks if b.kind == "markdown")
+    assert "[[term:histograma_boxplot]]" in intro.text
+    assert gc.has("histograma_boxplot")
+    glos = build_glosario(prof, {"glossary": gc})
+    entry = next(b for b in glos.blocks
+                 if getattr(b, "kind", "") == "glossary_entry"
+                 and b.key == "histograma_boxplot")
+    assert "boxplot" in entry.definition.lower()
+    assert "1,5·IQR" in entry.definition
+
+
+def test_llm_descripcion_y_unidad_por_columna():
+    # With an LLM dictionary, each numeric column whose name matches shows its
+    # business description and unit in a per-column markdown block.
+    prof = _profile(n_numeric=2)
+    prof["llm"] = {"dictionary": [
+        {"column": "precio", "description": "Precio de venta del producto",
+         "unit": "EUR"},
+        {"column": "alcohol", "business_meaning": "Grado alcohólico",
+         "unit": "% vol"},
+    ]}
+    ch = build_num_distr(prof, {})
+    md_all = " ".join(b.text for b in _flatten(ch.blocks)
+                      if b.kind == "markdown")
+    assert "Precio de venta" in md_all and "EUR" in md_all
+    assert "Grado alcohólico" in md_all and "% vol" in md_all
+
+
+def test_edge_sin_llm_no_anade_descripcion():
+    # Without an LLM block the per-column description markdown is simply omitted.
+    ch = build_num_distr(_profile(n_numeric=2), {})
+    md_all = " ".join(b.text for b in _flatten(ch.blocks)
+                      if b.kind == "markdown")
+    assert "Descripción" not in md_all
+
+
 def test_boxplot_stats_se_consumen_del_registry():
     # The chapter must feed build_boxplot_stats (group eda) and the resulting
     # box must carry the Tukey fences for the figure.

python/functions/datascience/automatic_eda/chapters/outliers.py

@@ -0,0 +1,593 @@
+"""Outliers chapter (OUTLIERS) — univariate + multivariate atypical values.
+
+Today the analysis of atypical values is scattered across the document: the
+NUM DISTR chapter mentions the per-column outlier count inside each distribution
+figure, and the MODELOS chapter runs Isolation Forest as one of several cheap
+models. This chapter gathers and deepens the whole outlier story in a single
+place, with its interpretation: an [[term:outlier]]outlier[[/term]] is **not
+necessarily an error** — it can be a legitimate, extreme but real observation —
+so the reading is exploratory (what to look at), never confirmatory (what to
+delete).
+
+Sections, in order:
+
+1. **Resumen univariante por columna** — for every numeric column, the number
+   and percentage of atypical values by two complementary criteria: Tukey's
+   1.5·IQR rule ([[term:tukey_fence]]vallas de Tukey[[/term]]) and the
+   [[term:zscore]]z-score[[/term]] rule (|z| > 3). The most contaminated columns
+   are flagged. The fences come from the pure registry function
+   ``build_boxplot_stats`` (derived from the profile percentiles); the per-column
+   counts use the raw sample in ``ctx['raw_numeric']`` when available (the exact
+   count), degrading to the profile's own z-score counts otherwise.
+2. **Boxplots** — a single figure with the Tukey boxplots of the most
+   contaminated columns (box, whiskers and atypical points), delegated to the
+   reusable registry helper ``build_boxplots_figure``.
+3. **Multivariante (filas anómalas)** — rows that are atypical considering ALL
+   columns at once, via the registry function ``isolation_forest_outliers``: the
+   count and percentage of anomalous rows, the most anomalous rows with their
+   score, and the dimensions that make each one rare (top columns by |z|, via
+   ``summarize_outlier_dims``). Run live on ``ctx['raw_numeric']`` (the same
+   numeric columns ``summarize_outlier_dims`` uses, so the row indexing stays
+   coherent and the dimension breakdown is correct); falls back to the
+   precomputed ``profile['models']['outliers']`` only when no raw sample is
+   available (e.g. the lite preset), where no per-row breakdown is shown.
+4. **Interpretación** — outlier ≠ error: how to tell a data-entry error from a
+   genuine extreme value, and what to do (inspect, winsorize, or re-express —
+   linking to the Tukey re-expression the profile already computes).
+
+The chapter activates whenever the table has at least one numeric column; with
+no numeric column it returns ``None`` and disappears from the document.
+
+Reads everything defensively (``.get``) and never raises: every registry
+delegation is imported lazily and degraded to an honest note on any failure.
+
+Contract: build_<id>(profile, ctx) -> Chapter | None ; CHAPTER_VERSION = "x.y.z".
+"""
+
+from __future__ import annotations
+
+from .. import model
+
+CHAPTER_VERSION = "1.0.0"
+CHAPTER_ID = "outliers"
+CHAPTER_TITLE = "Valores atípicos"
+
+# z-score threshold for the univariate z rule: |z| > 3 flags a value ~3 standard
+# deviations from the mean (≈99.7% of a normal distribution lies within ±3σ).
+_Z_THRESH = 3.0
+# How many columns to draw in the boxplots figure (most contaminated first) and
+# how many anomalous rows to list in the multivariate table.
+_TOP_BOX = 12
+_TOP_ROWS = 12
+# Cap on the raw atypical values passed as boxplot fliers, so a heavy-tailed
+# column does not flood the figure with thousands of points.
+_MAX_FLIERS = 200
+# How many columns flagged as "most contaminated" in the summary note.
+_TOP_FLAGGED = 3
+
+# Glossary terms this chapter explains (contract §11.1). Registered in the shared
+# collector and marked clickable on first appearance. ``isolation_forest`` and
+# ``zscore`` may also be registered by the MODELOS chapter — ``add`` is
+# idempotent (first definition wins), so registering them here is harmless and
+# keeps this chapter self-contained when MODELOS does not render.
+_TERM_DEFS = {
+    "outlier": (
+        "Valor atípico (outlier)",
+        "Una observación que se aparta mucho del grueso de los datos. Un atípico "
+        "NO es necesariamente un error: puede ser un fallo de medida o de "
+        "registro, pero también un dato real extremo (un cliente que gasta diez "
+        "veces la media, un día de ventas excepcional). Por eso se señalan para "
+        "revisarlos, no para borrarlos automáticamente.",
+    ),
+    "tukey_fence": (
+        "Vallas de Tukey (1,5·IQR)",
+        "Regla clásica para marcar atípicos a partir de los cuartiles: se calcula "
+        "el rango intercuartílico IQR = P75 − P25 y se trazan dos vallas, una "
+        "inferior en P25 − 1,5·IQR y otra superior en P75 + 1,5·IQR. Los valores "
+        "que caen fuera de esas vallas se consideran atípicos. Es robusta porque "
+        "se apoya en la mediana y los cuartiles, no en la media.",
+    ),
+    "zscore": (
+        "z-score (puntuación típica)",
+        "Mide a cuántas desviaciones típicas está un valor de la media de su "
+        "columna: z = (valor − media) / desviación típica. Un |z| grande (aquí > "
+        "3) señala un valor alejado del centro. A diferencia de las vallas de "
+        "Tukey, el z-score usa media y desviación, así que es más sensible a la "
+        "presencia de los propios atípicos.",
+    ),
+    "isolation_forest": (
+        "Isolation Forest (anomalías multivariantes)",
+        "Algoritmo de detección de anomalías que considera TODAS las columnas a "
+        "la vez: construye árboles que parten el espacio con cortes aleatorios y "
+        "mide cuántos cortes hacen falta para aislar cada fila. Las filas raras "
+        "se aíslan con muy pocos cortes y se marcan como atípicas según un umbral "
+        "de contaminación. Detecta combinaciones de valores poco frecuentes que "
+        "ninguna columna por separado revelaría.",
+    ),
+}
+
+
+# --------------------------------------------------------------------------- #
+# Lazy registry delegations (each degrades to None / no-op on any failure).
+# --------------------------------------------------------------------------- #
+def _load_build_boxplot_stats():
+    try:
+        from datascience.build_boxplot_stats import build_boxplot_stats
+        return build_boxplot_stats
+    except Exception:  # noqa: BLE001
+        return None
+
+
+def _load_detect_outliers():
+    # detect_outliers lives in the monolithic ``datascience.datascience`` module
+    # (file_path datascience.py), not in its own submodule — try both shapes.
+    try:
+        from datascience.datascience import detect_outliers
+        return detect_outliers
+    except Exception:  # noqa: BLE001
+        try:
+            from datascience import detect_outliers
+            return detect_outliers
+        except Exception:  # noqa: BLE001
+            return None
+
+
+def _load_isolation_forest():
+    try:
+        from datascience.isolation_forest_outliers import isolation_forest_outliers
+        return isolation_forest_outliers
+    except Exception:  # noqa: BLE001
+        return None
+
+
+def _load_summarize_dims():
+    try:
+        from datascience.summarize_outlier_dims import summarize_outlier_dims
+        return summarize_outlier_dims
+    except Exception:  # noqa: BLE001
+        return None
+
+
+# --------------------------------------------------------------------------- #
+# Defensive formatters (own copy: the chapter never imports siblings).
+# --------------------------------------------------------------------------- #
+def _fmt_num(value, decimals: int = 3) -> str:
+    if value is None:
+        return "—"
+    if isinstance(value, bool):
+        return "sí" if value else "no"
+    if isinstance(value, int):
+        return f"{value:,}".replace(",", ".")
+    if isinstance(value, float):
+        if value != value:  # NaN
+            return "—"
+        if value in (float("inf"), float("-inf")):
+            return str(value)
+        text = f"{value:.{decimals}f}".rstrip("0").rstrip(".")
+        return text if text else "0"
+    return model._safe_str(value)
+
+
+def _fmt_int(value) -> str:
+    if value is None:
+        return "—"
+    try:
+        return f"{int(round(float(value))):,}".replace(",", ".")
+    except (TypeError, ValueError):
+        return model._safe_str(value)
+
+
+def _fmt_pct(value, decimals: int = 2) -> str:
+    """Format an already-0-100 value as a percentage. None -> placeholder."""
+    if value is None:
+        return "—"
+    try:
+        return f"{float(value):.{decimals}f}%"
+    except (TypeError, ValueError):
+        return model._safe_str(value)
+
+
+def _term(mark: bool, key: str, text: str) -> str:
+    return f"[[term:{key}]]{text}[[/term]]" if mark else text
+
+
+def _is_dict(v) -> bool:
+    return isinstance(v, dict)
+
+
+# --------------------------------------------------------------------------- #
+# Profile reads.
+# --------------------------------------------------------------------------- #
+def _numeric_columns(profile: dict) -> list:
+    """Return [(name, numeric_dict)] for numeric columns with usable stats."""
+    out = []
+    for col in profile.get("columns") or []:
+        if not isinstance(col, dict):
+            continue
+        if col.get("inferred_type") != "numeric":
+            continue
+        num = col.get("numeric")
+        if not isinstance(num, dict) or not num:
+            continue
+        if num.get("mean") is None and num.get("median") is None:
+            continue
+        out.append((col.get("name") or "(columna)", num))
+    return out
+
+
+def _clean_values(raw):
+    """Return the finite float values of a raw column list (drop None/NaN/inf)."""
+    if not isinstance(raw, (list, tuple)):
+        return None
+    vals = []
+    for v in raw:
+        if v is None or isinstance(v, bool):
+            continue
+        try:
+            f = float(v)
+        except (TypeError, ValueError):
+            continue
+        if f != f or f in (float("inf"), float("-inf")):
+            continue
+        vals.append(f)
+    return vals
+
+
+# --------------------------------------------------------------------------- #
+# Per-column univariate summary.
+# --------------------------------------------------------------------------- #
+def _univariate_row(name, numeric, raw_vals, box_fn, detect_fn):
+    """Compute one univariate summary row + boxplot inputs for a column.
+
+    Returns a dict with the table cells and, when raw values are available, the
+    exact Tukey/z counts and the list of atypical (flier) values; otherwise it
+    degrades to the profile's own z-score counts and the fence flags.
+    """
+    box = {}
+    if box_fn is not None:
+        try:
+            box = box_fn(numeric) or {}
+        except Exception:  # noqa: BLE001
+            box = {}
+    lf = box.get("lower_fence")
+    uf = box.get("upper_fence")
+
+    vals = _clean_values(raw_vals)
+    n_tukey = pct_tukey = None
+    n_z = pct_z = None
+    low_extreme = high_extreme = None
+    fliers = []
+    contamination = None  # metric used to rank columns (prefer Tukey %).
+
+    if vals:
+        n = len(vals)
+        tukey_out = []
+        for v in vals:
+            below = (lf is not None and v < lf)
+            above = (uf is not None and v > uf)
+            if below or above:
+                tukey_out.append(v)
+        n_tukey = len(tukey_out)
+        pct_tukey = 100.0 * n_tukey / n if n else None
+        if tukey_out:
+            low_extreme = min(tukey_out)
+            high_extreme = max(tukey_out)
+            fliers = tukey_out[:_MAX_FLIERS]
+        # z-score rule via the registry function (returns parallel bools).
+        if detect_fn is not None:
+            try:
+                flags = detect_fn(vals, _Z_THRESH) or []
+                n_z = int(sum(1 for b in flags if b))
+                pct_z = 100.0 * n_z / n if n else None
+            except Exception:  # noqa: BLE001
+                n_z = pct_z = None
+        contamination = pct_tukey
+    else:
+        # Degrade: no raw sample for this column. The profile's own outlier
+        # count/pct come from the z-score block (build_boxplot_stats note); the
+        # Tukey count is unknown, only the fence flags are.
+        n_z = numeric.get("n_outliers")
+        pct_z = numeric.get("outlier_pct")
+        if box.get("has_low_outliers") and box.get("min") is not None:
+            low_extreme = box.get("min")
+        if box.get("has_high_outliers") and box.get("max") is not None:
+            high_extreme = box.get("max")
+        contamination = pct_z if isinstance(pct_z, (int, float)) else None
+
+    # Compact "extremos atípicos" cell: down/up arrows for the low/high tail.
+    extremes = []
+    if low_extreme is not None:
+        extremes.append(f"↓ {_fmt_num(low_extreme)}")
+    if high_extreme is not None:
+        extremes.append(f"↑ {_fmt_num(high_extreme)}")
+    extremes_cell = "  ".join(extremes) if extremes else "—"
+
+    return {
+        "name": model._safe_str(name),
+        "n_tukey": n_tukey,
+        "pct_tukey": pct_tukey,
+        "n_z": n_z,
+        "pct_z": pct_z,
+        "lower_fence": lf,
+        "upper_fence": uf,
+        "extremes": extremes_cell,
+        "box": box,
+        "fliers": fliers,
+        "has_raw": bool(vals),
+        "contamination": contamination if isinstance(contamination, (int, float)) else -1.0,
+    }
+
+
+def _univariate_table(rows: list) -> model.DataTable:
+    header = ["Columna", "Atípicos Tukey", "% Tukey", "Atípicos z", "% z",
+              "Valla inf.", "Valla sup.", "Extremos atípicos"]
+    table_rows = []
+    for r in rows:
+        table_rows.append([
+            r["name"],
+            _fmt_int(r["n_tukey"]) if r["n_tukey"] is not None else "—",
+            _fmt_pct(r["pct_tukey"]) if r["pct_tukey"] is not None else "—",
+            _fmt_int(r["n_z"]) if r["n_z"] is not None else "—",
+            _fmt_pct(r["pct_z"]) if r["pct_z"] is not None else "—",
+            _fmt_num(r["lower_fence"]),
+            _fmt_num(r["upper_fence"]),
+            r["extremes"],
+        ])
+    return model.DataTable(
+        header=header, rows=table_rows,
+        title="Valores atípicos por columna",
+        note="Tukey = fuera de las vallas 1,5·IQR · z = |z-score| > 3 · "
+             "ordenado de más a menos contaminada")
+
+
+# --------------------------------------------------------------------------- #
+# Multivariate (Isolation Forest) section.
+# --------------------------------------------------------------------------- #
+def _resolve_multivariate(profile: dict, ctx: dict, raw_numeric):
+    """Return (outliers_dict_or_None, source).
+
+    Prefers a LIVE Isolation Forest over ``raw_numeric`` so the detector and
+    ``summarize_outlier_dims`` use EXACTLY the same numeric columns and the same
+    valid-row indexing — otherwise the precomputed ``profile['models']
+    ['outliers']`` (run by MODELOS over a possibly different column subset) would
+    yield ``row_index`` values that no longer point at the rows
+    ``summarize_outlier_dims`` reconstructs, mislabelling the "dimensions that
+    make each row rare". Falls back to the precomputed block when no raw sample
+    is available (e.g. the lite preset drops ``raw_numeric``)."""
+    if _is_dict(raw_numeric) and raw_numeric:
+        iso = _load_isolation_forest()
+        if iso is not None:
+            try:
+                out = iso(raw_numeric)
+                if _is_dict(out) and out.get("n_outliers") is not None and out.get("n_rows_used"):
+                    return out, "live"
+            except Exception:  # noqa: BLE001
+                pass
+    # Fallback: the model the MODELOS chapter already computed (no raw sample to
+    # recompute against, so no per-row dimension breakdown either).
+    models = profile.get("models") if _is_dict(profile.get("models")) else {}
+    pre = models.get("outliers") if _is_dict(models) else None
+    if _is_dict(pre) and pre.get("n_outliers") is not None and pre.get("n_rows_used"):
+        return pre, "precomputed"
+    return None, "none"
+
+
+def _multivariate_blocks(outliers: dict, raw_numeric, mark: bool) -> list:
+    isof = _term(mark, "isolation_forest", "**Isolation Forest**")
+    blocks = [
+        model.Heading(text="Filas atípicas (multivariante)", level=2),
+        model.Markdown(text=(
+            f"Hasta aquí cada columna se ha mirado por separado. {isof} busca "
+            "filas raras considerando **todas las columnas a la vez**: una fila "
+            "puede ser normal en cada variable y aun así ser atípica por la "
+            "**combinación** de sus valores (p. ej. una edad baja con una tarifa "
+            "muy alta). La tabla resume cuántas filas se marcaron y el umbral de "
+            "decisión.")),
+        model.KVTable(rows=[
+            ("Filas analizadas", _fmt_int(outliers.get("n_rows_used"))),
+            ("Columnas consideradas", _fmt_int(outliers.get("n_features"))),
+            ("Filas atípicas", _fmt_int(outliers.get("n_outliers"))),
+            ("% filas atípicas", _fmt_pct(outliers.get("outlier_pct"))),
+            ("Umbral de decisión", _fmt_num(outliers.get("threshold"), 4)),
+        ], title="Anomalías multivariantes"),
+    ]
+
+    rows_in = outliers.get("outlier_rows") or []
+    if not rows_in:
+        return blocks
+
+    # Enrich each anomalous row with the dimensions that make it rare, when the
+    # raw sample is available (summarize_outlier_dims reconstructs the same
+    # valid-row indexing as isolation_forest_outliers).
+    dims_by_row = {}
+    if _is_dict(raw_numeric) and raw_numeric:
+        summ = _load_summarize_dims()
+        if summ is not None:
+            try:
+                enriched = summ(raw_numeric, rows_in, top_k=3) or []
+                for e in enriched:
+                    if _is_dict(e) and e.get("row_index") is not None:
+                        dims_by_row[e.get("row_index")] = e.get("dims") or []
+            except Exception:  # noqa: BLE001
+                dims_by_row = {}
+
+    has_dims = bool(dims_by_row)
+    header = ["Fila (entre válidas)", "Score"]
+    if has_dims:
+        header.append("Dimensiones que la hacen rara (col = valor, z)")
+    table_rows = []
+    for r in rows_in[:_TOP_ROWS]:
+        if not _is_dict(r):
+            continue
+        ridx = r.get("row_index")
+        cells = [_fmt_int(ridx), _fmt_num(r.get("score"), 4)]
+        if has_dims:
+            dims = dims_by_row.get(ridx) or []
+            parts = []
+            for d in dims:
+                if not _is_dict(d):
+                    continue
+                parts.append(
+                    f"{model._safe_str(d.get('col'))} = {_fmt_num(d.get('value'))} "
+                    f"(z {_fmt_num(d.get('z'), 2)})")
+            cells.append("; ".join(parts) if parts else "—")
+        table_rows.append(cells)
+
+    if table_rows:
+        shown = len(table_rows)
+        total = outliers.get("n_outliers")
+        note = "las filas más anómalas primero (score más bajo = más rara)"
+        if isinstance(total, int) and total > shown:
+            note += f" — top {shown} de {total}"
+        if not has_dims:
+            note += (" · no se pudo recuperar la muestra cruda para explicar las "
+                     "dimensiones de cada fila")
+        blocks.append(model.DataTable(
+            header=header, rows=table_rows,
+            title="Filas más atípicas", note=note))
+    return blocks
+
+
+# --------------------------------------------------------------------------- #
+# Interpretation section.
+# --------------------------------------------------------------------------- #
+def _interpretation_block(mark: bool) -> model.Markdown:
+    outlier = _term(mark, "outlier", "atípico")
+    text = (
+        f"**Un {outlier} no es necesariamente un error.** Conviene distinguir "
+        "dos casos antes de actuar:\n\n"
+        "- **Error de dato** (medida, registro o unidad equivocada): una edad de "
+        "200 años, un importe negativo donde no puede haberlo, un decimal "
+        "desplazado. Estos sí se corrigen o se eliminan, idealmente en el origen.\n"
+        "- **Dato real extremo**: una observación legítima de la cola de la "
+        "distribución (un cliente que gasta mucho más, una tarifa de lujo, un día "
+        "de ventas excepcional). Borrarla sesga el análisis y oculta información "
+        "valiosa.\n\n"
+        "**Qué hacer.** Primero, **revisar** los valores señalados arriba contra "
+        "su origen para decidir cuál de los dos casos es. Si son errores, "
+        "corregirlos. Si son datos reales que distorsionan medias y modelos, hay "
+        "alternativas a borrarlos: **winsorizar** (recortar los extremos a un "
+        "percentil), o **re-expresar** la variable (por ejemplo una "
+        "transformación logarítmica o la escalera de re-expresión de Tukey que "
+        "este mismo perfil ya calcula para las columnas asimétricas), que suele "
+        "domar la cola sin perder ninguna fila. La elección depende del objetivo: "
+        "esta lectura es **exploratoria** —orienta dónde mirar—, no una regla "
+        "automática de limpieza.")
+    return model.Markdown(text=text)
+
+
+# --------------------------------------------------------------------------- #
+# Entry point.
+# --------------------------------------------------------------------------- #
+def build_outliers(profile: dict, ctx: dict):
+    """Build the OUTLIERS Chapter, or None if the dataset has no numeric column."""
+    profile = profile or {}
+    ctx = ctx or {}
+    if not isinstance(profile, dict):
+        return None
+
+    numerics = _numeric_columns(profile)
+    if not numerics:
+        return None  # chapter does not apply to a dataset with no numerics.
+
+    # Register glossary terms (if a collector is present) and mark them clickable.
+    glossary = ctx.get("glossary")
+    mark = False
+    if isinstance(glossary, model.GlossaryCollector):
+        for key, (label, definition) in _TERM_DEFS.items():
+            glossary.add(key, label, definition)
+        mark = True
+
+    raw_numeric = ctx.get("raw_numeric")
+    raw_numeric = raw_numeric if isinstance(raw_numeric, dict) else {}
+
+    box_fn = _load_build_boxplot_stats()
+    detect_fn = _load_detect_outliers()
+
+    # --- Univariate summary ------------------------------------------------- #
+    uni_rows = []
+    for name, numeric in numerics:
+        uni_rows.append(_univariate_row(
+            name, numeric, raw_numeric.get(name), box_fn, detect_fn))
+    # Rank columns by contamination (Tukey % when available, else z %).
+    uni_rows.sort(key=lambda r: r.get("contamination", -1.0), reverse=True)
+
+    intro = (
+        "Este capítulo reúne en un solo sitio el análisis de los **valores "
+        "atípicos** de la tabla, que en el resto del informe aparecen dispersos. "
+        f"Un {_term(mark, 'outlier', 'atípico')} es una observación que se aparta "
+        "mucho del grueso de los datos. Cada columna numérica se evalúa con dos "
+        f"criterios complementarios: las {_term(mark, 'tukey_fence', 'vallas de Tukey')} "
+        "(fuera de P25−1,5·IQR o P75+1,5·IQR, robusto a la propia cola) y el "
+        f"{_term(mark, 'zscore', 'z-score')} (|z| > 3, sensible a la media). La "
+        "tabla está ordenada de la columna más contaminada a la menos.")
+
+    blocks = [
+        model.Heading(text=CHAPTER_TITLE, level=1),
+        model.Markdown(text=intro),
+        _univariate_table(uni_rows),
+    ]
+
+    # Flag the most contaminated columns explicitly.
+    flagged = [r["name"] for r in uni_rows
+               if r.get("contamination", -1.0) > 0][:_TOP_FLAGGED]
+    if flagged:
+        names = ", ".join(f"**{n}**" for n in flagged)
+        blocks.append(model.Markdown(text=(
+            f"Las columnas con mayor proporción de atípicos son {names}: "
+            "concentran el grueso de los valores fuera de las vallas y son las "
+            "primeras a revisar.")))
+
+    # --- Boxplots figure ---------------------------------------------------- #
+    box_entries = [
+        {"name": r["name"], "box": r["box"], "fliers": r.get("fliers")}
+        for r in uni_rows
+        if r.get("box")
+    ][:_TOP_BOX]
+    if box_entries:
+        def _boxplots_make(entries=box_entries):
+            try:
+                from datascience.build_boxplots_figure import build_boxplots_figure
+                return build_boxplots_figure(
+                    entries, title="Boxplots de Tukey por columna",
+                    max_boxes=_TOP_BOX)
+            except Exception:  # noqa: BLE001 — minimal fallback figure.
+                import matplotlib
+                matplotlib.use("Agg")
+                from matplotlib.figure import Figure
+                fig = Figure(figsize=(5.0, 2.2))
+                ax = fig.add_subplot(111)
+                ax.text(0.5, 0.5, "(boxplots no disponibles)",
+                        ha="center", va="center")
+                ax.axis("off")
+                return fig
+
+        blocks.append(model.Group(blocks=[
+            model.Heading(text="Boxplots", level=2),
+            model.Markdown(text=(
+                "Cada caja abarca del primer al tercer cuartil (P25–P75), la línea "
+                "interior es la mediana y los bigotes llegan hasta 1,5·IQR; los "
+                "puntos son los valores que caen fuera de las vallas (atípicos por "
+                "Tukey).")),
+            model.Figure(
+                make=_boxplots_make,
+                caption="Boxplots de Tukey de las columnas más contaminadas."),
+        ]))
+
+    # --- Multivariate ------------------------------------------------------- #
+    outliers, _src = _resolve_multivariate(profile, ctx, raw_numeric)
+    if outliers is not None:
+        blocks.extend(_multivariate_blocks(outliers, raw_numeric, mark))
+    else:
+        blocks.append(model.Heading(text="Filas atípicas (multivariante)", level=2))
+        blocks.append(model.Note(
+            "No se pudo analizar la anomalía multivariante: hacen falta al menos "
+            "dos columnas numéricas y la muestra cruda (o los modelos del perfil) "
+            "para correr Isolation Forest."))
+
+    # --- Interpretation ----------------------------------------------------- #
+    blocks.append(model.Heading(text="Cómo interpretar los atípicos", level=2))
+    blocks.append(_interpretation_block(mark))
+
+    return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
+                         version=CHAPTER_VERSION, blocks=blocks)

python/functions/datascience/automatic_eda/chapters/outliers_test.py

@@ -0,0 +1,304 @@
+"""Tests for the OUTLIERS chapter — DoD: golden + edges + error path.
+
+Self-contained: builds synthetic ``numeric`` blocks + a raw_numeric sample (no
+DuckDB) so the suite is fast and deterministic. Verifies that the chapter emits
+the univariate per-column table, a boxplots figure, the multivariate Isolation
+Forest section and the outlier≠error interpretation; that the most contaminated
+column is ranked first; that a profile with no numeric column yields None; that
+None/empty never raises; that the glossary terms are registered; and that the
+chapter renders into both PDF and PPTX without cutting its title.
+"""
+
+import math
+import os
+import re
+import tempfile
+
+from pypdf import PdfReader
+
+from datascience.automatic_eda.chapters.outliers import (
+    build_outliers, CHAPTER_VERSION, CHAPTER_TITLE, _TERM_DEFS,
+)
+from datascience.automatic_eda import model
+from datascience.render_automatic_eda_pdf import render_automatic_eda_pdf
+from datascience.render_automatic_eda_pptx import render_automatic_eda_pptx
+
+
+def _percentile(sorted_vals, q):
+    """Linear-interpolation percentile (q in 0..1) on an already-sorted list."""
+    if not sorted_vals:
+        return None
+    if len(sorted_vals) == 1:
+        return float(sorted_vals[0])
+    pos = q * (len(sorted_vals) - 1)
+    lo = int(math.floor(pos))
+    hi = int(math.ceil(pos))
+    if lo == hi:
+        return float(sorted_vals[lo])
+    frac = pos - lo
+    return float(sorted_vals[lo] * (1 - frac) + sorted_vals[hi] * frac)
+
+
+def _col_from_values(values, nbins=10):
+    """Build a ``numeric`` sub-block shaped like describe_numeric's output from a
+    concrete list of raw values, so the profile percentiles and the raw sample
+    are consistent (the boxplot fences match the crudo)."""
+    vals = [float(v) for v in values]
+    s = sorted(vals)
+    n = len(s)
+    mean = sum(vals) / n
+    var = sum((v - mean) ** 2 for v in vals) / n
+    std = math.sqrt(var)
+    median = _percentile(s, 0.5)
+    p25 = _percentile(s, 0.25)
+    p75 = _percentile(s, 0.75)
+    mn, mx = s[0], s[-1]
+    # z-score outlier count (population), what the profile's n_outliers carries.
+    n_out = sum(1 for v in vals if std > 0 and abs((v - mean) / std) > 3.0)
+    width = (mx - mn) / nbins if mx > mn else 1.0
+    hist = [{"lo": mn + i * width, "hi": mn + (i + 1) * width, "count": 1}
+            for i in range(nbins)]
+    return {
+        "min": mn, "max": mx, "mean": mean, "median": median, "std": std,
+        "p25": p25, "p50": median, "p75": p75, "iqr": (p75 - p25),
+        "n_outliers": n_out, "outlier_pct": 100.0 * n_out / n,
+        "distribution_type": "right-skewed", "histogram": hist,
+    }
+
+
+def _fare_values():
+    """A heavy-tailed column (most ~10-30, a few 200-512): clear Tukey/z outliers."""
+    base = [7.0 + (i % 25) for i in range(120)]      # bulk 7..31
+    tail = [180.0, 210.0, 263.0, 512.0]              # extreme upper tail
+    return base + tail
+
+
+def _age_values():
+    """A roughly symmetric column with one extreme low value."""
+    base = [22.0 + (i % 40) for i in range(120)]     # 22..61
+    return base + [80.0, 0.5, 74.0, 1.0]
+
+
+def _quiet_values():
+    """A clean column with no atypical values."""
+    return [50.0 + (i % 5) for i in range(124)]
+
+
+def _profile_and_ctx(with_models=True, with_raw=True):
+    fare = _fare_values()
+    age = _age_values()
+    quiet = _quiet_values()
+    cols = [
+        {"name": "Fare", "inferred_type": "numeric", "numeric": _col_from_values(fare)},
+        {"name": "Age", "inferred_type": "numeric", "numeric": _col_from_values(age)},
+        {"name": "Quiet", "inferred_type": "numeric", "numeric": _col_from_values(quiet)},
+        {"name": "Sexo", "inferred_type": "categorical",
+         "categorical": {"top": [{"value": "male", "count": 80}]}},
+    ]
+    profile = {"table": "titanic", "n_rows": len(fare), "n_cols": len(cols),
+               "columns": cols}
+    if with_models:
+        profile["models"] = {
+            "outliers": {
+                "n_outliers": 4, "outlier_pct": 3.2,
+                "outlier_rows": [
+                    {"row_index": 123, "score": -0.21},
+                    {"row_index": 121, "score": -0.15},
+                ],
+                "threshold": -0.02, "n_rows_used": 124, "n_features": 3,
+            }
+        }
+    ctx = {}
+    if with_raw:
+        ctx["raw_numeric"] = {"Fare": fare, "Age": age, "Quiet": quiet}
+    return profile, ctx
+
+
+def _pdf_text(path: str) -> str:
+    txt = "".join((pg.extract_text() or "") for pg in PdfReader(path).pages)
+    return re.sub(r"\s+", " ", txt)
+
+
+def _flatten(blocks):
+    out = []
+    for b in blocks:
+        if getattr(b, "kind", "") == "group":
+            out.extend(_flatten(getattr(b, "blocks", []) or []))
+        else:
+            out.append(b)
+    return out
+
+
+# --------------------------------------------------------------------------- #
+# Golden.
+# --------------------------------------------------------------------------- #
+def test_golden_estructura_y_secciones():
+    profile, ctx = _profile_and_ctx()
+    ctx["glossary"] = model.GlossaryCollector()
+    ch = build_outliers(profile, ctx)
+    assert ch is not None
+    assert ch.id == "outliers"
+    assert ch.version == CHAPTER_VERSION
+
+    flat = _flatten(ch.blocks)
+    kinds = [b.kind for b in flat]
+    # Title heading + univariate DataTable + boxplots Figure + multivariate
+    # KVTable + interpretation Markdown.
+    assert kinds[0] == "heading" and flat[0].text == CHAPTER_TITLE
+    tables = [b for b in flat if b.kind == "data_table"]
+    titles = [t.title for t in tables]
+    assert any(t and "atípicos por columna" in t for t in titles)
+    assert any(b.kind == "figure" for b in flat), "falta la figura de boxplots"
+    assert any(b.kind == "kv_table" for b in flat), "falta el resumen multivariante"
+
+    # The boxplots figure maker yields a real matplotlib figure (or its fallback).
+    fig = next(b for b in flat if b.kind == "figure").make()
+    assert fig is not None
+    import matplotlib.pyplot as plt
+    plt.close(fig)
+
+
+def test_golden_fare_es_la_mas_contaminada():
+    # The univariate table must rank Fare (heavy tail) first and report a
+    # non-zero Tukey percentage for it.
+    profile, ctx = _profile_and_ctx()
+    ch = build_outliers(profile, ctx)
+    table = next(b for b in _flatten(ch.blocks)
+                 if b.kind == "data_table" and b.title
+                 and "atípicos por columna" in b.title)
+    first_col = table.rows[0][0]
+    assert first_col == "Fare", f"esperaba Fare primera, fue {first_col}"
+    # % Tukey column (index 2) of the first row must be > 0.
+    pct_cell = table.rows[0][2]
+    assert pct_cell not in ("—", "0%", "0.00%"), f"% Tukey de Fare vacío: {pct_cell}"
+    # The z-score rule (detect_outliers) must actually run with raw_numeric: at
+    # least one column reports a non-empty z count/percentage (regression guard
+    # for the detect_outliers import path).
+    z_pcts = [r[4] for r in table.rows]
+    assert any(c not in ("—",) for c in z_pcts), f"columna z toda vacía: {z_pcts}"
+    z_counts = [r[3] for r in table.rows]
+    assert any(c not in ("—",) for c in z_counts), f"conteo z vacío: {z_counts}"
+
+
+def test_golden_interpretacion_outlier_no_es_error():
+    profile, ctx = _profile_and_ctx()
+    ch = build_outliers(profile, ctx)
+    md = " ".join(b.text for b in _flatten(ch.blocks) if b.kind == "markdown")
+    assert "no es necesariamente un error" in md.lower()
+    # Mentions the actionable options (winsorize / re-express).
+    assert "winsoriz" in md.lower()
+    assert "re-expres" in md.lower() or "logarítmic" in md.lower()
+
+
+def test_golden_terminos_glosario_registrados():
+    profile, ctx = _profile_and_ctx()
+    gloss = model.GlossaryCollector()
+    ctx["glossary"] = gloss
+    build_outliers(profile, ctx)
+    for key in _TERM_DEFS:
+        assert gloss.has(key), f"término '{key}' no registrado en el glosario"
+    # Terms are marked clickable in the body text.
+    md = " ".join(b.text for b in _flatten(build_outliers(profile, ctx).blocks)
+                  if b.kind == "markdown")
+    assert "[[term:outlier]]" in md and "[[term:tukey_fence]]" in md
+
+
+# --------------------------------------------------------------------------- #
+# Multivariate.
+# --------------------------------------------------------------------------- #
+def test_multivariante_live_con_raw_y_dims():
+    # With a raw sample the chapter runs Isolation Forest live (over the same
+    # columns summarize_outlier_dims uses) and lists the anomalous rows with the
+    # dimensions that make each one rare.
+    profile, ctx = _profile_and_ctx(with_models=False, with_raw=True)
+    ch = build_outliers(profile, ctx)
+    flat = _flatten(ch.blocks)
+    kv = next(b for b in flat if b.kind == "kv_table")
+    flat_kv = " ".join(f"{k} {v}" for (k, v) in kv.rows)
+    assert "Filas atípicas" in flat_kv
+    # A non-zero number of anomalous rows is reported.
+    n_cell = dict(kv.rows).get("Filas atípicas")
+    assert n_cell not in (None, "—", "0"), f"sin filas atípicas: {n_cell}"
+    # The anomalous-rows table carries the per-row dimension breakdown.
+    tbls = [b for b in flat if b.kind == "data_table" and b.title
+            and "más atípicas" in b.title]
+    assert tbls, "falta la tabla de filas más atípicas"
+    assert any("hacen rara" in h for h in tbls[0].header), \
+        f"falta la columna de dimensiones: {tbls[0].header}"
+
+
+def test_multivariante_precomputed_sin_raw():
+    # Without a raw sample the chapter falls back to profile['models']['outliers']
+    # (lite preset path); the precomputed n_outliers (4) surfaces in the KV table.
+    profile, ctx = _profile_and_ctx(with_models=True, with_raw=False)
+    ch = build_outliers(profile, ctx)
+    kv = next(b for b in _flatten(ch.blocks) if b.kind == "kv_table")
+    assert any("4" in str(v) for (k, v) in kv.rows)
+
+
+def test_multivariante_ausente_degrada_a_nota():
+    # No models and no raw sample → an honest note, never a crash.
+    profile, ctx = _profile_and_ctx(with_models=False, with_raw=False)
+    ch = build_outliers(profile, ctx)
+    assert ch is not None
+    notes = [b.text for b in _flatten(ch.blocks) if b.kind == "note"]
+    assert any("Isolation Forest" in n for n in notes)
+
+
+# --------------------------------------------------------------------------- #
+# Edges / error path.
+# --------------------------------------------------------------------------- #
+def test_edge_sin_columnas_numericas_devuelve_none():
+    prof = {"columns": [{"name": "c", "inferred_type": "categorical",
+                         "categorical": {"top": [{"value": "x", "count": 3}]}}]}
+    assert build_outliers(prof, {}) is None
+
+
+def test_edge_solo_texto_sintetico_devuelve_none():
+    # A text-only synthetic table (no numeric column) yields None (does not break).
+    prof = {"table": "notas", "n_rows": 3, "n_cols": 1,
+            "columns": [{"name": "comentario", "inferred_type": "text",
+                         "text": {"n_docs": 3}}]}
+    assert build_outliers(prof, {}) is None
+
+
+def test_edge_profile_none_y_vacio_no_revienta():
+    assert build_outliers(None, None) is None
+    assert build_outliers({}, {}) is None
+    assert build_outliers({"columns": []}, {}) is None
+
+
+def test_edge_sin_raw_numeric_degrada_a_perfil():
+    # Without raw_numeric the chapter still builds, using the profile z-score
+    # counts; the univariate table exists and Tukey counts degrade to '—'.
+    profile, ctx = _profile_and_ctx(with_models=True, with_raw=False)
+    ch = build_outliers(profile, ctx)
+    assert ch is not None
+    table = next(b for b in _flatten(ch.blocks)
+                 if b.kind == "data_table" and b.title
+                 and "atípicos por columna" in b.title)
+    # z column comes from the profile; Tukey count is unknown ('—').
+    assert all(len(r) == 8 for r in table.rows)
+
+
+# --------------------------------------------------------------------------- #
+# Anti-cut render.
+# --------------------------------------------------------------------------- #
+def test_render_pdf_y_pptx_incluyen_el_capitulo():
+    profile, ctx = _profile_and_ctx()
+    # The renderers build the whole document; the chapter is reached via the
+    # registry. Render the chapter standalone through a one-chapter document by
+    # passing the profile directly (the renderers run the full chapter registry).
+    with tempfile.TemporaryDirectory() as d:
+        pdf = os.path.join(d, "out.pdf")
+        res_pdf = render_automatic_eda_pdf(profile, pdf,
+                                           {"write_manifest": False, "ctx": ctx})
+        assert res_pdf["path"] == pdf
+        txt = _pdf_text(pdf)
+        assert CHAPTER_TITLE in txt, "el capítulo OUTLIERS no aparece en el PDF"
+        assert "Fare" in txt
+        pptx = os.path.join(d, "out.pptx")
+        res_pptx = render_automatic_eda_pptx(profile, pptx,
+                                             {"write_manifest": False, "ctx": ctx})
+        assert res_pptx["path"] == pptx
+        assert res_pptx["n_slides"] >= 1

python/functions/datascience/automatic_eda/chapters/overview.py

@@ -7,11 +7,21 @@ as needed, the renderers paginate):
    NOT carry the raw head, so this is read from ``ctx['head_rows']`` /
    ``profile['head_rows']`` (a list of row dicts). When absent the chapter shows
    an honest placeholder documenting the missing key instead of inventing data.
-2. Column dictionary — name / type / nulls / non-null examples. Examples come
+2. Column dictionary — name / type / nulls / non-null examples plus, when the
+   LLM layer ran, the business **description** and **unit** of each column so the
+   reader knows at a glance what every column is and in which unit. Examples come
    from ``columns[i]['examples']`` when present; otherwise they are derived from
    real non-null profile values (categorical top values, numeric min/median/max)
    so the cell is never empty nor fabricated.
-3. ``df.describe`` — mean / median / min / max / std for every numeric column.
+3. ``df.describe`` — mean / median / min / max / std for every numeric column,
+   plus its **unit** (same LLM source) so the stats read in context.
+
+The description/unit come from the ``llm`` block that ``eda_llm_insights`` (group
+``eda``) already stored in the profile (``profile['llm']['dictionary']``, a list
+of ``{"column","description","business_meaning","unit"}`` entries) — this chapter
+only **consumes** it, matching by column name; it never calls the LLM nor
+recomputes anything. When the block is absent (``run_llm`` did not run) those
+cells degrade to ``"—"`` and the tables still render.
 
 Contract: build_<id>(profile, ctx) -> Chapter | None ; CHAPTER_VERSION = "x.y.z".
 """
@@ -20,13 +30,59 @@ from __future__ import annotations
 
 from .. import model
 
-CHAPTER_VERSION = "1.1.0"
+CHAPTER_VERSION = "1.2.0"
 CHAPTER_ID = "overview"
 CHAPTER_TITLE = "Overview"
 
 # Profile/ctx keys the calculation phase must add for a full head + examples.
 HEAD_KEY = "head_rows"          # list[dict] — df.head(n)
 EXAMPLES_KEY = "examples"       # per column: list of non-null sample values
+LLM_KEY = "llm"                 # interpretive block from eda_llm_insights
+
+
+def _llm_dict_index(profile: dict, ctx: dict) -> dict:
+    """Map column name -> its LLM dictionary entry (description/unit/...).
+
+    Reads the ``llm.dictionary`` list that ``eda_llm_insights`` stored in the
+    profile (``profile['llm']``; falls back to ``ctx['llm']``). Returns an empty
+    dict when no LLM block ran, so the caller degrades to "—" cells. Fully
+    defensive: never raises on malformed input.
+    """
+    llm = profile.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        llm = ctx.get(LLM_KEY)
+    if not isinstance(llm, dict):
+        return {}
+    entries = llm.get("dictionary")
+    if not isinstance(entries, (list, tuple)):
+        return {}
+    index: dict = {}
+    for e in entries:
+        if not isinstance(e, dict):
+            continue
+        col = e.get("column")
+        if col is None:
+            continue
+        index[model._safe_str(col)] = e
+    return index
+
+
+def _llm_desc(entry) -> str:
+    """Business description of a column from its LLM entry, or "—"."""
+    if not isinstance(entry, dict):
+        return "—"
+    raw = entry.get("description") or entry.get("business_meaning")
+    text = " ".join(model._safe_str(raw).split()) if raw is not None else ""
+    return text or "—"
+
+
+def _llm_unit(entry) -> str:
+    """Unit of a column from its LLM entry, or "—"."""
+    if not isinstance(entry, dict):
+        return "—"
+    raw = entry.get("unit")
+    text = " ".join(model._safe_str(raw).split()) if raw is not None else ""
+    return text or "—"
 
 
 def _fmt_num(value, decimals: int = 3) -> str:
@@ -104,9 +160,12 @@ def _head_block(profile: dict, ctx: dict):
         "pasarlo en ctx['head_rows'] para mostrar las primeras filas.")
 
 
-def _columns_block(profile: dict):
+def _columns_block(profile: dict, llm_index: dict):
     cols = profile.get("columns") or []
-    header = ["Columna", "Tipo", "Nulos", "Ejemplos (no nulos)"]
+    # Descripción / Unidad come from the LLM dictionary (matched by column name);
+    # they read "—" when run_llm did not run, so the table always renders.
+    header = ["Columna", "Tipo", "Nulos", "Ejemplos (no nulos)",
+              "Descripción", "Unidad"]
     rows = []
     for c in cols:
         if not isinstance(c, dict):
@@ -126,15 +185,18 @@ def _columns_block(profile: dict):
             nulls = str(null_count)
         else:
             nulls = "—"
-        rows.append([name, ctype, nulls, _examples_for(c)])
+        entry = llm_index.get(model._safe_str(name))
+        rows.append([name, ctype, nulls, _examples_for(c),
+                     _llm_desc(entry), _llm_unit(entry)])
     if not rows:
         return None
     return model.DataTable(header=header, rows=rows, title="Columnas")
 
 
-def _describe_block(profile: dict):
+def _describe_block(profile: dict, llm_index: dict):
     cols = profile.get("columns") or []
-    header = ["Columna", "mean", "median", "min", "max", "std"]
+    # "Unidad" (LLM source) lets the reader know in which unit each stat is.
+    header = ["Columna", "mean", "median", "min", "max", "std", "Unidad"]
     rows = []
     for c in cols:
         if not isinstance(c, dict) or c.get("inferred_type") != "numeric":
@@ -142,13 +204,16 @@ def _describe_block(profile: dict):
         num = c.get("numeric") or {}
         if not num:
             continue
+        name = c.get("name") or "(col)"
+        entry = llm_index.get(model._safe_str(name))
         rows.append([
-            c.get("name") or "(col)",
+            name,
             _fmt_num(num.get("mean")),
             _fmt_num(num.get("median")),
             _fmt_num(num.get("min")),
             _fmt_num(num.get("max")),
             _fmt_num(num.get("std")),
+            _llm_unit(entry),
         ])
     if not rows:
         return None
@@ -163,16 +228,18 @@ def build_overview(profile: dict, ctx: dict):
     if not cols and not (ctx.get(HEAD_KEY) or profile.get(HEAD_KEY)):
         return None
 
+    llm_index = _llm_dict_index(profile, ctx)
+
     blocks = [
         model.Heading(text="Primeras filas (df.head)", level=2),
         _head_block(profile, ctx),
     ]
-    cols_block = _columns_block(profile)
+    cols_block = _columns_block(profile, llm_index)
     if cols_block is not None:
         blocks.append(model.Heading(
             text="Diccionario de columnas", level=2))
         blocks.append(cols_block)
-    desc_block = _describe_block(profile)
+    desc_block = _describe_block(profile, llm_index)
     if desc_block is not None:
         blocks.append(model.Heading(
             text="Resumen estadístico numérico", level=2))

python/functions/datascience/automatic_eda/chapters/overview_test.py

@@ -56,7 +56,21 @@ def _head_rows() -> list:
     ]
 
 
-def _profile(with_head: bool = True) -> dict:
+def _llm() -> dict:
+    """Interpretive block as eda_llm_insights stores it under profile['llm']."""
+    return {
+        "summary": "Pasajeros del Titanic.",
+        "dictionary": [
+            {"column": "PassengerId", "description": "Identificador del pasajero",
+             "business_meaning": "Clave única de cada pasajero", "unit": "id"},
+            {"column": "Pclass", "description": "Clase del billete",
+             "business_meaning": "Clase socioeconómica", "unit": "clase (1-3)"},
+            # No entry for Survived/Name/Sex on purpose -> they degrade to "—".
+        ],
+    }
+
+
+def _profile(with_head: bool = True, with_llm: bool = False) -> dict:
     prof = {
         "table": "titanic",
         "source": "/data/titanic.csv",
@@ -68,6 +82,8 @@ def _profile(with_head: bool = True) -> dict:
     }
     if with_head:
         prof["head_rows"] = _head_rows()
+    if with_llm:
+        prof["llm"] = _llm()
     return prof
 
 
@@ -185,3 +201,70 @@ def test_edge_none_y_vacio_no_rompen():
     assert ch is not None
     tables = [b for b in _flatten(ch.blocks) if isinstance(b, DataTable)]
     assert tables and len(tables[0].rows) == 3
+
+
+def _table_by_header(blocks, marker: str):
+    """Return the first DataTable whose header contains ``marker``."""
+    for b in _flatten(blocks):
+        if isinstance(b, DataTable) and marker in b.header:
+            return b
+    return None
+
+
+def test_golden_diccionario_lleva_descripcion_y_unidad_del_llm():
+    # With run_llm: the column dictionary gains "Descripción" and "Unidad"
+    # columns populated from profile['llm']['dictionary'], matched by name.
+    ch = build_overview(_profile(with_llm=True), {})
+    assert ch is not None
+    dic = _table_by_header(ch.blocks, "Descripción")
+    assert dic is not None
+    assert dic.header == ["Columna", "Tipo", "Nulos", "Ejemplos (no nulos)",
+                          "Descripción", "Unidad"]
+    by_name = {row[0]: row for row in dic.rows}
+    # PassengerId has an LLM entry -> description + unit populated.
+    assert by_name["PassengerId"][4] == "Identificador del pasajero"
+    assert by_name["PassengerId"][5] == "id"
+    assert by_name["Pclass"][5] == "clase (1-3)"
+    # Columns with no LLM entry degrade to "—" without breaking the row.
+    assert by_name["Survived"][4] == "—" and by_name["Survived"][5] == "—"
+
+
+def test_golden_describe_lleva_unidad_del_llm():
+    ch = build_overview(_profile(with_llm=True), {})
+    desc = _table_by_header(ch.blocks, "std")
+    assert desc is not None
+    assert desc.header[-1] == "Unidad"
+    by_name = {row[0]: row for row in desc.rows}
+    assert by_name["PassengerId"][-1] == "id"
+    assert by_name["Pclass"][-1] == "clase (1-3)"
+    # Numeric column with no LLM unit still renders, unit "—".
+    assert by_name["Survived"][-1] == "—"
+
+
+def test_edge_sin_llm_descripcion_unidad_son_guion():
+    # No profile['llm'] at all: the new cells degrade to "—" and nothing breaks.
+    ch = build_overview(_profile(), {})
+    assert ch is not None
+    dic = _table_by_header(ch.blocks, "Unidad")
+    assert dic is not None
+    for row in dic.rows:
+        assert row[4] == "—" and row[5] == "—"
+    desc = _table_by_header(ch.blocks, "std")
+    assert all(row[-1] == "—" for row in desc.rows)
+
+
+def test_golden_llm_via_ctx_tambien_funciona():
+    # LLM block arriving through ctx['llm'] (fallback path) is consumed too.
+    ch = build_overview(_profile(with_llm=False), {"llm": _llm()})
+    dic = _table_by_header(ch.blocks, "Descripción")
+    by_name = {row[0]: row for row in dic.rows}
+    assert by_name["PassengerId"][5] == "id"
+
+
+def test_golden_render_pdf_muestra_descripcion_y_unidad():
+    with tempfile.TemporaryDirectory() as d:
+        out = os.path.join(d, "eda.pdf")
+        render_automatic_eda_pdf(_profile(with_llm=True), out, {"title": "EDA"})
+        txt = _pdf_text(out)
+        assert "Descripción" in txt and "Unidad" in txt
+        assert "Identificador del pasajero" in txt

python/functions/datascience/automatic_eda/chapters/portada.py

@@ -26,7 +26,7 @@ from datetime import datetime, timezone
 
 from .. import model
 
-CHAPTER_VERSION = "1.2.0"
+CHAPTER_VERSION = "1.4.0"
 CHAPTER_ID = "portada"
 CHAPTER_TITLE = "Portada"
 
@@ -35,12 +35,9 @@ CHAPTER_TITLE = "Portada"
 # row represents) from it when the LLM layer ran (``run_llm``).
 _LLM_KEY = "llm"
 
-# Default human description of what the table quality score measures. Chapters
+# Font size (pt) for the dataset name on the PPTX cover slide — notably larger
-# can override it via ctx["quality_criteria"].
+# than the default H1 so the dataset name stands out (shown underlined too).
-_DEFAULT_QUALITY_CRITERIA = (
+_PPTX_TITLE_PT = 44.0
-    "media de los scores por columna (0–100): completitud (sin nulos/vacíos), "
-    "validez (tipo y rango coherentes) y consistencia (sin duplicados/constantes)."
-)
 
 
 def _storage_from_source(source: str) -> str:
@@ -120,11 +117,20 @@ def _summary_blocks(summary) -> list:
 
     blocks = [model.Heading(text="Resumen del análisis", level=2)]
     if rows:
-        blocks.append(model.KVTable(rows=rows))
+        # Values pinned to the right margin (numbers flush right, label left).
+        blocks.append(model.KVTable(rows=rows, value_align="right"))
     if titles:
-        bullets = "\n".join(f"- {model._safe_str(t)}" for t in titles)
+        # Clickable index ("Índice"): one TocEntry per chapter title. Each entry
-        blocks.append(model.Markdown(
+        # becomes a real jump to that chapter's first page/slide once the document
-            text="Este informe incluye los siguientes capítulos:\n" + bullets))
+        # is laid out (the renderers register every chapter start and wire the
+        # links; ``target_id`` is matched against the chapter title). The cover only
+        # knows chapter titles, so the title doubles as the link target.
+        blocks.append(model.Heading(text="Índice", level=2))
+        for t in titles:
+            label = model._safe_str(t)
+            if not label:
+                continue
+            blocks.append(model.TocEntry(label=label, target_id=label))
     return blocks
 
 
@@ -213,9 +219,7 @@ def _derive_description(profile: dict, ctx: dict) -> str:
     score = profile.get("quality_score")
     if score is not None:
         parts.append(f"Calidad media estimada: {score}/100.")
-    parts.append(
+    parts.append("Resumen derivado del perfil.")
-        "Resumen derivado del perfil; active la interpretación LLM (`run_llm`) "
-        "para una descripción de negocio más rica.")
     return " ".join(parts)
 
 
@@ -259,7 +263,6 @@ def build_portada(profile: dict, ctx: dict):
     shape = f"{_fmt_int(n_rows)} filas × {_fmt_int(n_cols)} columnas"
 
     score = profile.get("quality_score")
-    quality_criteria = ctx.get("quality_criteria") or _DEFAULT_QUALITY_CRITERIA
     quality_value = "—" if score is None else f"{score} / 100"
 
     llm = _llm_block(profile, ctx)
@@ -282,8 +285,11 @@ def build_portada(profile: dict, ctx: dict):
 
     # Title + dataset size shown together and BIG (Heading) at the top, kept on
     # the same page (Group). The size is no longer buried in the metadata table.
+    # The dataset name is shown big and underlined on the PPTX cover slide
+    # (size_pt/underline are honoured by the PPTX renderer; the PDF ignores them).
     cover = [
-        model.Heading(text=str(dataset_name), level=1),
+        model.Heading(text=str(dataset_name), level=1, underline=True,
+                      size_pt=_PPTX_TITLE_PT),
         model.Markdown(text="**Automatic-EDA** · informe exploratorio automático"),
         model.Heading(text=shape, level=2),
     ]
@@ -295,7 +301,6 @@ def build_portada(profile: dict, ctx: dict):
             ("Almacenamiento", storage),
             ("Generado", when),
             ("Calidad", quality_value),
-            ("Criterios de calidad", quality_criteria),
         ]),
         model.Heading(text="Descripción", level=2),
         model.Markdown(text=str(description)),

python/functions/datascience/automatic_eda/chapters/text_distr.py

@@ -0,0 +1,559 @@
+"""Free-text / NLP distributions chapter (TEXT DISTR) for AutomaticEDA.
+
+First chapter for **non-tabular** content: it profiles the linguistic content of
+any column holding long free text (reviews, descriptions, comments, tickets) that
+the categorical chapter cannot meaningfully summarize (high cardinality, many
+words per value). It is the cheap, model-free counterpart to ``cat_distr`` for
+columns that are prose rather than discrete labels.
+
+Activation (returns ``None`` when it does not apply):
+
+1. Cheap gate from the aggregated profile: at least one non-numeric column whose
+   ``categorical.len_mean`` (mean character length) is ``>= _MIN_LEN_CHARS``.
+   A dataset whose only string columns are short labels (e.g. titanic's
+   ``Name``, ~27 chars) never passes this gate, so the chapter disappears with
+   zero extra work and the existing report is untouched.
+2. Confirmation from a raw sample: each candidate column is sampled (push-down
+   ``extract_text_sample`` over ``ctx['db_path']``/``ctx['table']``, or an
+   in-memory ``ctx['text_raw']`` for tests) and kept only if the **median word
+   count is ``>= _MIN_WORDS``** — i.e. it is genuinely long text, not a long
+   single token. If no column survives, the chapter returns ``None``.
+
+Per surviving column the chapter emits, kept together on its own page/slide
+(``Group(page_break_before=...)``):
+
+- a key/value summary (documents, length percentiles, vocabulary richness with
+  **[[term:ttr]]TTR[[/term]]** and **[[term:hapax]]hapax legomena[[/term]]**,
+  dominant language, exact-duplicate %, readability when available);
+- a word-count histogram figure;
+- a top-terms table + a horizontal bar figure;
+- bigram and trigram frequency tables;
+- a detected-language bar figure (when ``langdetect`` is available);
+- an optional word-cloud figure (only when ``wordcloud`` is installed);
+- a closing note on duplicates / readability degradation.
+
+Every metric is delegated to pure ``eda`` registry functions
+(``compute_text_length_stats``, ``compute_vocabulary_stats``,
+``compute_top_ngrams``, ``detect_corpus_language``, ``compute_text_duplicates``,
+``compute_text_readability``) and the raw sample to ``extract_text_sample``; all
+are imported defensively so a missing function or optional library degrades that
+single piece to a note instead of aborting the chapter. Optional libraries
+(``langdetect``, ``textstat``, ``wordcloud``, ``datasketch``) are never required:
+the piece is silently omitted when they are absent.
+
+Contract: build_<id>(profile, ctx) -> Chapter | None ; CHAPTER_VERSION = "x.y.z".
+"""
+
+from __future__ import annotations
+
+from .. import model
+
+CHAPTER_VERSION = "1.0.0"
+CHAPTER_ID = "text_distr"
+CHAPTER_TITLE = "Texto libre (NLP)"
+
+# Cheap activation gate (characters): a non-numeric column whose mean string
+# length reaches this is a candidate for "long text". Short labels (titanic's
+# Name ≈ 27 chars) stay below it, so the chapter does not fire on them.
+_MIN_LEN_CHARS = 50
+# Confirmation gate (words): a candidate is kept only if its median document has
+# at least this many words — genuine prose, not a long id/URL token.
+_MIN_WORDS = 20
+# Bound the document so very wide datasets stay readable.
+_MAX_TEXT_COLS = 5
+# Raw text rows to sample per column when the chapter must extract them itself.
+_SAMPLE_ROWS = 2000
+# Rows shown in the frequency tables.
+_TOP_TERMS = 15
+_TOP_NGRAMS = 10
+
+# Glossary terms this chapter explains (registered in the shared collector and
+# marked clickable on first appearance — same mechanism as cat_distr's entropía).
+_TERMS = {
+    "ttr": (
+        "TTR (type-token ratio)",
+        "Riqueza léxica de un texto: número de palabras distintas (tipos) "
+        "dividido por el número total de palabras (tokens). Vale 1 cuando no se "
+        "repite ninguna palabra (máxima variedad) y baja hacia 0 cuando el "
+        "vocabulario se repite mucho. Depende de la longitud del corpus, así que "
+        "compara mejor textos de tamaño parecido."),
+    "hapax": (
+        "Hapax legomena",
+        "Palabras que aparecen una sola vez en todo el corpus. Un porcentaje "
+        "alto de hapax indica vocabulario muy variado o, a veces, ruido "
+        "(erratas, identificadores, tokens raros). Se expresa como porcentaje "
+        "sobre el número de palabras distintas."),
+}
+
+
+def _fmt_int(value) -> str:
+    if value is None:
+        return "—"
+    try:
+        return f"{int(value):,}".replace(",", ".")
+    except (TypeError, ValueError):
+        return str(value)
+
+
+def _fmt_num(value, decimals: int = 2) -> str:
+    if value is None:
+        return "—"
+    if isinstance(value, bool):
+        return str(value)
+    if isinstance(value, int):
+        return f"{value:,}".replace(",", ".")
+    if isinstance(value, float):
+        if value != value:  # NaN
+            return "NaN"
+        if value in (float("inf"), float("-inf")):
+            return str(value)
+        text = f"{value:.{decimals}f}".rstrip("0").rstrip(".")
+        return text if text else "0"
+    return str(value)
+
+
+def _fmt_pct(value, decimals: int = 1) -> str:
+    if value is None:
+        return "—"
+    try:
+        return f"{float(value):.{decimals}f}%"
+    except (TypeError, ValueError):
+        return str(value)
+
+
+def _truncate(text, limit: int = 40) -> str:
+    s = model._safe_str(text)
+    return s if len(s) <= limit else s[: max(1, limit - 1)].rstrip() + "…"
+
+
+# --------------------------------------------------------------------------- #
+# Defensive wrappers around the registry functions: each returns the function's
+# output dict or a safe empty default, never raising and never importing at
+# module load (so the chapter stays importable even if a function is missing).
+# --------------------------------------------------------------------------- #
+def _length_stats(texts) -> dict:
+    try:
+        from datascience.compute_text_length_stats import compute_text_length_stats
+        out = compute_text_length_stats(texts)
+        if isinstance(out, dict):
+            return out
+    except Exception:  # noqa: BLE001
+        pass
+    return {}
+
+
+def _vocab_stats(texts) -> dict:
+    try:
+        from datascience.compute_vocabulary_stats import compute_vocabulary_stats
+        out = compute_vocabulary_stats(texts, top_k=_TOP_TERMS)
+        if isinstance(out, dict):
+            return out
+    except Exception:  # noqa: BLE001
+        pass
+    return {}
+
+
+def _ngrams(texts, n) -> list:
+    try:
+        from datascience.compute_top_ngrams import compute_top_ngrams
+        out = compute_top_ngrams(texts, n=n, top_k=_TOP_NGRAMS)
+        if isinstance(out, dict):
+            return out.get("top") or []
+    except Exception:  # noqa: BLE001
+        pass
+    return []
+
+
+def _language(texts) -> dict:
+    try:
+        from datascience.detect_corpus_language import detect_corpus_language
+        out = detect_corpus_language(texts)
+        if isinstance(out, dict):
+            return out
+    except Exception:  # noqa: BLE001
+        pass
+    return {"available": False, "distribution": [], "dominant": None}
+
+
+def _duplicates(texts) -> dict:
+    try:
+        from datascience.compute_text_duplicates import compute_text_duplicates
+        out = compute_text_duplicates(texts)
+        if isinstance(out, dict):
+            return out
+    except Exception:  # noqa: BLE001
+        pass
+    return {}
+
+
+def _readability(texts) -> dict:
+    try:
+        from datascience.compute_text_readability import compute_text_readability
+        out = compute_text_readability(texts)
+        if isinstance(out, dict):
+            return out
+    except Exception:  # noqa: BLE001
+        pass
+    return {"available": False, "flesch": {}}
+
+
+# --------------------------------------------------------------------------- #
+# Candidate detection + raw sample acquisition.
+# --------------------------------------------------------------------------- #
+def _candidate_columns(profile: dict) -> list:
+    """Cheap gate: non-numeric columns whose mean char length reaches the
+    threshold. Returns the list of column names (possibly empty)."""
+    out = []
+    for col in profile.get("columns") or []:
+        if not isinstance(col, dict):
+            continue
+        if col.get("inferred_type") == "numeric":
+            continue
+        cat = col.get("categorical")
+        if not isinstance(cat, dict):
+            continue
+        len_mean = cat.get("len_mean")
+        if isinstance(len_mean, (int, float)) and not isinstance(len_mean, bool) \
+                and len_mean >= _MIN_LEN_CHARS:
+            name = col.get("name")
+            if name:
+                out.append(str(name))
+    return out
+
+
+def _get_samples(profile: dict, ctx: dict, columns: list) -> dict:
+    """Return {col: [str, ...]} raw text samples for the candidate columns.
+
+    Prefers an in-memory ``ctx['text_raw']`` (used by tests); otherwise pushes a
+    sample down to the database via ``extract_text_sample`` using ctx db_path /
+    table. Never raises: returns {} when no sample can be obtained."""
+    text_raw = ctx.get("text_raw")
+    if isinstance(text_raw, dict) and text_raw:
+        return {c: [str(v) for v in (text_raw.get(c) or []) if v is not None]
+                for c in columns if text_raw.get(c)}
+
+    db_path = ctx.get("db_path")
+    table = ctx.get("table")
+    if not db_path or not table:
+        return {}
+    backend = ctx.get("backend") or "duckdb"
+    sample = ctx.get("sample") or _SAMPLE_ROWS
+    try:
+        from datascience.extract_text_sample import extract_text_sample
+        out = extract_text_sample(db_path, table, columns, backend=backend,
+                                  sample=sample)
+        if isinstance(out, dict) and out.get("status") == "ok":
+            cols = out.get("columns")
+            if isinstance(cols, dict):
+                return {c: list(v) for c, v in cols.items() if v}
+    except Exception:  # noqa: BLE001 — dict-no-throw: no sample → chapter omits.
+        pass
+    return {}
+
+
+def _confirm_long_text(samples: dict) -> dict:
+    """Keep only columns whose median word count reaches _MIN_WORDS. Returns
+    {col: length_stats_dict} for the survivors, in input order."""
+    survivors = {}
+    for col, texts in samples.items():
+        stats = _length_stats(texts)
+        words = stats.get("words") if isinstance(stats, dict) else None
+        median = words.get("p50") if isinstance(words, dict) else None
+        if isinstance(median, (int, float)) and not isinstance(median, bool) \
+                and median >= _MIN_WORDS:
+            survivors[col] = stats
+    return survivors
+
+
+# --------------------------------------------------------------------------- #
+# Figures (lazy matplotlib, scaled by the renderers — same style as num_distr).
+# --------------------------------------------------------------------------- #
+def _hist_figure(name: str, length_stats: dict):
+    def make():
+        import matplotlib
+        matplotlib.use("Agg")
+        from matplotlib.figure import Figure
+        fig = Figure(figsize=(6.2, 3.0))
+        ax = fig.add_subplot(111)
+        bins = (length_stats or {}).get("word_hist") or []
+        drew = False
+        for b in bins:
+            if not isinstance(b, dict):
+                continue
+            lo, hi, count = b.get("lo"), b.get("hi"), b.get("count") or 0
+            if lo is None or hi is None:
+                continue
+            width = (hi - lo) if hi > lo else max(abs(lo) * 1e-3, 1e-6)
+            ax.bar(lo, count, width=width, align="edge", color="#9ec6df",
+                   edgecolor="#5b8aa6", linewidth=0.4)
+            drew = True
+        if not drew:
+            ax.text(0.5, 0.5, "(sin datos de longitud)", ha="center",
+                    va="center", color="#8a8a8a", transform=ax.transAxes)
+        ax.set_xlabel("palabras por documento", fontsize=8)
+        ax.set_ylabel("nº de documentos", fontsize=8)
+        ax.tick_params(labelsize=7)
+        for spine in ("top", "right"):
+            ax.spines[spine].set_visible(False)
+        ax.set_title(f"Longitud de «{_truncate(name, 30)}»", fontsize=10,
+                     loc="left")
+        fig.tight_layout()
+        return fig
+    return make
+
+
+def _barh_figure(title: str, items: list, label_key: str, value_key: str,
+                 xlabel: str):
+    """Horizontal bar chart from [{label_key:..., value_key:...}, ...]."""
+    def make():
+        import matplotlib
+        matplotlib.use("Agg")
+        from matplotlib.figure import Figure
+        rows = [it for it in (items or []) if isinstance(it, dict)
+                and isinstance(it.get(value_key), (int, float))]
+        rows = rows[:12]
+        fig = Figure(figsize=(6.2, max(2.2, 0.32 * len(rows) + 0.8)))
+        ax = fig.add_subplot(111)
+        if not rows:
+            ax.text(0.5, 0.5, "(sin datos)", ha="center", va="center",
+                    color="#8a8a8a", transform=ax.transAxes)
+            ax.axis("off")
+            return fig
+        labels = [_truncate(r.get(label_key), 28) for r in rows][::-1]
+        values = [float(r.get(value_key) or 0) for r in rows][::-1]
+        ypos = range(len(rows))
+        ax.barh(list(ypos), values, color="#9ec6df", edgecolor="#5b8aa6",
+                linewidth=0.4)
+        ax.set_yticks(list(ypos))
+        ax.set_yticklabels(labels, fontsize=7)
+        ax.set_xlabel(xlabel, fontsize=8)
+        ax.tick_params(labelsize=7)
+        for spine in ("top", "right"):
+            ax.spines[spine].set_visible(False)
+        ax.set_title(_truncate(title, 44), fontsize=10, loc="left")
+        fig.tight_layout()
+        return fig
+    return make
+
+
+def _wordcloud_figure(texts):
+    """Word-cloud figure callable, or None if wordcloud is not installed."""
+    try:
+        import wordcloud  # noqa: F401
+    except Exception:  # noqa: BLE001 — optional dependency: omit the figure.
+        return None
+
+    def make():
+        import matplotlib
+        matplotlib.use("Agg")
+        from matplotlib.figure import Figure
+        from wordcloud import WordCloud
+        fig = Figure(figsize=(6.2, 3.2))
+        ax = fig.add_subplot(111)
+        joined = " ".join(t for t in texts if isinstance(t, str))
+        try:
+            wc = WordCloud(width=800, height=400, background_color="white",
+                           colormap="viridis").generate(joined)
+            ax.imshow(wc, interpolation="bilinear")
+        except Exception:  # noqa: BLE001
+            ax.text(0.5, 0.5, "(nube de palabras no disponible)", ha="center",
+                    va="center", color="#8a8a8a", transform=ax.transAxes)
+        ax.axis("off")
+        fig.tight_layout()
+        return fig
+    return make
+
+
+# --------------------------------------------------------------------------- #
+# Per-column block assembly.
+# --------------------------------------------------------------------------- #
+def _summary_kv(n_docs, length_stats, vocab, lang, dup, read):
+    chars = (length_stats or {}).get("chars") or {}
+    words = (length_stats or {}).get("words") or {}
+    sents = (length_stats or {}).get("sentences") or {}
+    rows = [
+        ("Documentos", _fmt_int(n_docs)),
+        ("Caracteres (media · p50 · p90 · p99)",
+         f"{_fmt_num(chars.get('mean'))} · {_fmt_int(chars.get('p50'))} · "
+         f"{_fmt_int(chars.get('p90'))} · {_fmt_int(chars.get('p99'))}"),
+        ("Palabras (media · p50 · p90 · p99)",
+         f"{_fmt_num(words.get('mean'))} · {_fmt_int(words.get('p50'))} · "
+         f"{_fmt_int(words.get('p90'))} · {_fmt_int(words.get('p99'))}"),
+        ("Frases (media · máx)",
+         f"{_fmt_num(sents.get('mean'))} · {_fmt_int(sents.get('max'))}"),
+        ("Vocabulario (tokens · tipos · TTR)",
+         f"{_fmt_int(vocab.get('n_tokens'))} · {_fmt_int(vocab.get('n_types'))} "
+         f"· {_fmt_num(vocab.get('ttr'), 3)}"),
+        ("Hapax legomena",
+         f"{_fmt_int(vocab.get('n_hapax'))} ({_fmt_pct(vocab.get('hapax_pct'))})"),
+    ]
+    if isinstance(lang, dict) and lang.get("available"):
+        dom = lang.get("dominant")
+        n_langs = len(lang.get("distribution") or [])
+        rows.append(("Idioma dominante · nº idiomas",
+                     f"{model._safe_str(dom) or '—'} · {_fmt_int(n_langs)}"))
+    if isinstance(dup, dict) and dup.get("n_docs"):
+        rows.append(("Duplicados exactos",
+                     f"{_fmt_int(dup.get('n_exact_dup'))} "
+                     f"({_fmt_pct(dup.get('exact_dup_pct'))})"))
+    if isinstance(read, dict) and read.get("available"):
+        flesch = read.get("flesch") or {}
+        rows.append(("Legibilidad Flesch (media)",
+                     _fmt_num(flesch.get("mean"), 1)))
+    return model.KVTable(rows=rows, title="Resumen del texto")
+
+
+def _terms_table(vocab) -> "model.DataTable | None":
+    top = (vocab or {}).get("top_terms") or []
+    rows = [[_truncate(t.get("term"), 32), _fmt_int(t.get("count")),
+             _fmt_pct(t.get("pct"))]
+            for t in top[:_TOP_TERMS] if isinstance(t, dict)]
+    if not rows:
+        return None
+    return model.DataTable(header=["Término", "Conteo", "% tokens"], rows=rows,
+                           title="Términos más frecuentes",
+                           note="stopwords ES+EN eliminadas")
+
+
+def _ngram_table(items, n_label) -> "model.DataTable | None":
+    rows = [[_truncate(it.get("ngram"), 40), _fmt_int(it.get("count"))]
+            for it in (items or [])[:_TOP_NGRAMS] if isinstance(it, dict)]
+    if not rows:
+        return None
+    return model.DataTable(header=[n_label, "Conteo"], rows=rows,
+                           title=f"{n_label} más frecuentes")
+
+
+def _dup_note(dup, lang, read) -> "model.Note | None":
+    bits = []
+    if isinstance(dup, dict):
+        nd = dup.get("near_dup") or {}
+        if nd.get("available"):
+            bits.append(
+                f"casi-duplicados detectados (MinHash, umbral "
+                f"{_fmt_num(nd.get('threshold'))}): "
+                f"{_fmt_int(nd.get('n_near_dup_docs'))} documentos")
+        else:
+            bits.append("near-duplicados no calculados (datasketch no instalado; "
+                        "se reportan solo los duplicados exactos por hash)")
+    if isinstance(lang, dict) and not lang.get("available"):
+        bits.append("detección de idioma omitida (langdetect no instalado)")
+    if isinstance(read, dict) and not read.get("available"):
+        bits.append("legibilidad omitida (textstat no instalado)")
+    if not bits:
+        return None
+    return model.Note(" · ".join(bits))
+
+
+def _column_group(name, texts, length_stats, idx, mark_terms):
+    vocab = _vocab_stats(texts)
+    lang = _language(texts)
+    dup = _duplicates(texts)
+    read = _readability(texts)
+    n_docs = (length_stats or {}).get("n_docs")
+
+    blocks = [
+        model.Heading(text=str(name), level=2),
+        _summary_kv(n_docs, length_stats, vocab, lang, dup, read),
+        model.Figure(make=_hist_figure(name, length_stats),
+                     caption=f"Distribución de la longitud (palabras) de "
+                             f"«{_truncate(name, 30)}»."),
+    ]
+
+    terms_tbl = _terms_table(vocab)
+    if terms_tbl is not None:
+        blocks.append(terms_tbl)
+        blocks.append(model.Figure(
+            make=_barh_figure(f"Top términos de «{_truncate(name, 24)}»",
+                              vocab.get("top_terms"), "term", "count",
+                              "conteo"),
+            caption="Términos más frecuentes (barras)."))
+
+    bi_tbl = _ngram_table(_ngrams(texts, 2), "Bigrama")
+    if bi_tbl is not None:
+        blocks.append(bi_tbl)
+    tri_tbl = _ngram_table(_ngrams(texts, 3), "Trigrama")
+    if tri_tbl is not None:
+        blocks.append(tri_tbl)
+
+    if isinstance(lang, dict) and lang.get("available") \
+            and lang.get("distribution"):
+        blocks.append(model.Figure(
+            make=_barh_figure(f"Idiomas detectados en «{_truncate(name, 24)}»",
+                              lang.get("distribution"), "lang", "count",
+                              "documentos"),
+            caption="Distribución de idiomas detectados (langdetect)."))
+
+    wc = _wordcloud_figure(texts)
+    if wc is not None:
+        blocks.append(model.Figure(
+            make=wc, caption=f"Nube de palabras de «{_truncate(name, 30)}»."))
+
+    note = _dup_note(dup, lang, read)
+    if note is not None:
+        blocks.append(note)
+
+    return model.Group(blocks=blocks, page_break_before=(idx > 0))
+
+
+def _intro_blocks(n_cols, mark_terms):
+    ttr = ("[[term:ttr]]TTR[[/term]]" if mark_terms else "TTR")
+    hapax = ("[[term:hapax]]hapax legomena[[/term]]" if mark_terms
+             else "hapax legomena")
+    text = (
+        f"Este capítulo perfila las columnas de **texto libre largo** del "
+        f"dataset (reseñas, descripciones, comentarios): contenido lingüístico "
+        f"que la distribución categórica no resume bien. Para cada columna se "
+        f"muestran la longitud de los documentos, la riqueza de vocabulario "
+        f"(incluido el {ttr} y el porcentaje de {hapax}), los términos y "
+        f"n-gramas más frecuentes, los idiomas detectados y el nivel de "
+        f"duplicación. Las métricas son baratas y sin modelos pesados; las "
+        f"piezas que dependen de una librería opcional se omiten si no está "
+        f"instalada.")
+    return [
+        model.Heading(text=CHAPTER_TITLE, level=1),
+        model.Markdown(text=text),
+    ]
+
+
+def build_text_distr(profile: dict, ctx: dict):
+    """Build the free-text Chapter, or None if no long-text column applies."""
+    profile = profile or {}
+    ctx = ctx or {}
+
+    # 1) Cheap gate from the profile (no DB access yet).
+    candidates = _candidate_columns(profile)
+    if not candidates:
+        return None
+
+    # 2) Raw sample + 3) confirm genuine long text (median words >= threshold).
+    samples = _get_samples(profile, ctx, candidates)
+    if not samples:
+        return None
+    survivors = _confirm_long_text(samples)
+    if not survivors:
+        return None
+
+    # Register glossary terms (clickable) once we know the chapter applies.
+    glossary = ctx.get("glossary")
+    mark_terms = False
+    if isinstance(glossary, model.GlossaryCollector):
+        for key, (label, definition) in _TERMS.items():
+            glossary.add(key, label, definition)
+        mark_terms = True
+
+    blocks = list(_intro_blocks(len(survivors), mark_terms))
+
+    rendered = list(survivors.items())[:_MAX_TEXT_COLS]
+    for idx, (name, length_stats) in enumerate(rendered):
+        texts = samples.get(name) or []
+        blocks.append(_column_group(name, texts, length_stats, idx, mark_terms))
+
+    if len(survivors) > len(rendered):
+        omitted = len(survivors) - len(rendered)
+        blocks.append(model.Note(
+            f"Se muestran las primeras {len(rendered)} columnas de texto; "
+            f"quedan {omitted} sin mostrar para mantener acotado el informe."))
+
+    return model.Chapter(id=CHAPTER_ID, title=CHAPTER_TITLE,
+                         version=CHAPTER_VERSION, blocks=blocks)

python/functions/datascience/automatic_eda/chapters/text_distr_test.py

@@ -0,0 +1,256 @@
+"""Tests for the TEXT DISTR chapter — DoD: golden + edges + degradation.
+
+Self-contained: builds synthetic TableProfiles and feeds the raw text sample
+in-memory through ``ctx['text_raw']`` (no DuckDB needed), so the suite is fast
+and deterministic. Verifies that ``build_text_distr``:
+
+- GOLDEN: with a long-text column, emits the chapter with its key blocks
+  (length summary, word histogram, top-terms table, n-gram tables, language
+  bars) and registers the clickable glossary terms; and that it renders inside
+  the full document to both PDF and PPTX showing that content.
+- EDGE (None): a dataset whose only string column is short labels (titanic-like
+  ``Name``) yields ``None`` without raising — the existing report is untouched.
+- EDGE (None): a column that passes the cheap char gate but whose documents are
+  short (median words below the threshold) is rejected at the confirmation step.
+- DEGRADATION: with ``langdetect`` / ``textstat`` / ``wordcloud`` unavailable,
+  the chapter still builds (those pieces are omitted) and never raises.
+"""
+
+import builtins
+import os
+import tempfile
+
+from pypdf import PdfReader
+from pptx import Presentation
+
+from datascience.automatic_eda.model import (
+    DataTable, Figure, GlossaryCollector, Group, Heading, KVTable, Markdown,
+    Note,
+)
+from datascience.automatic_eda.chapters.text_distr import (
+    CHAPTER_ID, CHAPTER_VERSION, build_text_distr,
+)
+from datascience.automatic_eda.chapters_registry import build_document
+from datascience.render_automatic_eda_pdf import render_automatic_eda_pdf
+from datascience.render_automatic_eda_pptx import render_automatic_eda_pptx
+
+
+# --------------------------------------------------------------------------- #
+# Synthetic corpus + profiles.
+# --------------------------------------------------------------------------- #
+_ES = [
+    "El producto llegó en perfecto estado y mucho antes de lo previsto por la tienda",
+    "La calidad de los materiales es realmente excelente y se nota la diferencia al usarlo",
+    "No me convenció del todo porque esperaba bastante más por el precio que pagué finalmente",
+    "El servicio de atención al cliente fue rápido amable y resolvió mi problema sin demora",
+    "Lo recomiendo totalmente ya que ha superado con creces todas mis expectativas iniciales",
+]
+_EN = [
+    "The product arrived in perfect condition and much earlier than the store had promised me",
+    "The build quality is genuinely outstanding and you can really feel the difference using it",
+    "I was not fully convinced because I expected quite a lot more for the price i finally paid",
+    "Customer support was fast friendly and solved my whole problem without any delay at all",
+    "I highly recommend it since it has exceeded by far every one of my initial expectations",
+]
+
+
+def _long_reviews(n=40) -> list:
+    """A corpus of long multi-sentence reviews (>= 20 words each), mixing two
+    languages and including a few exact duplicates."""
+    out = []
+    for i in range(n):
+        base = _ES if i % 3 != 0 else _EN  # mostly ES, some EN
+        a = base[i % len(base)]
+        b = base[(i + 2) % len(base)]
+        out.append(f"{a}. {b}.")
+    # Inject a couple of exact duplicates.
+    out.append(out[0])
+    out.append(out[1])
+    return out
+
+
+def _text_profile() -> dict:
+    """Profile with a long free-text column (review) + a numeric + a short cat."""
+    return {
+        "table": "reviews",
+        "source": "/data/reviews.duckdb",
+        "profiled_at": "2026-06-30T10:00:00+00:00",
+        "n_rows": 42,
+        "n_cols": 3,
+        "quality_score": 88.0,
+        "columns": [
+            {
+                "name": "review",
+                "inferred_type": "categorical",
+                "categorical": {
+                    "top": [{"value": "x", "count": 2, "pct": 0.05}],
+                    "n_distinct": 40,
+                    "len_mean": 180.0,
+                    "len_min": 80,
+                    "len_max": 220,
+                },
+            },
+            {
+                "name": "rating",
+                "inferred_type": "numeric",
+                "numeric": {"mean": 3.1, "median": 3.0, "std": 1.2,
+                            "min": 1, "max": 5},
+            },
+            {
+                "name": "product",
+                "inferred_type": "categorical",
+                "categorical": {
+                    "top": [{"value": "teclado", "count": 10, "pct": 0.25}],
+                    "n_distinct": 6,
+                    "len_mean": 7.0,
+                    "len_min": 5, "len_max": 11,
+                },
+            },
+        ],
+    }
+
+
+def _no_text_profile() -> dict:
+    """titanic-like: the only string column is short labels (Name ≈ 27 chars)."""
+    return {
+        "table": "titanic",
+        "n_rows": 891,
+        "n_cols": 3,
+        "columns": [
+            {"name": "Age", "inferred_type": "numeric",
+             "numeric": {"mean": 29.7, "median": 28.0, "std": 14.5}},
+            {"name": "Name", "inferred_type": "categorical",
+             "categorical": {"top": [{"value": "Braund, Mr. Owen Harris",
+                                      "count": 1, "pct": 0.001}],
+                             "n_distinct": 891, "len_mean": 27.0,
+                             "len_min": 12, "len_max": 82}},
+            {"name": "Sex", "inferred_type": "categorical",
+             "categorical": {"top": [{"value": "male", "count": 577,
+                                      "pct": 0.65}],
+                             "n_distinct": 2, "len_mean": 4.6,
+                             "len_min": 4, "len_max": 6}},
+        ],
+    }
+
+
+def _flatten(blocks) -> list:
+    """Recursively flatten Group blocks so tests can inspect leaf blocks."""
+    out = []
+    for b in blocks:
+        if isinstance(b, Group):
+            out.extend(_flatten(b.blocks))
+        else:
+            out.append(b)
+    return out
+
+
+# --------------------------------------------------------------------------- #
+# Golden.
+# --------------------------------------------------------------------------- #
+def test_golden_activa_con_texto():
+    glossary = GlossaryCollector()
+    ctx = {"text_raw": {"review": _long_reviews()}, "glossary": glossary}
+    ch = build_text_distr(_text_profile(), ctx)
+
+    assert ch is not None, "el capítulo debe activarse con una columna de texto largo"
+    assert ch.id == CHAPTER_ID
+    assert ch.version == CHAPTER_VERSION
+    leaves = _flatten(ch.blocks)
+    kinds = [b.kind for b in leaves]
+    assert "heading" in kinds
+    assert "kv_table" in kinds          # summary
+    assert "figure" in kinds            # histogram / bars
+    assert "data_table" in kinds        # top terms + n-grams
+
+    # KV summary mentions vocabulary metrics.
+    kv = next(b for b in leaves if isinstance(b, KVTable))
+    labels = " ".join(str(r[0]) for r in kv.rows)
+    assert "TTR" in labels
+    assert "Hapax" in labels or "hapax" in labels
+
+    # There is a terms table and at least one n-gram table.
+    titles = [getattr(b, "title", "") or "" for b in leaves
+              if isinstance(b, DataTable)]
+    assert any("Términos" in t for t in titles)
+    assert any("Bigrama" in t for t in titles)
+
+    # Glossary terms were registered (clickable destinations).
+    assert glossary.has("ttr")
+    assert glossary.has("hapax")
+
+
+def test_golden_render_pdf_pptx():
+    profile = _text_profile()
+    ctx = {"text_raw": {"review": _long_reviews()},
+           "dataset_name": "reviews"}
+    chapters = build_document(profile, ctx)
+    ids = [c.id for c in chapters]
+    assert "text_distr" in ids, f"text_distr ausente en {ids}"
+
+    with tempfile.TemporaryDirectory() as d:
+        pdf = os.path.join(d, "t.pdf")
+        pptx = os.path.join(d, "t.pptx")
+        rp = render_automatic_eda_pdf(profile, pdf, {"title": "EDA", "ctx": ctx})
+        rx = render_automatic_eda_pptx(profile, pptx, {"title": "EDA", "ctx": ctx})
+        assert rp.get("path") and os.path.exists(pdf)
+        assert rx.get("path") and os.path.exists(pptx)
+
+        text = "\n".join(p.extract_text() or "" for p in PdfReader(pdf).pages)
+        assert "Texto libre" in text or "TTR" in text
+
+        prs = Presentation(pptx)
+        ptext = []
+        for slide in prs.slides:
+            for shp in slide.shapes:
+                if shp.has_text_frame:
+                    ptext.append(shp.text_frame.text)
+        joined = "\n".join(ptext)
+        assert "Texto libre" in joined or "TTR" in joined
+
+
+# --------------------------------------------------------------------------- #
+# Edges — None.
+# --------------------------------------------------------------------------- #
+def test_edge_none_sin_texto_largo():
+    # titanic-like: short labels only → chapter must not apply.
+    assert build_text_distr(_no_text_profile(), {}) is None
+
+
+def test_edge_none_palabras_cortas():
+    # Char gate passes (len_mean high) but documents are short → confirmation
+    # rejects them (median words below threshold).
+    profile = _text_profile()
+    short = ["palabra " * 3] * 30  # 3 words each, < _MIN_WORDS
+    ctx = {"text_raw": {"review": short}}
+    assert build_text_distr(profile, ctx) is None
+
+
+def test_edge_none_empty_profile():
+    assert build_text_distr({}, {}) is None
+    assert build_text_distr(None, None) is None
+
+
+# --------------------------------------------------------------------------- #
+# Degradation — optional libs absent.
+# --------------------------------------------------------------------------- #
+def test_degradacion_sin_libs(monkeypatch):
+    real_import = builtins.__import__
+    blocked = ("langdetect", "textstat", "wordcloud", "datasketch")
+
+    def fake_import(name, *a, **k):
+        if name in blocked or any(name.startswith(b + ".") for b in blocked):
+            raise ImportError(f"simulado: {name}")
+        return real_import(name, *a, **k)
+
+    monkeypatch.setattr(builtins, "__import__", fake_import)
+
+    ctx = {"text_raw": {"review": _long_reviews()}}
+    ch = build_text_distr(_text_profile(), ctx)
+    # Still builds (the cheap, stdlib-only pieces remain) and never raises.
+    assert ch is not None
+    leaves = _flatten(ch.blocks)
+    assert any(isinstance(b, KVTable) for b in leaves)
+    assert any(isinstance(b, DataTable) for b in leaves)
+    # A degradation note is present mentioning the missing optional libs.
+    notes = " ".join(b.text for b in leaves if isinstance(b, Note))
+    assert "langdetect" in notes or "textstat" in notes or "datasketch" in notes

python/functions/datascience/automatic_eda/chapters_registry.py

@@ -31,7 +31,10 @@ CHAPTER_ORDER = [
     "analisis_llm",  # LLM interpretation — sits next to overview (user request)
     "num_distr",     # numeric distributions
     "cat_distr",     # categorical distributions
+    "text_distr",    # free-text / NLP distributions (non-tabular content)
     "calidad",       # data quality
+    "missingness",   # missing-data patterns (co-occurrence of absences; MCAR/MAR)
+    "outliers",      # atypical values: univariate (Tukey/z) + multivariate (IsolationForest)
     "correlacion",   # correlations / associations
     "relaciones",    # key relations: declared/candidate PK + FK (inter/intra-table)
     "modelos",       # cheap models (PCA/KMeans/outliers)
@@ -70,24 +73,51 @@ def build_chapter(chapter_id: str, profile: dict, ctx: dict):
     return model.as_chapter(result)
 
 
-def build_document(profile: dict, ctx: dict = None) -> list:
+def build_document(profile: dict, ctx: dict = None, only: list = None) -> list:
-    """Build the full ordered list of chapters for a TableProfile.
+    """Build the ordered list of chapters for a TableProfile.
 
     Args:
         profile: the ``eda`` group TableProfile dict (may be None/empty).
         ctx: optional context dict carrying presentation metadata not present in
             the profile (dataset_name, source_origin, storage, generated_at,
             description, granularity, quality_criteria, head_rows, ...).
+        only: optional list of chapter ids to render. ``None`` (default) keeps
+            the historical behaviour — every implemented & applicable chapter in
+            canonical order. A list restricts the BODY to just those ids (in
+            canonical order), but the cover (``portada``) and glossary
+            (``glosario``) are ALWAYS included so the document stays valid and
+            the clickable terms keep a destination — so passing ``only=["x"]``
+            yields portada + x + glosario. Unknown ids are simply skipped (the
+            caller is responsible for strict validation). ``only=[]`` yields the
+            minimal document (portada + glosario only). This argument is additive
+            and backward-compatible: the signature is unchanged for existing
+            callers (default ``None``).
 
     Returns:
         list[Chapter] in canonical order, containing only the chapters that are
-        implemented and applicable. Never raises.
+        implemented, applicable and selected. Never raises.
     """
     if not isinstance(profile, dict):
         profile = {}
     # Copy ctx so the shared collector / summary we add do not leak to the caller.
     ctx = dict(ctx) if isinstance(ctx, dict) else {}
 
+    # only=None -> all body chapters (historical). only=list -> restrict body to
+    # that selection (portada/glosario are added unconditionally below). The
+    # renderers call build_document(profile, meta['ctx']) without an `only`
+    # argument, so the pipeline forwards the selection through a reserved ctx key
+    # (``_only_chapters``); an explicit `only` argument always wins. The key is
+    # popped from the local ctx copy so it never reaches the chapters.
+    if only is None:
+        _carried = ctx.pop("_only_chapters", None)
+        if isinstance(_carried, (list, tuple, set)):
+            only = list(_carried)
+    else:
+        ctx.pop("_only_chapters", None)
+    # A set makes the membership test cheap; the iteration order stays
+    # CHAPTER_ORDER. only=[] is a valid (empty) selection -> minimal document.
+    only_set = set(only) if isinstance(only, (list, tuple, set)) else None
+
     # A single glossary collector is shared by every chapter via ctx['glossary'].
     # Chapters call ctx['glossary'].add(key, label, definition) and mark in-text
     # appearances with [[term:key]]…[[/term]]; the glosario chapter renders the
@@ -103,6 +133,10 @@ def build_document(profile: dict, ctx: dict = None) -> list:
     for cid in CHAPTER_ORDER:
         if cid in (_PORTADA, _GLOSARIO):
             continue
+        # When a selection is given, skip body chapters outside it. portada and
+        # glosario are never filtered (handled out of this loop).
+        if only_set is not None and cid not in only_set:
+            continue
         ch = build_chapter(cid, profile, ctx)
         if ch is not None and ch.blocks:
             body.append(ch)

python/functions/datascience/automatic_eda/md_completeness_test.py

@@ -0,0 +1,253 @@
+"""Tests for the Markdown completeness appendix (report 2053).
+
+The AutomaticEDA Markdown is the output meant to be *pasted into an LLM*, so it
+must carry EVERYTHING the engine computed — even the numbers the human-facing
+chapters (shared with the PDF/PPTX) drop for readability. ``render_md`` appends a
+full-data appendix built from ``meta['profile']`` that closes the six losses the
+evaluation found:
+
+1. the complete association matrix (every pair, incl. correlation_ratio /
+   cramers_v) — not just the top extremes;
+2. every numeric statistic for every numeric column (skew/kurtosis/percentiles);
+3. the concrete recommended re-expression;
+4. KMeans ``scores_by_k``;
+5. the normality test statistics;
+6. correct headers for bar/scree figure tables (not ``Desde/Hasta/Frecuencia``).
+
+Self-contained: a synthetic profile, no DuckDB, no heavy renderer.
+"""
+
+import os
+import sys
+
+import pytest  # noqa: F401
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_FUNCTIONS = os.path.abspath(os.path.join(_HERE, "..", "..", ".."))  # python/functions
+if _FUNCTIONS not in sys.path:
+    sys.path.insert(0, _FUNCTIONS)
+
+from datascience.automatic_eda import model  # noqa: E402
+from datascience.automatic_eda.render_md_impl import (  # noqa: E402
+    _bars_table,
+    _is_histogram_caption,
+    _profile_appendix,
+    render_md,
+)
+
+
+# --------------------------------------------------------------------------- #
+# Synthetic profile fixtures.
+# --------------------------------------------------------------------------- #
+def _numeric(skew, kurtosis):
+    """A numeric stat block with every key the appendix serializes."""
+    return {
+        "count": 100, "min": 0.0, "max": 10.0, "mean": 5.0, "median": 5.0,
+        "mode": 4.0, "std": 2.0, "variance": 4.0, "cv": 0.4,
+        "p1": 0.1, "p5": 0.5, "p25": 2.5, "p50": 5.0, "p75": 7.5,
+        "p95": 9.5, "p99": 9.9, "iqr": 5.0, "skew": skew, "kurtosis": kurtosis,
+        "n_outliers": 1, "distribution_type": "normal",
+    }
+
+
+def _profile():
+    """A small but structurally faithful TableProfile (3 numeric, 2 categorical)."""
+    pairs = [
+        {"a": "A", "b": "B", "a_type": "numeric", "b_type": "numeric",
+         "method": "pearson/spearman", "value": 0.8,
+         "p_value": 1e-9, "p_value_adjusted": 2e-9, "significant": True},
+        {"a": "A", "b": "C", "a_type": "numeric", "b_type": "numeric",
+         "method": "pearson/spearman", "value": -0.3,
+         "p_value": 0.01, "p_value_adjusted": 0.02, "significant": True},
+        {"a": "A", "b": "Cat1", "a_type": "numeric", "b_type": "categorical",
+         "method": "correlation_ratio", "value": 0.45,
+         "p_value": 0.001, "p_value_adjusted": 0.002, "significant": True},
+        # The single cat-cat pair the human chapter never shows.
+        {"a": "Cat1", "b": "Cat2", "a_type": "categorical",
+         "b_type": "categorical", "method": "cramers_v", "value": 0.11,
+         "p_value": 0.04, "p_value_adjusted": 0.05, "significant": False},
+    ]
+    return {
+        "correlations": {
+            "pairs": pairs,
+            "multiple_testing": {"method": "bh", "n_tests": 4, "n_rejected": 3},
+        },
+        "columns": [
+            {"name": "A", "count": 100, "numeric": _numeric(0.0, -1.2),
+             "reexpression": {"recommended": "none", "ladder_power": 1.0,
+                              "reason": "symmetric", "alternatives": []}},
+            {"name": "B", "count": 100, "numeric": _numeric(4.77, 33.1),
+             "reexpression": {"recommended": "log1p", "ladder_power": 0.0,
+                              "reason": "skew 4.77 with zeros",
+                              "alternatives": [{"transform": "yeo-johnson"},
+                                               {"transform": "sqrt"}]}},
+            {"name": "C", "count": 100, "numeric": _numeric(-0.6, 0.2)},
+            {"name": "Cat1", "categorical": {"top": [], "mode": "x"}},
+            {"name": "Cat2", "categorical": {"top": [], "mode": "y"}},
+        ],
+        "models": {
+            "kmeans": {
+                "best_k": 3,
+                "scores_by_k": [
+                    {"k": 2, "silhouette": 0.46, "inertia": 900.0},
+                    {"k": 3, "silhouette": 0.50, "inertia": 550.0},
+                    {"k": 4, "silhouette": 0.38, "inertia": 430.0},
+                ],
+                "cluster_sizes": [40, 35, 25],
+            },
+            "normality": {
+                "A": {"n": 100,
+                      "jarque_bera": {"stat": 18.7, "p": 8e-5, "normal": False},
+                      "dagostino": {"stat": 18.1, "p": 1e-4, "normal": False},
+                      "shapiro": {"stat": 0.98, "p": 7e-8, "normal": False},
+                      "is_normal": False},
+                "C": {"n": 100,
+                      "jarque_bera": {"stat": 2.1, "p": 0.35, "normal": True},
+                      "dagostino": {"stat": 1.9, "p": 0.38, "normal": True},
+                      "shapiro": {"stat": 0.99, "p": 0.12, "normal": True},
+                      "is_normal": True},
+            },
+        },
+    }
+
+
+def _dummy_chapters():
+    """A minimal one-chapter document so render_md does not early-return empty."""
+    return model.as_chapters([
+        {"id": "intro", "title": "Intro",
+         "blocks": [{"kind": "markdown", "text": "cuerpo del informe"}]},
+    ])
+
+
+def _render(tmp_path, profile):
+    out = os.path.join(str(tmp_path), "out.md")
+    res = render_md(_dummy_chapters(), out, {"title": "EDA — t", "profile": profile})
+    assert res["path"] == out
+    return open(out, encoding="utf-8").read()
+
+
+def _table_rows(md, section_title):
+    """Count data rows of the first Markdown table under ``section_title``."""
+    seg = md.split(section_title, 1)[1]
+    rows, in_t, seen_sep = 0, False, False
+    for ln in seg.splitlines():
+        if ln.startswith("|"):
+            in_t = True
+            stripped = ln.replace("|", "").replace(" ", "")
+            if stripped and set(stripped) == {"-"}:
+                seen_sep = True
+                continue
+            if seen_sep:
+                rows += 1
+        elif in_t and not ln.strip():
+            break
+    return rows
+
+
+# --------------------------------------------------------------------------- #
+# Golden: every datum the profile holds reaches the .md.
+# --------------------------------------------------------------------------- #
+def test_appendix_lists_all_correlation_pairs(tmp_path):
+    md = _render(tmp_path, _profile())
+    assert "## Apéndice — Datos completos del perfil" in md
+    # All 4 pairs (the real titanic profile has 28; here 4 synthetic).
+    assert _table_rows(md, "### Matriz de asociación") == 4
+    # The cat-cat Cramér's V pair the human chapter drops is present.
+    assert "Cat1 ↔ Cat2" in md
+    assert "cramers_v" in md
+    assert "correlation_ratio" in md
+
+
+def test_appendix_has_skew_kurtosis_for_every_numeric(tmp_path):
+    md = _render(tmp_path, _profile())
+    seg = md.split("### Estadísticos numéricos completos", 1)[1].split("###", 1)[0]
+    lines = [l for l in seg.splitlines() if l.startswith("|")]
+    header = [h.strip() for h in lines[0].strip("|").split("|")]
+    assert "skew" in header and "kurtosis" in header
+    ski, kui = header.index("skew"), header.index("kurtosis")
+    data = lines[2:]  # skip header + separator
+    assert len(data) == 3  # exactly the 3 numeric columns
+    for row in data:
+        cells = [c.strip() for c in row.strip("|").split("|")]
+        assert cells[ski] != "", f"missing skew in {cells[0]}"
+        assert cells[kui] != "", f"missing kurtosis in {cells[0]}"
+
+
+def test_appendix_has_extended_percentiles(tmp_path):
+    md = _render(tmp_path, _profile())
+    seg = md.split("### Estadísticos numéricos completos", 1)[1]
+    header = [h.strip() for h in seg.splitlines()[2].strip("|").split("|")]
+    for p in ("p1", "p5", "p25", "p75", "p95", "p99"):
+        assert p in header, f"percentile {p} missing from describe header"
+
+
+def test_appendix_names_concrete_reexpression(tmp_path):
+    md = _render(tmp_path, _profile())
+    assert "### Re-expresión recomendada" in md
+    assert "log1p" in md  # the concrete transform, not just "consider re-expressing"
+    assert "yeo-johnson" in md  # alternatives listed too
+
+
+def test_appendix_has_kmeans_scores_by_k(tmp_path):
+    md = _render(tmp_path, _profile())
+    assert "scores_by_k" in md
+    assert _table_rows(md, "#### KMeans — selección de k") == 3  # k=2,3,4
+
+
+def test_appendix_has_normality_statistics(tmp_path):
+    md = _render(tmp_path, _profile())
+    assert "JB stat" in md  # the statistic, not only the p-value
+    assert "Shapiro stat" in md
+    assert _table_rows(md, "#### Tests de normalidad") == 2  # cols A and C
+
+
+# --------------------------------------------------------------------------- #
+# Edge: a profile missing models / correlations degrades, never raises.
+# --------------------------------------------------------------------------- #
+def test_lite_profile_without_models(tmp_path):
+    prof = _profile()
+    prof.pop("models")  # lite: no KMeans/normality
+    md = _render(tmp_path, prof)
+    assert "scores_by_k" not in md  # section skipped
+    assert "Matriz de asociación" in md  # correlations still dumped
+    assert "## Apéndice" in md
+
+
+def test_profile_without_correlations(tmp_path):
+    prof = _profile()
+    prof.pop("correlations")
+    md = _render(tmp_path, prof)  # must not raise
+    assert "Matriz de asociación" not in md
+    assert "Estadísticos numéricos completos" in md  # numeric section still there
+
+
+def test_no_profile_means_no_appendix(tmp_path):
+    out = os.path.join(str(tmp_path), "noprof.md")
+    res = render_md(_dummy_chapters(), out, {"title": "x"})
+    assert res["path"] == out
+    assert "## Apéndice" not in open(out, encoding="utf-8").read()
+
+
+def test_appendix_helper_is_defensive():
+    assert _profile_appendix(None) == ""
+    assert _profile_appendix({}) == ""
+    assert _profile_appendix({"columns": []}) == ""
+
+
+# --------------------------------------------------------------------------- #
+# Loss #6: bar/scree figure tables get a non-misleading header.
+# --------------------------------------------------------------------------- #
+def test_histogram_caption_detection():
+    assert _is_histogram_caption("Histograma de Age")
+    assert _is_histogram_caption("Distribución de Fare")
+    assert not _is_histogram_caption("Media de Survived por Sex")
+    assert not _is_histogram_caption("Varianza explicada (scree PCA)")
+
+
+def test_bars_table_custom_header():
+    bars = [(0.0, 1.0, 5.0), (1.0, 2.0, 3.0)]
+    hist = _bars_table(bars)  # default histogram header
+    assert "| Desde | Hasta | Frecuencia |" in hist
+    bar = _bars_table(bars, ("Inicio", "Fin", "Valor"))
+    assert "| Inicio | Fin | Valor |" in bar
+    assert "Frecuencia" not in bar

python/functions/datascience/automatic_eda/model.py

@@ -38,10 +38,18 @@ ENGINE_NAME = "AutomaticEDA"
 # --------------------------------------------------------------------------- #
 @dataclass
 class Heading:
-    """A section heading. ``level`` 1 (largest) .. 3 (smallest)."""
+    """A section heading. ``level`` 1 (largest) .. 3 (smallest).
+
+    ``underline`` and ``size_pt`` are optional emphasis hints honoured by the
+    PPTX renderer (the cover uses them to show the dataset name big and
+    underlined). ``size_pt`` overrides the per-level font size when set; the PDF
+    renderer ignores both so its layout is unchanged.
+    """
 
     text: str = ""
     level: int = 1
+    underline: bool = False
+    size_pt: Optional[float] = None
     kind: str = field(default="heading", init=False)
 
 
@@ -62,10 +70,17 @@ class Markdown:
 
 @dataclass
 class KVTable:
-    """A two-column key/value table. ``rows`` is a list of ``(label, value)``."""
+    """A two-column key/value table. ``rows`` is a list of ``(label, value)``.
+
+    ``value_align`` controls the horizontal alignment of the value column in the
+    PDF renderer: ``"left"`` (default) keeps values next to the label column;
+    ``"right"`` pins them to the right margin (used by the cover's analysis
+    summary so the numbers line up flush right).
+    """
 
     rows: list = field(default_factory=list)
     title: Optional[str] = None
+    value_align: str = "left"
     kind: str = field(default="kv_table", init=False)
 
 
@@ -145,11 +160,21 @@ class Group:
     a chapter can give each unit its own page — e.g. one categorical column per
     page (see CAT DISTR). It is purely additive: the default False keeps the plain
     keep-together behaviour for every existing chapter.
+
+    ``layout`` is a hint for how the group's children are arranged:
+    ``"stack"`` (default) keeps the historical top-to-bottom flow; ``"side_by_side"``
+    asks the PPTX renderer to place the group's table to the LEFT and its figure to
+    the RIGHT of the same slide (table ~55% width, figure ~45%), measuring so both
+    fit and falling back to stacking when they do not. The PDF renderer treats
+    ``"side_by_side"`` exactly like ``"stack"`` (the A5 mobile page is too narrow for
+    two readable columns). Unknown values degrade to ``"stack"``. Purely additive:
+    the default keeps every existing chapter unchanged.
     """
 
     blocks: list = field(default_factory=list)
     title: Optional[str] = None
     page_break_before: bool = False
+    layout: str = "stack"
     kind: str = field(default="group", init=False)
 
 
@@ -168,6 +193,22 @@ class GlossaryEntry:
     kind: str = field(default="glossary_entry", init=False)
 
 
+@dataclass
+class TocEntry:
+    """One clickable index (table-of-contents) entry shown on the cover.
+
+    Rendered as a single line — the chapter ``label`` in the accent link colour —
+    that, once the document is laid out, becomes a real click jumping to the first
+    page/slide of the target chapter (PDF link annotation via PyMuPDF; PPTX native
+    slide jump). ``target_id`` is matched against each chapter's ``id`` *and* its
+    ``title`` (the cover only knows chapter titles), so either resolves. If the
+    target cannot be resolved the entry still renders as plain text (never cut)."""
+
+    label: str = ""
+    target_id: str = ""
+    kind: str = field(default="toc_entry", init=False)
+
+
 @dataclass
 class Chapter:
     """An ordered set of blocks with an id, a title and a generation version."""
@@ -192,13 +233,14 @@ _BLOCK_BY_KIND = {
     "note": Note,
     "group": Group,
     "glossary_entry": GlossaryEntry,
+    "toc_entry": TocEntry,
 }
 
 
 def as_block(obj: Any):
     """Coerce a value into a block dataclass. Unknown values become a Note."""
     if isinstance(obj, (Heading, Markdown, KVTable, DataTable, Figure, Image,
-                        Caption, Note, Group, GlossaryEntry)):
+                        Caption, Note, Group, GlossaryEntry, TocEntry)):
         if isinstance(obj, Group):
             obj.blocks = as_blocks(obj.blocks)
         return obj
@@ -210,13 +252,20 @@ def as_block(obj: Any):
         # Build only with fields the dataclass accepts (ignore extras).
         try:
             if cls is Heading:
+                size_pt = obj.get("size_pt")
                 return Heading(text=_safe_str(obj.get("text")),
-                               level=int(obj.get("level", 1) or 1))
+                               level=int(obj.get("level", 1) or 1),
+                               underline=bool(obj.get("underline", False)),
+                               size_pt=(float(size_pt)
+                                        if isinstance(size_pt, (int, float))
+                                        else None))
             if cls is Markdown:
                 return Markdown(text=_safe_str(obj.get("text")))
             if cls is KVTable:
                 return KVTable(rows=list(obj.get("rows") or []),
-                               title=obj.get("title"))
+                               title=obj.get("title"),
+                               value_align=_safe_str(
+                                   obj.get("value_align")) or "left")
             if cls is DataTable:
                 return DataTable(header=list(obj.get("header") or []),
                                  rows=list(obj.get("rows") or []),
@@ -237,11 +286,15 @@ def as_block(obj: Any):
                 return Group(blocks=as_blocks(obj.get("blocks")),
                              title=obj.get("title"),
                              page_break_before=bool(
-                                 obj.get("page_break_before", False)))
+                                 obj.get("page_break_before", False)),
+                             layout=_safe_str(obj.get("layout")) or "stack")
             if cls is GlossaryEntry:
                 return GlossaryEntry(key=_safe_str(obj.get("key")),
                                      label=_safe_str(obj.get("label")),
                                      definition=_safe_str(obj.get("definition")))
+            if cls is TocEntry:
+                return TocEntry(label=_safe_str(obj.get("label")),
+                                target_id=_safe_str(obj.get("target_id")))
         except Exception:  # noqa: BLE001 — never raise on a malformed block.
             return Note(text=_safe_str(obj))
     return Note(text=_safe_str(obj))

python/functions/datascience/automatic_eda/render_features_test.py

@@ -298,11 +298,16 @@ def test_cover_first_glossary_last_with_summary():
     headings = [b.text for b in cover.blocks if b.kind == "heading"]
     assert any("Resumen" in h for h in headings), \
         "la portada no incluye el resumen agregado"
-    # The summary reflects the body chapters (e.g. the numeric/categorical ones).
+    # The index ("Índice") is now a clickable list of TocEntry blocks (one per
-    cover_text = " ".join(
+    # body chapter), not a markdown bullet list. Verify both the heading and that
-        b.text for b in cover.blocks if getattr(b, "kind", "") == "markdown")
+    # the entries name the body chapters.
-    assert "Distribuciones" in cover_text, \
+    assert any("Índice" in h for h in headings), \
-        "el resumen de portada no menciona los capítulos del cuerpo"
+        "la portada no incluye la sección Índice"
+    toc_labels = " ".join(
+        getattr(b, "label", "") for b in cover.blocks
+        if getattr(b, "kind", "") == "toc_entry")
+    assert "Distribuciones" in toc_labels, \
+        "el índice de portada no menciona los capítulos del cuerpo"
 
 
 # --------------------------------------------------------------------------- #

python/functions/datascience/automatic_eda/render_md_impl.py

@@ -178,9 +178,17 @@ def _md_data_table(block) -> str:
     return "\n".join(lines)
 
 
-def _bars_table(bars: list) -> str:
+def _bars_table(bars: list, header: tuple = ("Desde", "Hasta", "Frecuencia")) -> str:
-    """Render extracted bar/histogram data as a Markdown table (Desde/Hasta/Frec)."""
+    """Render extracted bar/histogram data as a Markdown table.
-    lines = ["| Desde | Hasta | Frecuencia |", "| --- | --- | --- |"]
+
+    ``header`` is the 3-column header to use. Histogram bars are
+    ``(Desde, Hasta, Frecuencia)``; bar/scree charts (means by group, PCA
+    explained variance) are *not* bins, so the caller passes a semantically
+    correct header (e.g. ``(Inicio, Fin, Valor)``) to avoid the misleading
+    "Frecuencia" label — see report 2053, loss #6.
+    """
+    h0, h1, h2 = header
+    lines = [f"| {h0} | {h1} | {h2} |", "| --- | --- | --- |"]
     shown = bars[:_MAX_BAR_ROWS]
     for x0, x1, h in shown:
         lines.append(f"| {_fmt_num(x0)} | {_fmt_num(x1)} | {_fmt_num(h)} |")
@@ -191,6 +199,18 @@ def _bars_table(bars: list) -> str:
     return out
 
 
+def _is_histogram_caption(caption: str) -> bool:
+    """True when a figure caption describes a histogram (genuine numeric bins).
+
+    Histograms are the only figures whose bars are real ``[Desde, Hasta)`` bins
+    with a frequency count. Bar charts (means by group) and the PCA scree plot
+    carry per-category / per-component values, not bins — they must not inherit
+    the ``Desde/Hasta/Frecuencia`` header.
+    """
+    c = (caption or "").lower()
+    return "histograma" in c or "distribución" in c or "distribucion" in c
+
+
 def _extract_bars(fig) -> list:
     """Collect (x_from, x_to, height) of the rectangular bars of a matplotlib fig.
 
@@ -253,7 +273,13 @@ def _md_figure(block, meta: dict, out_path: str, counter: list) -> str:
         if fig is not None:
             bars = _extract_bars(fig)
             if bars:
-                parts.append(_bars_table(bars))
+                # A histogram's bars are genuine numeric bins (Desde/Hasta/
+                # Frecuencia). Bar charts and the PCA scree plot are not bins —
+                # give them a header that does not lie about "Frecuencia".
+                header = (("Desde", "Hasta", "Frecuencia")
+                          if _is_histogram_caption(caption)
+                          else ("Inicio", "Fin", "Valor"))
+                parts.append(_bars_table(bars, header))
             if meta.get("embed_figures"):
                 png = _embed_png(fig, out_path, counter)
                 if png:
@@ -354,6 +380,258 @@ def _serialize_block(block, meta: dict, out_path: str, counter: list) -> str:
     return _md_note(model.Note(text=model._safe_str(block)))
 
 
+# --------------------------------------------------------------------------- #
+# Profile appendix — the data the human-facing chapters drop.
+#
+# The chapter document (shared with the PDF/PPTX renderers) is designed for human
+# reading and intentionally omits raw numbers: the correlation matrix shows only
+# the top extremes, the numeric blocks skip skew/kurtosis/extended percentiles,
+# the model chapter does not list ``scores_by_k`` or the normality test
+# statistics. But the Markdown is meant to be *pasted into an LLM*, so it should
+# carry EVERYTHING the engine computed. This appendix serializes the full
+# ``profile`` (passed via ``meta['profile']``) as Markdown tables, additively:
+# the PDF/PPTX are untouched, the .md simply has more than they do. Each section
+# is emitted only when its source data is present, so a ``lite`` profile (no
+# models) or a profile without correlations degrades cleanly instead of raising.
+# See report 2053 for the six losses this closes.
+# --------------------------------------------------------------------------- #
+def _pair_types(a_type, b_type) -> str:
+    """Short ``num↔cat`` label for an association pair's variable types."""
+    def short(t):
+        t = model._safe_str(t).lower()
+        if t.startswith("num"):
+            return "num"
+        if t.startswith("cat"):
+            return "cat"
+        return t or "?"
+    return f"{short(a_type)}↔{short(b_type)}"
+
+
+def _app_correlations(corr: dict) -> str:
+    """Loss #1 — every association pair (not just the top extremes).
+
+    Dumps all of ``correlations['pairs']`` as a table (pair · types · method ·
+    value · p · p-FDR · significant), ordered by |value| desc so the strongest
+    associations lead while nothing is cut. Includes the ``correlation_ratio``
+    (num↔cat) and ``cramers_v`` (cat↔cat) pairs the human chapter never shows.
+    """
+    pairs = list(corr.get("pairs", []) or [])
+    if not pairs:
+        return ""
+    def keyfn(p):
+        try:
+            return -abs(float(p.get("value")))
+        except Exception:  # noqa: BLE001
+            return 0.0
+    pairs_sorted = sorted(pairs, key=keyfn)
+    lines = ["### Matriz de asociación — todos los pares",
+             "",
+             ("| Par | Tipos | Método | Valor | p-value | p-ajustado (FDR) "
+              "| ¿Sig? |"),
+             "| --- | --- | --- | --- | --- | --- | --- |"]
+    for p in pairs_sorted:
+        par = f"{_cell(p.get('a'))} ↔ {_cell(p.get('b'))}"
+        types = _pair_types(p.get("a_type"), p.get("b_type"))
+        method = _cell(p.get("method"))
+        val = _fmt_num(p.get("value"))
+        pv = _fmt_num(p.get("p_value")) if p.get("p_value") is not None else ""
+        padj = (_fmt_num(p.get("p_value_adjusted"))
+                if p.get("p_value_adjusted") is not None else "")
+        sig = "sí" if p.get("significant") else "no"
+        lines.append(
+            f"| {par} | {types} | {method} | {val} | {pv} | {padj} | {sig} |")
+    mt = corr.get("multiple_testing") or {}
+    n_tests = mt.get("n_tests", corr.get("n_tests"))
+    n_rej = mt.get("n_rejected")
+    note_bits = [f"{len(pairs)} pares en total"]
+    if n_tests is not None and n_rej is not None:
+        note_bits.append(
+            f"{n_rej} de {n_tests} significativos tras corrección "
+            f"{model._safe_str(mt.get('method', 'FDR')).upper()}")
+    lines.append("")
+    lines.append(f"*{'; '.join(note_bits)}.*")
+    return "\n".join(lines)
+
+
+# Numeric statistics, in serialization order: (profile key, column header).
+_NUM_STATS = [
+    ("count", "n"), ("mean", "mean"), ("median", "median"), ("mode", "mode"),
+    ("std", "std"), ("variance", "variance"), ("cv", "cv"),
+    ("skew", "skew"), ("kurtosis", "kurtosis"),
+    ("min", "min"), ("p1", "p1"), ("p5", "p5"), ("p25", "p25"), ("p50", "p50"),
+    ("p75", "p75"), ("p95", "p95"), ("p99", "p99"), ("iqr", "iqr"),
+    ("max", "max"), ("n_outliers", "outliers"),
+    ("distribution_type", "distribución"),
+]
+
+
+def _app_numeric_describe(columns: list) -> str:
+    """Loss #2 — every numeric statistic for every numeric column.
+
+    One row per numeric column with the full describe: mean/median/mode/std/
+    variance/cv, skew & kurtosis (for ALL columns, not only the skewed ones),
+    p1/p5/p25/p50/p75/p95/p99, iqr, min/max, outliers and distribution_type.
+    """
+    rows = []
+    for info in (columns or []):
+        num = info.get("numeric") if isinstance(info, dict) else None
+        if not num:
+            continue
+        name = _cell(info.get("name"))
+        cells = [name]
+        for key, _hdr in _NUM_STATS:
+            v = num.get("count" if key == "count" else key)
+            if key == "count":
+                v = num.get("count", info.get("count"))
+            if key == "distribution_type":
+                cells.append(_cell(v))
+            else:
+                cells.append(_fmt_num(v) if v is not None else "")
+        rows.append(cells)
+    if not rows:
+        return ""
+    header = ["Columna"] + [hdr for _k, hdr in _NUM_STATS]
+    lines = ["### Estadísticos numéricos completos (describe)",
+             "",
+             "| " + " | ".join(header) + " |",
+             "| " + " | ".join(["---"] * len(header)) + " |"]
+    for cells in rows:
+        lines.append("| " + " | ".join(cells) + " |")
+    return "\n".join(lines)
+
+
+def _app_reexpression(columns: list) -> str:
+    """Loss #3 — the concrete recommended re-expression per column.
+
+    Names the transform (log1p/sqrt/yeo-johnson/none) instead of a vague
+    "consider re-expressing", with the ladder power, reason and alternatives.
+    """
+    rows = []
+    for info in (columns or []):
+        rx = info.get("reexpression") if isinstance(info, dict) else None
+        if not rx or not isinstance(rx, dict):
+            continue
+        rec = model._safe_str(rx.get("recommended")).strip()
+        if not rec:
+            continue
+        alts = rx.get("alternatives") or []
+        alt_txt = ", ".join(
+            model._safe_str(a.get("transform")) for a in alts
+            if isinstance(a, dict) and a.get("transform")) or "—"
+        rows.append([
+            _cell(info.get("name")), _cell(rec),
+            _fmt_num(rx.get("ladder_power")) if rx.get("ladder_power") is not None else "",
+            _cell(rx.get("reason")), _cell(alt_txt),
+        ])
+    if not rows:
+        return ""
+    lines = ["### Re-expresión recomendada (escalera de Tukey)",
+             "",
+             "| Columna | Recomendada | Potencia | Razón | Alternativas |",
+             "| --- | --- | --- | --- | --- |"]
+    for r in rows:
+        lines.append("| " + " | ".join(r) + " |")
+    return "\n".join(lines)
+
+
+def _app_kmeans_scores(kmeans: dict) -> str:
+    """Loss #4 — KMeans silhouette + inertia per k (justifies the chosen k)."""
+    scores = list(kmeans.get("scores_by_k", []) or [])
+    if not scores:
+        return ""
+    best_k = kmeans.get("best_k")
+    lines = ["#### KMeans — selección de k (`scores_by_k`)",
+             "",
+             "| k | Silhouette | Inercia | Elegido |",
+             "| --- | --- | --- | --- |"]
+    for s in scores:
+        if not isinstance(s, dict):
+            continue
+        k = s.get("k")
+        chosen = "✓" if best_k is not None and k == best_k else ""
+        lines.append(
+            f"| {_fmt_num(k)} | {_fmt_num(s.get('silhouette'))} "
+            f"| {_fmt_num(s.get('inertia'))} | {chosen} |")
+    return "\n".join(lines)
+
+
+def _app_normality(normality: dict) -> str:
+    """Loss #5 — each normality test's statistic next to its p-value."""
+    if not isinstance(normality, dict) or not normality:
+        return ""
+    lines = ["#### Tests de normalidad (estadístico + p-value)",
+             "",
+             ("| Columna | n | JB stat | JB p | D'Agostino stat | D'Agostino p "
+              "| Shapiro stat | Shapiro p | ¿Normal? |"),
+             "| --- | --- | --- | --- | --- | --- | --- | --- | --- |"]
+    any_row = False
+    for col, res in normality.items():
+        if not isinstance(res, dict):
+            continue
+        jb = res.get("jarque_bera") or {}
+        da = res.get("dagostino") or {}
+        sh = res.get("shapiro") or {}
+        is_norm = "sí" if res.get("is_normal") else "no"
+        lines.append(
+            f"| {_cell(col)} | {_fmt_num(res.get('n')) if res.get('n') is not None else ''} "
+            f"| {_fmt_num(jb.get('stat'))} | {_fmt_num(jb.get('p'))} "
+            f"| {_fmt_num(da.get('stat'))} | {_fmt_num(da.get('p'))} "
+            f"| {_fmt_num(sh.get('stat'))} | {_fmt_num(sh.get('p'))} | {is_norm} |")
+        any_row = True
+    return "\n".join(lines) if any_row else ""
+
+
+def _profile_appendix(profile: dict) -> str:
+    """Build the full-data appendix from a TableProfile dict (additive).
+
+    Returns a Markdown ``## Apéndice`` section with one sub-table per loss the
+    human chapters drop, or ``""`` when the profile carries none of them. Never
+    raises: a missing/oddly-shaped section is skipped, not fatal.
+    """
+    if not isinstance(profile, dict):
+        return ""
+    sections: list = []
+    try:
+        corr = profile.get("correlations") or {}
+        seg = _app_correlations(corr) if isinstance(corr, dict) else ""
+        if seg:
+            sections.append(seg)
+    except Exception:  # noqa: BLE001
+        pass
+    try:
+        columns = profile.get("columns") or []
+        seg = _app_numeric_describe(columns)
+        if seg:
+            sections.append(seg)
+        seg = _app_reexpression(columns)
+        if seg:
+            sections.append(seg)
+    except Exception:  # noqa: BLE001
+        pass
+    try:
+        models = profile.get("models") or {}
+        if isinstance(models, dict):
+            model_segs = []
+            seg = _app_kmeans_scores(models.get("kmeans") or {})
+            if seg:
+                model_segs.append(seg)
+            seg = _app_normality(models.get("normality") or {})
+            if seg:
+                model_segs.append(seg)
+            if model_segs:
+                sections.append(
+                    "### Modelos — detalle\n\n" + "\n\n".join(model_segs))
+    except Exception:  # noqa: BLE001
+        pass
+    if not sections:
+        return ""
+    intro = ("Volcado completo de los datos que el motor computó y que los "
+             "capítulos (pensados para lectura humana / PDF) resumen. "
+             "Pensado para que un LLM reconstruya el análisis entero.")
+    return ("## Apéndice — Datos completos del perfil\n\n"
+            f"*{intro}*\n\n" + "\n\n".join(sections))
+
+
 # --------------------------------------------------------------------------- #
 # Entry point.
 # --------------------------------------------------------------------------- #
@@ -437,6 +715,18 @@ def render_md(chapters: list, out_path: str, meta: dict = None) -> dict:
                 segments.append(seg)
         chapters_meta.append({"id": ch.id, "version": ch.version})
 
+    # Full-data appendix: dump everything the profile holds that the human
+    # chapters drop (additive — the .md ends up with more than the PDF/PPTX).
+    # Emitted only when a profile is supplied via meta['profile']; never fatal.
+    try:
+        appendix = _profile_appendix(meta.get("profile"))
+    except Exception as e:  # noqa: BLE001
+        appendix = ""
+        notes.append(f"apéndice de perfil omitido: {e}")
+    if appendix:
+        segments.append("---")
+        segments.append(appendix)
+
     content = "\n\n".join(segments) + "\n"
     note = f"{len(content)} caracteres"
     if notes:

python/functions/datascience/automatic_eda/render_pdf_impl.py

@@ -46,11 +46,23 @@ _MUTED = "#8a8a8a"
 _RULE = "#cccccc"
 _HEAD_BG = "#eef3f6"
 
+# Rasterization DPI for every embedded raster (figure/table image) AND for the
+# page save itself. Raised from the old 150/default-100 to 220 so a reader can
+# pinch-zoom on a phone and still see crisp detail (axis labels, table cells)
+# without pixelation. Text stays vectorial (pdf.fonttype=42) so it remains
+# selectable regardless of DPI — only the embedded images gain resolution. 220 is
+# a deliberate balance: noticeably sharper than 150 while keeping the file size
+# reasonable. ``savefig.dpi`` matters because matplotlib re-rasterizes each
+# ``imshow`` when PdfPages writes the page; without it the final image would land
+# at ~100 dpi no matter how sharp the intermediate PNG was.
+_RASTER_DPI = 220
+
 _RC = {
     "font.size": 10,
     "font.family": "sans-serif",
     "figure.facecolor": "white",
     "savefig.facecolor": "white",
+    "savefig.dpi": _RASTER_DPI,
     "pdf.fonttype": 42,  # embed TrueType — text stays selectable on mobile.
 }
 
@@ -80,6 +92,10 @@ class _PdfState:
         # points (1/72") with a top-left origin — same convention as PyMuPDF.
         self.term_sources = []       # [{key, page, rect:[x0,y0,x1,y1]}]
         self.term_dests = {}         # key -> {page, point:[x,y]}
+        # Clickable index (cover → chapter). Sources are the cover's TocEntry
+        # rects; chapter_starts maps a chapter id AND its title to its first page.
+        self.toc_sources = []        # [{target_id, page, rect:[x0,y0,x1,y1]}]
+        self.chapter_starts = {}     # id|title -> {page, point:[x,y]}
 
 
 # --------------------------------------------------------------------------- #
@@ -317,10 +333,18 @@ def _place_kv_table(st: _PdfState, block) -> None:
     if title:
         _place_heading(st, model.Heading(title, level=2))
     rows = getattr(block, "rows", []) or []
+    # ``value_align="right"`` pins the value column to the right margin (label
+    # left, number flush right) — used by the cover's analysis summary.
+    right = str(getattr(block, "value_align", "left")).lower() == "right"
     key_w = 1.9  # inches reserved for the label column.
+    # Right-aligned values wrap against the full usable width minus the label
+    # column; left-aligned values wrap against the value column only.
     val_chars = tl.chars_per_line(_USABLE_W - key_w - 0.1, _FS_BODY)
     lh = tl.line_height_in(_FS_BODY)
-    for row in rows:
+    # ``data_idx`` is the 0-based logical row index: even rows (1-based) are
+    # zebra-shaded → 0-based odd indices, matching the data-table convention so
+    # every table in the document carries the same striping.
+    for data_idx, row in enumerate(rows):
         try:
             label, value = row[0], row[1]
         except Exception:  # noqa: BLE001
@@ -329,11 +353,25 @@ def _place_kv_table(st: _PdfState, block) -> None:
         row_h = lh * len(v_lines) + _ROW_VPAD
         _ensure_space(st, row_h)
         y0 = st.y
+        # Faint zebra fill for even rows, drawn first (zorder 0) so striping
+        # never hides the text/value drawn on top.
+        if data_idx % 2 == 1:
+            st.fig.add_artist(Rectangle(
+                (_xf(_ML), _yf(y0 + row_h)), _xf(_ML + _USABLE_W) - _xf(_ML),
+                _yf(y0) - _yf(y0 + row_h), transform=st.fig.transFigure,
+                color=_ZEBRA, lw=0, zorder=0))
         st.fig.text(_xf(_ML), _yf(y0), tl.strip_inline_md(model._safe_str(label)),
-                    fontsize=_FS_BODY, color=_MUTED, ha="left", va="top")
+                    fontsize=_FS_BODY, color=_MUTED, ha="left", va="top",
+                    zorder=2)
         for k, vl in enumerate(v_lines):
-            st.fig.text(_xf(_ML + key_w), _yf(y0 + k * lh), vl,
+            if right:
-                        fontsize=_FS_BODY, color=_INK, ha="left", va="top")
+                st.fig.text(_xf(_ML + _USABLE_W), _yf(y0 + k * lh), vl,
+                            fontsize=_FS_BODY, color=_INK, ha="right",
+                            va="top", zorder=2)
+            else:
+                st.fig.text(_xf(_ML + key_w), _yf(y0 + k * lh), vl,
+                            fontsize=_FS_BODY, color=_INK, ha="left",
+                            va="top", zorder=2)
         st.y = y0 + row_h
     st.y += _GAP
 
@@ -363,6 +401,57 @@ def _col_widths(header: list, rows: list, fs: float) -> list:
     return widths
 
 
+# Minimal legible characters reserved per column when deciding whether a table
+# can be shown as selectable text. Below this width per column the cells become
+# unreadable, so the table is rasterized to a zoomable high-res image instead.
+_MIN_LEGIBLE_CHARS = 8
+
+
+def _table_fits_as_text(header: list, rows: list) -> bool:
+    """True when the table fits the usable width as readable text.
+
+    A table whose columns cannot each get a minimal legible width within the A5
+    usable width (typically many columns, e.g. a 19-column ``df.head``) is flagged
+    so it is rendered as a single high-resolution image — the reader zooms in on
+    the phone and reads every cell, nothing cut — instead of being squeezed until
+    unreadable. Narrow tables (few columns) keep the selectable-text rendering."""
+    header = header or []
+    rows = rows or []
+    ncol = len(header) if header else (len(rows[0]) if rows else 1)
+    ncol = max(1, ncol)
+    cw = tl.avg_char_width_in(_FS_CELL)
+    min_needed = ncol * (_MIN_LEGIBLE_CHARS * cw + _CELL_PAD * 2)
+    return min_needed <= _USABLE_W
+
+
+def _table_figure_block(block):
+    """Wrap a too-wide table as a lazily-rasterized Figure (cached on the block).
+
+    The table is drawn once via ``render_table_as_figure`` (header shading + zebra)
+    and embedded as one high-res image scaled to fit entirely. The same Figure is
+    reused for measuring and placing so keep-together stays consistent. The table
+    title/note are drawn inside the image (self-describing when zoomed/shared), so
+    the block-level caption is left empty to avoid a duplicate title."""
+    cached = getattr(block, "_aeda_tablefig", None)
+    if cached is not None:
+        return cached
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    title = getattr(block, "title", None)
+    note = getattr(block, "note", None)
+
+    def _make():
+        from datascience.render_table_as_figure import render_table_as_figure
+        return render_table_as_figure(header, rows, title=title, note=note)
+
+    fig = model.Figure(make=_make, caption=None)
+    try:
+        block._aeda_tablefig = fig
+    except Exception:  # noqa: BLE001 — block may reject attributes; degrade.
+        pass
+    return fig
+
+
 def _wrap_row(cells: list, widths: list, fs: float) -> list:
     """Wrap each cell to its column width → list of line-lists per cell."""
     out = []
@@ -402,11 +491,16 @@ def _draw_table_row(st: _PdfState, cells_lines: list, widths: list, fs: float,
 
 
 def _place_data_table(st: _PdfState, block) -> None:
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    # Too many columns to be legible as text → render the whole table as one
+    # high-res image, scaled to fit entirely (the reader zooms to read it).
+    if not _table_fits_as_text(header, rows):
+        _place_figure(st, _table_figure_block(block))
+        return
     title = getattr(block, "title", None)
     if title:
         _place_heading(st, model.Heading(title, level=2))
-    header = list(getattr(block, "header", []) or [])
-    rows = list(getattr(block, "rows", []) or [])
     fs = _FS_CELL
     widths = _col_widths(header, rows, fs)
     header_lines = _wrap_row(header, widths, fs) if header else None
@@ -464,8 +558,11 @@ def _resolve_figure(block):
 
 
 def _png_from_figure(fig) -> bytes:
+    # ``bbox_inches='tight'`` is kept so the real aspect ratio is what we measure
+    # and place. The page save (savefig.dpi in _RC) re-rasterizes this at the same
+    # high DPI, so the embedded image stays crisp for phone zoom.
     buf = io.BytesIO()
-    fig.savefig(buf, format="png", dpi=150, bbox_inches="tight")
+    fig.savefig(buf, format="png", dpi=_RASTER_DPI, bbox_inches="tight")
     buf.seek(0)
     return buf.read()
 
@@ -707,12 +804,16 @@ def _measure_data_table(block) -> float:
     Counts the optional title heading, the wrapped header row, every wrapped data
     row (per-column wrap via the same ``_col_widths``/``_wrap_row`` the placer
     uses) and the optional note. Keep this in sync with ``_place_data_table``."""
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    # Mirror the placer: a too-wide table is drawn as a single image, so its
+    # keep-together height is the image's, not the (squeezed) text layout's.
+    if not _table_fits_as_text(header, rows):
+        return _measure_figure_like(_table_figure_block(block))
     h = 0.0
     title = getattr(block, "title", None)
     if title:
         h += _measure_heading_text(title, 2)
-    header = list(getattr(block, "header", []) or [])
-    rows = list(getattr(block, "rows", []) or [])
     fs = _FS_CELL
     widths = _col_widths(header, rows, fs)
     lh = tl.line_height_in(fs)
@@ -744,6 +845,10 @@ def _measure_block(st: _PdfState, block) -> float:
             lines = tl.wrap(getattr(block, "text", ""),
                             tl.chars_per_line(_USABLE_W, _FS_NOTE))
             return tl.line_height_in(_FS_NOTE) * len(lines) + _GAP
+        if kind == "toc_entry":
+            lines = tl.wrap(tl.strip_inline_md(getattr(block, "label", "")),
+                            tl.chars_per_line(_USABLE_W - 0.22, _FS_BODY)) or [""]
+            return tl.line_height_in(_FS_BODY) * len(lines) + _GAP * 0.4
         if kind == "kv_table":
             return _measure_kv_table(block)
         if kind == "data_table":
@@ -828,6 +933,38 @@ def _place_glossary_entry(st: _PdfState, block) -> None:
     st.y += _GAP * 0.5
 
 
+def _place_toc_entry(st: _PdfState, block) -> None:
+    """Render one clickable index line and record it as a link source.
+
+    Drawn as a bulleted line in the accent link colour; its rectangle is recorded
+    in ``st.toc_sources`` so the post-processor turns it into a real jump to the
+    target chapter's first page. If the target is never resolved the line still
+    shows as plain (accent) text — never cut, never broken."""
+    label = tl.strip_inline_md(getattr(block, "label", "")) or ""
+    target_id = getattr(block, "target_id", "") or ""
+    fs = _FS_BODY
+    lh = tl.line_height_in(fs)
+    bullet = "•  "
+    indent = 0.22
+    max_chars = tl.chars_per_line(_USABLE_W - indent, fs)
+    lines = tl.wrap(label, max_chars) or [""]
+    for idx, ln in enumerate(lines):
+        _ensure_space(st, lh)
+        x = _ML
+        st.fig.text(_xf(x), _yf(st.y), bullet if idx == 0 else "   ",
+                    fontsize=fs, color=_LINK, ha="left", va="top")
+        x += indent
+        w = _text_width_in(st, ln, fs, False)
+        st.fig.text(_xf(x), _yf(st.y), ln, fontsize=fs, color=_LINK,
+                    ha="left", va="top")
+        if target_id and idx == 0:
+            st.toc_sources.append({
+                "target_id": target_id, "page": st.page - 1,
+                "rect": _pt_rect(_ML, st.y, x + w, st.y + lh)})
+        st.y += lh
+    st.y += _GAP * 0.4
+
+
 _PLACERS = {
     "heading": _place_heading,
     "markdown": _place_markdown,
@@ -839,6 +976,7 @@ _PLACERS = {
     "note": _place_note,
     "group": _place_group,
     "glossary_entry": _place_glossary_entry,
+    "toc_entry": _place_toc_entry,
 }
 
 
@@ -870,6 +1008,15 @@ def render_pdf(chapters: list, out_path: str, meta: dict = None) -> dict:
                     st.chapter = ch
                     st.chapter_pages = 0
                     _new_page(st)  # each chapter starts on a fresh page.
+                    # Record this chapter's first page as a link target for the
+                    # cover index (keyed by id AND title, since the cover only
+                    # knows titles). Point is the top of the content area.
+                    _start = {"page": st.page - 1,
+                              "point": [_ML * 72.0, _CONTENT_TOP * 72.0]}
+                    if ch.id:
+                        st.chapter_starts[ch.id] = _start
+                    if getattr(ch, "title", ""):
+                        st.chapter_starts.setdefault(ch.title, _start)
                     for block in ch.blocks:
                         placer = _PLACERS.get(getattr(block, "kind", ""),
                                               _place_note)
@@ -902,7 +1049,7 @@ def render_pdf(chapters: list, out_path: str, meta: dict = None) -> dict:
 
     note = f"{n_pages} páginas"
     if n_links:
-        note += f" · {n_links} enlaces de glosario"
+        note += f" · {n_links} enlaces internos"
     if notes:
         note += " · " + "; ".join(notes)
     return {"path": out_path, "n_pages": n_pages, "chapters": chapters_meta,
@@ -910,9 +1057,11 @@ def render_pdf(chapters: list, out_path: str, meta: dict = None) -> dict:
 
 
 def _wire_glossary_links(st: _PdfState, out_path: str, notes: list) -> int:
-    """Build {source rect → glossary dest} links and apply them via PyMuPDF.
+    """Apply internal PDF links via PyMuPDF: glossary terms + the cover index.
 
-    Returns the number of links applied (0 if there is nothing to wire or the
+    Builds two sets of GOTO links — every in-text glossary term → its entry, and
+    every cover ``TocEntry`` → its chapter's first page — and applies them in one
+    pass. Returns the number of links applied (0 if there is nothing to wire or the
     post-processor is unavailable). Never raises."""
     try:
         links = []
@@ -923,6 +1072,14 @@ def _wire_glossary_links(st: _PdfState, out_path: str, notes: list) -> int:
             links.append({
                 "src_page": src["page"], "src_rect": src["rect"],
                 "dst_page": dest["page"], "dst_point": dest["point"]})
+        # Cover index → chapter first page (clickable, navigable table of contents).
+        for src in st.toc_sources:
+            dest = st.chapter_starts.get(src.get("target_id"))
+            if not dest:
+                continue
+            links.append({
+                "src_page": src["page"], "src_rect": src["rect"],
+                "dst_page": dest["page"], "dst_point": dest["point"]})
         if not links:
             return 0
         from datascience.add_pdf_internal_links import add_pdf_internal_links
@@ -930,7 +1087,7 @@ def _wire_glossary_links(st: _PdfState, out_path: str, notes: list) -> int:
         if isinstance(res, dict) and res.get("status") == "ok":
             return int(res.get("n_links") or 0)
         if isinstance(res, dict) and res.get("error"):
-            notes.append(f"glosario sin enlaces: {res.get('error')}")
+            notes.append(f"enlaces internos no aplicados: {res.get('error')}")
     except Exception as e:  # noqa: BLE001 — links are best-effort.
-        notes.append(f"glosario sin enlaces: {e}")
+        notes.append(f"enlaces internos no aplicados: {e}")
     return 0

python/functions/datascience/automatic_eda/render_pptx_impl.py

@@ -51,6 +51,12 @@ _FS_H1, _FS_H2, _FS_H3 = 20, 16, 13
 _FS_BODY, _FS_CELL, _FS_NOTE = 14, 11, 11
 _GAP = 0.12
 
+# Rasterization DPI for every embedded figure/table image. Raised from 150 to 220
+# so a viewer can zoom into a slide (or a shared picture) and read crisp detail —
+# axis labels, table cells — without pixelation. Kept moderate so the deck size
+# stays reasonable. Same value as the PDF renderer.
+_RASTER_DPI = 220
+
 
 class _PptxState:
     def __init__(self, prs, title: str):
@@ -65,6 +71,10 @@ class _PptxState:
         # Glossary wiring (mejora 6): runs to link and per-term target slide.
         self.term_runs = []           # [(key, run)]
         self.term_anchor_slide = {}   # key -> Slide (glossary entry)
+        # Clickable index (cover → chapter). toc_runs are the cover's index runs;
+        # chapter_starts maps a chapter id AND its title to its first slide.
+        self.toc_runs = []            # [(target_id, run, src_slide)]
+        self.chapter_starts = {}      # id|title -> Slide (chapter first slide)
 
 
 def _rgb(c):
@@ -135,7 +145,7 @@ def _ensure(st: _PptxState, height: float) -> None:
 
 
 def _add_text(st: _PptxState, lines: list, fs: float, color, bold=False,
-              italic=False, indent=0.0, bullet=False) -> None:
+              italic=False, indent=0.0, bullet=False, underline=False) -> None:
     lh = tl.line_height_in(fs)
     height = lh * len(lines) + 0.05
     _ensure(st, height)
@@ -153,6 +163,7 @@ def _add_text(st: _PptxState, lines: list, fs: float, color, bold=False,
         run.font.size = Pt(fs)
         run.font.bold = bold
         run.font.italic = italic
+        run.font.underline = underline
         run.font.color.rgb = _rgb(color)
     st.y += height
 
@@ -206,10 +217,16 @@ def _add_rich_text(st: _PptxState, rich_lines: list, fs: float, color,
 def _place_heading(st: _PptxState, block) -> None:
     level = max(1, min(3, int(getattr(block, "level", 1) or 1)))
     fs = {1: _FS_H1, 2: _FS_H2, 3: _FS_H3}[level]
+    # Optional per-heading emphasis (cover dataset name): a larger font and an
+    # underline. ``size_pt`` overrides the per-level size when set.
+    size_override = getattr(block, "size_pt", None)
+    if isinstance(size_override, (int, float)) and size_override > 0:
+        fs = float(size_override)
+    underline = bool(getattr(block, "underline", False))
     text = tl.strip_inline_md(getattr(block, "text", ""))
     st.last_heading = text or st.last_heading
     lines = tl.wrap(text, tl.chars_per_line(_USABLE_W, fs))
-    _add_text(st, lines, fs, _INK, bold=True)
+    _add_text(st, lines, fs, _INK, bold=True, underline=underline)
     st.y += 0.04
 
 
@@ -302,6 +319,58 @@ def _col_widths(header, rows):
     return [_USABLE_W * w / total for w in clamped]
 
 
+# Minimal legible characters reserved per column when deciding whether a table
+# can be shown as a native (selectable) PowerPoint table. Below this width per
+# column the cells become unreadable, so the table is rasterized to a zoomable
+# high-res image instead. The 16:9 slide is wide, so more columns fit than on A5.
+_MIN_LEGIBLE_CHARS = 8
+_CELL_PAD = 0.05
+
+
+def _table_fits_as_text(header: list, rows: list) -> bool:
+    """True when the table fits the usable slide width as a readable table.
+
+    A table whose columns cannot each get a minimal legible width within the slide
+    usable width (typically many columns, e.g. a 19-column ``df.head``) is flagged
+    so it is rendered as one high-resolution image — the viewer zooms in and reads
+    every cell — instead of being squeezed unreadable. Narrow tables keep the
+    native selectable table."""
+    header = header or []
+    rows = rows or []
+    ncol = len(header) if header else (len(rows[0]) if rows else 1)
+    ncol = max(1, ncol)
+    cw = tl.avg_char_width_in(_FS_CELL)
+    min_needed = ncol * (_MIN_LEGIBLE_CHARS * cw + _CELL_PAD * 2)
+    return min_needed <= _USABLE_W
+
+
+def _table_figure_block(block):
+    """Wrap a too-wide table as a lazily-rasterized Figure (cached on the block).
+
+    Drawn once via ``render_table_as_figure`` (header shading + zebra) and embedded
+    as one high-res image scaled to fit entirely. The title/note are drawn inside
+    the image (self-describing when zoomed/shared), so no separate caption is
+    emitted. Reused for measuring and placing so keep-together stays consistent."""
+    cached = getattr(block, "_aeda_tablefig", None)
+    if cached is not None:
+        return cached
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    title = getattr(block, "title", None)
+    note = getattr(block, "note", None)
+
+    def _make():
+        from datascience.render_table_as_figure import render_table_as_figure
+        return render_table_as_figure(header, rows, title=title, note=note)
+
+    fig = model.Figure(make=_make, caption=None)
+    try:
+        block._aeda_tablefig = fig
+    except Exception:  # noqa: BLE001 — block may reject attributes; degrade.
+        pass
+    return fig
+
+
 def _row_height_in(cells, widths, fs) -> float:
     lh = tl.line_height_in(fs)
     maxlines = 1
@@ -365,11 +434,27 @@ def _style_cell(cell, fs, color, bold, fill) -> None:
 
 def _place_data_table(st: _PptxState, block, shaded_header=True,
                       key_value=False) -> None:
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    # Too many columns to be legible as a native table → render the whole table as
+    # one high-res picture, scaled to fit entirely (the viewer zooms to read it).
+    # KVTables (rendered here as a 2-column Campo/Valor table) are excluded: they
+    # always fit in width and stay as a selectable table.
+    if not key_value and not _table_fits_as_text(header, rows):
+        figblock = _table_figure_block(block)
+        data, _asp = _figure_bytes_cached(figblock)
+        if data is None:
+            _add_text(st, ["(tabla no disponible)"], _FS_NOTE, _MUTED,
+                      italic=True)
+            st.y += _GAP
+            return
+        _place_picture_bytes(st, data, None,
+                             max_h_in=getattr(figblock, "height_in", None),
+                             force_caption=False)
+        return
     title = getattr(block, "title", None)
     if title:
         _place_heading(st, model.Heading(title, level=2))
-    header = list(getattr(block, "header", []) or [])
-    rows = list(getattr(block, "rows", []) or [])
     fs = _FS_CELL
     widths = _col_widths(header, rows)
     header_h = _row_height_in(header, widths, fs) if header else 0.0
@@ -429,7 +514,7 @@ def _resolve_png(block):
     try:
         import matplotlib.pyplot as plt
         buf = io.BytesIO()
-        f.savefig(buf, format="png", dpi=150, bbox_inches="tight")
+        f.savefig(buf, format="png", dpi=_RASTER_DPI, bbox_inches="tight")
         buf.seek(0)
         return buf.read()
     except Exception:  # noqa: BLE001
@@ -476,12 +561,15 @@ def _figure_bytes_cached(block):
 
 
 def _place_picture_bytes(st: _PptxState, data: bytes, caption,
-                         max_h_in=None) -> None:
+                         max_h_in=None, force_caption=True) -> None:
     # Mejora 4 — every figure on a slide carries a visible caption/title. If the
     # block has no caption, fall back to the current section heading, then to a
-    # generic label, so no image is ever shown untitled.
+    # generic label, so no image is ever shown untitled. ``force_caption=False``
-    caption = (model._safe_str(caption).strip()
+    # suppresses that fallback (used for table images, whose title is inside the
-               or model._safe_str(st.last_heading).strip() or "Figura")
+    # picture) so no redundant caption is drawn.
+    caption = model._safe_str(caption).strip()
+    if not caption and force_caption:
+        caption = model._safe_str(st.last_heading).strip() or "Figura"
     w_px, h_px = _img_size_px(data)
     aspect = (h_px / w_px) if w_px else 0.66
     # Reserve the caption's REAL (possibly multi-line) height FIRST, then scale
@@ -489,9 +577,11 @@ def _place_picture_bytes(st: _PptxState, data: bytes, caption,
     # so its caption always fits on the SAME slide and no image is untitled.
     # cap_real = what _add_text consumes; cap_reserve adds the post-image gap and
     # a small cushion so the caption never spills to the next slide.
-    cap_lines = tl.wrap(caption, tl.chars_per_line(_USABLE_W, _FS_NOTE))
+    cap_lines = tl.wrap(caption, tl.chars_per_line(_USABLE_W, _FS_NOTE)) \
-    cap_real = tl.line_height_in(_FS_NOTE) * len(cap_lines) + 0.05
+        if caption else []
-    cap_reserve = cap_real + 0.05 + 0.10
+    cap_real = (tl.line_height_in(_FS_NOTE) * len(cap_lines) + 0.05) \
+        if cap_lines else 0.0
+    cap_reserve = (cap_real + 0.05 + 0.10) if cap_lines else 0.05
     max_h = _CONTENT_BOTTOM - _CONTENT_TOP
     # height_in hint (model.Figure/Image): cap the target height so a figure in a
     # keep-together Group shrinks to leave room for its heading and text.
@@ -510,7 +600,8 @@ def _place_picture_bytes(st: _PptxState, data: bytes, caption,
     st.slide.shapes.add_picture(io.BytesIO(data), Inches(left), Inches(st.y),
                                 width=Inches(target_w), height=Inches(target_h))
     st.y += target_h + 0.05
-    _add_text(st, cap_lines, _FS_NOTE, _MUTED, italic=True)
+    if cap_lines:
+        _add_text(st, cap_lines, _FS_NOTE, _MUTED, italic=True)
     st.y += _GAP
 
 
@@ -552,9 +643,11 @@ def _place_note(st: _PptxState, block) -> None:
 # WITHOUT drawing it so a Group can move whole to the next slide before drawing.
 # Over-estimating only triggers an earlier slide break, never a content cut.
 # --------------------------------------------------------------------------- #
-def _measure_heading_text(text: str, level: int) -> float:
+def _measure_heading_text(text: str, level: int, size_pt=None) -> float:
     level = max(1, min(3, int(level or 1)))
     fs = {1: _FS_H1, 2: _FS_H2, 3: _FS_H3}[level]
+    if isinstance(size_pt, (int, float)) and size_pt > 0:
+        fs = float(size_pt)
     lines = tl.wrap(tl.strip_inline_md(text), tl.chars_per_line(_USABLE_W, fs))
     return tl.line_height_in(fs) * len(lines) + 0.05 + 0.04
 
@@ -654,12 +747,16 @@ def _measure_kv_table(block) -> float:
 def _measure_data_table(block) -> float:
     """Faithful DataTable height — matches ``_place_data_table`` (title heading +
     wrapped header + every wrapped row + optional note). Keep in sync."""
+    header = list(getattr(block, "header", []) or [])
+    rows = list(getattr(block, "rows", []) or [])
+    # Mirror the placer: a too-wide table is drawn as one image, so its
+    # keep-together height is the image's, not the (squeezed) table layout's.
+    if not _table_fits_as_text(header, rows):
+        return _measure_figure_like(_table_figure_block(block))
     h = 0.0
     title = getattr(block, "title", None)
     if title:
         h += _measure_heading_text(title, 2)
-    header = list(getattr(block, "header", []) or [])
-    rows = list(getattr(block, "rows", []) or [])
     fs = _FS_CELL
     widths = _col_widths(header, rows)
     if header:
@@ -679,7 +776,8 @@ def _measure_block(st: _PptxState, block) -> float:
     try:
         if kind == "heading":
             return _measure_heading_text(getattr(block, "text", ""),
-                                         getattr(block, "level", 1))
+                                         getattr(block, "level", 1),
+                                         size_pt=getattr(block, "size_pt", None))
         if kind == "markdown":
             return _measure_markdown(block)
         if kind in ("figure", "image"):
@@ -688,6 +786,10 @@ def _measure_block(st: _PptxState, block) -> float:
             lines = tl.wrap(getattr(block, "text", ""),
                             tl.chars_per_line(_USABLE_W, _FS_NOTE))
             return tl.line_height_in(_FS_NOTE) * len(lines) + 0.05 + _GAP
+        if kind == "toc_entry":
+            lines = tl.wrap(tl.strip_inline_md(getattr(block, "label", "")),
+                            tl.chars_per_line(_USABLE_W - 0.3, _FS_BODY)) or [""]
+            return tl.line_height_in(_FS_BODY) * len(lines) + 0.05
         if kind == "kv_table":
             return _measure_kv_table(block)
         if kind == "data_table":
@@ -800,6 +902,73 @@ def _fit_group_blocks(st: _PptxState, blocks: list, avail_full: float) -> list:
     return out
 
 
+def _fit_img(width_col: float, aspect: float, max_h: float):
+    """Scale an image to ``width_col`` then clamp to ``max_h`` keeping aspect."""
+    w = width_col
+    h = w * aspect
+    if h > max_h:
+        h = max_h
+        w = (h / aspect) if aspect else width_col
+    return w, h
+
+
+def _place_group_side_by_side(st: _PptxState, block, avail_full: float) -> bool:
+    """Place a Group's table (left ~55%) next to its figure (right ~45%).
+
+    Both the table and the figure are rasterized to high-res images and placed in
+    two columns of the SAME slide; any other blocks (e.g. a heading) render full
+    width above the pair, the rest below. Returns True on success; returns False
+    (so the caller falls back to stacking) when the group has no table+figure pair
+    or the pair cannot fit side by side on one slide. Never raises by itself."""
+    blocks = getattr(block, "blocks", []) or []
+    tbl = next((b for b in blocks
+                if getattr(b, "kind", "") in ("data_table", "kv_table")), None)
+    fig = next((b for b in blocks
+                if getattr(b, "kind", "") in ("figure", "image")), None)
+    if tbl is None or fig is None:
+        return False
+    gap_col = 0.3
+    left_w = _USABLE_W * 0.55 - gap_col / 2.0
+    right_w = _USABLE_W * 0.45 - gap_col / 2.0
+    if left_w <= 1.0 or right_w <= 1.0:
+        return False
+    tdata, tasp = _figure_bytes_cached(_table_figure_block(tbl))
+    fdata, fasp = _figure_bytes_cached(fig)
+    if not tdata or not fdata:
+        return False
+    ti, fi = blocks.index(tbl), blocks.index(fig)
+    lo = min(ti, fi)
+    lead = list(blocks[:lo])
+    rest = [b for b in blocks[lo + 1:] if b is not tbl and b is not fig]
+    lead_h = sum(_measure_block(st, b) for b in lead)
+    rest_h = sum(_measure_block(st, b) for b in rest)
+    col_max_h = avail_full - lead_h - rest_h - _GAP * 2
+    if col_max_h < 1.2:
+        return False  # not enough vertical room to put the pair side by side.
+    tw, th = _fit_img(left_w, tasp, col_max_h)
+    fw, fh = _fit_img(right_w, fasp, col_max_h)
+    band = max(th, fh)
+    needed = lead_h + band + rest_h + _GAP * 2
+    if needed > avail_full:
+        return False  # taller than a whole slide even side by side → stack.
+    if needed > _remaining(st):
+        _new_slide(st, cont=True)
+    for b in lead:
+        _PLACERS.get(getattr(b, "kind", ""), _place_note)(st, b)
+    top = st.y
+    f_left = _ML + left_w + gap_col
+    st.slide.shapes.add_picture(
+        io.BytesIO(tdata), Inches(_ML + (left_w - tw) / 2.0),
+        Inches(top + (band - th) / 2.0), width=Inches(tw), height=Inches(th))
+    st.slide.shapes.add_picture(
+        io.BytesIO(fdata), Inches(f_left + (right_w - fw) / 2.0),
+        Inches(top + (band - fh) / 2.0), width=Inches(fw), height=Inches(fh))
+    st.y = top + band + _GAP
+    for b in rest:
+        _PLACERS.get(getattr(b, "kind", ""), _place_note)(st, b)
+    return True
+
+
 def _place_group(st: _PptxState, block) -> None:
     """Render a keep-together Group: move it whole to the next slide if needed."""
     blocks = getattr(block, "blocks", []) or []
@@ -810,6 +979,14 @@ def _place_group(st: _PptxState, block) -> None:
     if getattr(block, "page_break_before", False) and st.y > _CONTENT_TOP + 1e-6:
         _new_slide(st, cont=True)
     avail_full = _CONTENT_BOTTOM - _CONTENT_TOP
+    # layout="side_by_side": try table-left / figure-right on one slide; on any
+    # reason it can't, fall through to the normal stacked keep-together below.
+    if str(getattr(block, "layout", "stack")).lower() == "side_by_side":
+        try:
+            if _place_group_side_by_side(st, block, avail_full):
+                return
+        except Exception:  # noqa: BLE001 — degrade to stacking, never abort.
+            pass
     # Trim oversized tables first (keeps the chart on the same slide), then shrink
     # the figure to share the remaining room.
     blocks = _fit_group_blocks(st, blocks, avail_full)
@@ -843,6 +1020,44 @@ def _place_glossary_entry(st: _PptxState, block) -> None:
     st.y += _GAP
 
 
+def _place_toc_entry(st: _PptxState, block) -> None:
+    """Render one clickable index line and record its run as a link source.
+
+    Drawn as a bulleted line in the accent link colour; the run is recorded in
+    ``st.toc_runs`` so it later becomes a native slide-jump to the target chapter's
+    first slide. If the target is never resolved the line still shows as plain
+    (accent) text — never cut."""
+    label = tl.strip_inline_md(getattr(block, "label", "")) or ""
+    target_id = getattr(block, "target_id", "") or ""
+    fs = _FS_BODY
+    lines = tl.wrap(label, tl.chars_per_line(_USABLE_W - 0.3, fs)) or [""]
+    lh = tl.line_height_in(fs)
+    height = lh * len(lines) + 0.05
+    _ensure(st, height)
+    box = st.slide.shapes.add_textbox(
+        Inches(_ML), Inches(st.y), Inches(_USABLE_W), Inches(height))
+    tf = box.text_frame
+    tf.word_wrap = True
+    first = True
+    link_run = None
+    for idx, ln in enumerate(lines):
+        p = tf.paragraphs[0] if first else tf.add_paragraph()
+        first = False
+        r0 = p.add_run()
+        r0.text = "•  " if idx == 0 else "   "
+        r0.font.size = Pt(fs)
+        r0.font.color.rgb = _rgb(_LINK)
+        run = p.add_run()
+        run.text = ln
+        run.font.size = Pt(fs)
+        run.font.color.rgb = _rgb(_LINK)
+        if idx == 0:
+            link_run = run
+    if target_id and link_run is not None:
+        st.toc_runs.append((target_id, link_run, st.slide))
+    st.y += height
+
+
 _PLACERS = {
     "heading": _place_heading,
     "markdown": _place_markdown,
@@ -854,6 +1069,7 @@ _PLACERS = {
     "note": _place_note,
     "group": _place_group,
     "glossary_entry": _place_glossary_entry,
+    "toc_entry": _place_toc_entry,
 }
 
 
@@ -889,6 +1105,12 @@ def render_pptx(chapters: list, out_path: str, meta: dict = None) -> dict:
             st.chapter = ch
             st.chapter_slides = 0
             _new_slide(st, cont=False)
+            # Record this chapter's first slide as a link target for the cover
+            # index (keyed by id AND title, since the cover only knows titles).
+            if ch.id:
+                st.chapter_starts[ch.id] = st.slide
+            if getattr(ch, "title", ""):
+                st.chapter_starts.setdefault(ch.title, st.slide)
             for block in ch.blocks:
                 placer = _PLACERS.get(getattr(block, "kind", ""), _place_note)
                 try:
@@ -916,7 +1138,7 @@ def render_pptx(chapters: list, out_path: str, meta: dict = None) -> dict:
 
     note = f"{n_slides} slides"
     if n_links:
-        note += f" · {n_links} enlaces de glosario"
+        note += f" · {n_links} enlaces internos"
     if notes:
         note += " · " + "; ".join(notes)
     return {"path": out_path, "n_slides": n_slides, "chapters": chapters_meta,
@@ -924,19 +1146,21 @@ def render_pptx(chapters: list, out_path: str, meta: dict = None) -> dict:
 
 
 def _wire_glossary_links(st: _PptxState, notes: list) -> int:
-    """Turn each recorded term run into a native jump to its glossary slide.
+    """Apply native slide-jumps: glossary terms + the cover index.
 
-    Returns the number of links applied. A term whose only appearance is inside
+    Each in-text glossary term run jumps to its glossary entry slide, and each
-    its own glossary entry (source slide == target slide) is skipped. Never
+    cover ``TocEntry`` run jumps to its chapter's first slide. Returns the total
+    number of links applied. A run whose target is its own slide is skipped. Never
     raises."""
-    if not st.term_runs or not st.term_anchor_slide:
+    if not (st.term_runs and st.term_anchor_slide) and not (
+            st.toc_runs and st.chapter_starts):
         return 0
-    linked = 0
     try:
         from datascience.pptx_link_run_to_slide import pptx_link_run_to_slide
     except Exception as e:  # noqa: BLE001
-        notes.append(f"glosario sin enlaces: {e}")
+        notes.append(f"enlaces internos no aplicados: {e}")
         return 0
+    linked = 0
     for key, run, src_slide in st.term_runs:
         tgt = st.term_anchor_slide.get(key)
         if tgt is None or tgt is src_slide:
@@ -946,4 +1170,14 @@ def _wire_glossary_links(st: _PptxState, notes: list) -> int:
                 linked += 1
         except Exception:  # noqa: BLE001 — links are best-effort.
             pass
+    # Cover index → chapter first slide (clickable, navigable table of contents).
+    for target_id, run, src_slide in st.toc_runs:
+        tgt = st.chapter_starts.get(target_id)
+        if tgt is None or tgt is src_slide:
+            continue
+        try:
+            if pptx_link_run_to_slide(run, src_slide, tgt):
+                linked += 1
+        except Exception:  # noqa: BLE001 — links are best-effort.
+            pass
     return linked

python/functions/datascience/automatic_eda/render_quality_test.py

@@ -0,0 +1,283 @@
+"""Golden tests for the global render-quality features (issue: eda-render-quality).
+
+Covers, with executable evidence:
+  * High DPI: every embedded figure is rasterized at 220 dpi, so a phone reader
+    can zoom in and still see crisp detail.
+  * Wide table → image: a table too wide to be legible as text (e.g. a 19-column
+    df.head) is rendered as one high-res image that scales to fit entirely, while
+    a narrow table keeps its selectable-text/native-table rendering.
+  * ``Group(layout="side_by_side")``: in PPTX the table and figure are placed in
+    two columns of the same slide; in PDF the same group stacks vertically.
+  * Backward compatibility: a Group without ``layout`` defaults to ``"stack"`` and
+    a fitting table renders exactly as before.
+
+Renderers are invoked for real; PDFs are inspected with PyMuPDF and PPTX decks
+with python-pptx.
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+
+import matplotlib
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt  # noqa: E402
+
+import pytest  # noqa: E402
+
+from datascience.automatic_eda import model  # noqa: E402
+from datascience.automatic_eda.render_pdf_impl import (  # noqa: E402
+    render_pdf, _RASTER_DPI as _PDF_DPI, _table_fits_as_text as _pdf_fits)
+from datascience.automatic_eda.render_pptx_impl import (  # noqa: E402
+    render_pptx, _RASTER_DPI as _PPTX_DPI, _table_fits_as_text as _pptx_fits)
+
+
+# --------------------------------------------------------------------------- #
+# Helpers.
+# --------------------------------------------------------------------------- #
+def _simple_fig():
+    """A small, real matplotlib figure for the figure blocks."""
+    fig, ax = plt.subplots(figsize=(4, 3))
+    ax.plot([0, 1, 2, 3], [1, 3, 2, 4])
+    ax.set_title("demo")
+    return fig
+
+
+def _wide_table(n_cols=19, n_rows=5):
+    header = [f"columna_{i}" for i in range(n_cols)]
+    rows = [[f"v{r}_{c}" for c in range(n_cols)] for r in range(n_rows)]
+    return model.DataTable(header=header, rows=rows, title="Primeras filas")
+
+
+def _narrow_table():
+    return model.DataTable(header=["a", "b", "c"],
+                           rows=[["1", "2", "3"], ["4", "5", "6"]],
+                           title="Tabla estrecha")
+
+
+def _chapter(blocks, cid="cap", title="Capítulo"):
+    return [model.Chapter(id=cid, title=title, version="1.0.0", blocks=blocks)]
+
+
+# --------------------------------------------------------------------------- #
+# 1) High DPI — the unit constant and a real embedded image.
+# --------------------------------------------------------------------------- #
+def test_raster_dpi_is_high_both_renderers():
+    assert _PDF_DPI >= 200, "el DPI del PDF debe ser alto (>=200)"
+    assert _PPTX_DPI >= 200, "el DPI del PPTX debe ser alto (>=200)"
+
+
+def test_pdf_embedded_figure_is_high_resolution(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "fig.pdf")
+    res = render_pdf(_chapter([model.Figure(make=_simple_fig, caption="demo")]),
+                     out, {"title": "T"})
+    assert res["path"] == out
+    doc = fitz.open(out)
+    try:
+        widths = []
+        for page in doc:
+            for img in page.get_images(full=True):
+                xref = img[0]
+                info = doc.extract_image(xref)
+                widths.append(info.get("width", 0))
+        assert widths, "no se incrustó ninguna imagen en el PDF"
+        # A ~4" figure rasterized at 220 dpi is ~ >850 px wide. At the old 150 dpi
+        # it would be ~600 px. The high-res threshold proves the DPI bump.
+        assert max(widths) >= 800, \
+            f"la figura embebida no es de alta resolución: {max(widths)} px"
+    finally:
+        doc.close()
+
+
+# --------------------------------------------------------------------------- #
+# 2) Wide table → image (PDF and PPTX); narrow table stays text.
+# --------------------------------------------------------------------------- #
+def test_fit_criterion_flags_wide_and_keeps_narrow():
+    wide = _wide_table()
+    narrow = _narrow_table()
+    assert not _pdf_fits(wide.header, wide.rows), \
+        "una tabla de 19 columnas debería NO caber como texto en A5"
+    assert not _pptx_fits(wide.header, wide.rows), \
+        "una tabla de 19 columnas debería NO caber como tabla nativa en 16:9"
+    assert _pdf_fits(narrow.header, narrow.rows), \
+        "una tabla de 3 columnas debería caber como texto en A5"
+    assert _pptx_fits(narrow.header, narrow.rows), \
+        "una tabla de 3 columnas debería caber como tabla nativa en 16:9"
+
+
+def test_wide_table_rendered_as_image_pdf(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "wide.pdf")
+    res = render_pdf(_chapter([_wide_table()]), out, {"title": "T"})
+    assert res["path"] == out
+    doc = fitz.open(out)
+    try:
+        n_images = sum(len(page.get_images(full=True)) for page in doc)
+        text = "".join(page.get_text() for page in doc)
+    finally:
+        doc.close()
+    assert n_images >= 1, "la tabla ancha no se rasterizó como imagen en el PDF"
+    # The cells are now inside the image, not selectable text. A unique cell value
+    # must therefore NOT appear as extractable text (it lives in the picture).
+    assert "v4_18" not in text, \
+        "la tabla ancha sigue como texto seleccionable (no se hizo imagen)"
+
+
+def test_narrow_table_stays_selectable_text_pdf(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "narrow.pdf")
+    render_pdf(_chapter([_narrow_table()]), out, {"title": "T"})
+    doc = fitz.open(out)
+    try:
+        text = "".join(page.get_text() for page in doc)
+    finally:
+        doc.close()
+    # Narrow table is selectable text: its header/cells are extractable.
+    for v in ("a", "b", "c", "1", "6"):
+        assert v in text, f"la celda '{v}' debería ser texto seleccionable"
+
+
+def test_wide_table_rendered_as_picture_pptx(tmp_path):
+    pptx = pytest.importorskip("pptx")
+    from pptx.enum.shapes import MSO_SHAPE_TYPE
+    out = str(tmp_path / "wide.pptx")
+    res = render_pptx(_chapter([_wide_table()]), out, {"title": "T"})
+    assert res["path"] == out
+    prs = pptx.Presentation(out)
+    pics = sum(1 for s in prs.slides for sh in s.shapes
+               if sh.shape_type == MSO_SHAPE_TYPE.PICTURE)
+    assert pics >= 1, "la tabla ancha no se colocó como imagen en el PPTX"
+
+
+# --------------------------------------------------------------------------- #
+# 3) Group(layout="side_by_side"): two columns in PPTX, stacked in PDF.
+# --------------------------------------------------------------------------- #
+def _side_by_side_group():
+    return model.Group(
+        blocks=[model.Heading(text="Columna X", level=2),
+                _narrow_table(),
+                model.Figure(make=_simple_fig, caption="grafico")],
+        layout="side_by_side")
+
+
+def test_side_by_side_places_two_columns_pptx(tmp_path):
+    pptx = pytest.importorskip("pptx")
+    from pptx.enum.shapes import MSO_SHAPE_TYPE
+    from pptx.util import Inches
+    out = str(tmp_path / "sbs.pptx")
+    render_pptx(_chapter([_side_by_side_group()]), out, {"title": "T"})
+    prs = pptx.Presentation(out)
+    # Find the slide that holds the pair (table image + figure image).
+    centre_emu = int(Inches(13.333 / 2.0))
+    placed = False
+    for s in prs.slides:
+        lefts = [sh.left for sh in s.shapes
+                 if sh.shape_type == MSO_SHAPE_TYPE.PICTURE
+                 and sh.left is not None]
+        if len(lefts) >= 2:
+            # one picture starts in the left half, another in the right half.
+            if min(lefts) < centre_emu and max(lefts) > centre_emu:
+                placed = True
+                break
+    assert placed, \
+        "side_by_side no colocó tabla y figura en dos columnas de la misma slide"
+
+
+def test_side_by_side_stacks_in_pdf(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "sbs.pdf")
+    res = render_pdf(_chapter([_side_by_side_group()]), out, {"title": "T"})
+    assert res["path"] == out and res["n_pages"] >= 1
+    doc = fitz.open(out)
+    try:
+        n_images = sum(len(page.get_images(full=True)) for page in doc)
+        text = "".join(page.get_text() for page in doc)
+    finally:
+        doc.close()
+    # PDF stacks: the narrow table stays selectable text (1 of its cells is
+    # extractable) and the figure is the single embedded image — not a 2-column
+    # pair of pictures like PPTX.
+    assert n_images == 1, "el PDF no debería usar el layout de dos imágenes"
+    assert "Columna X" in text and "1" in text, \
+        "la tabla del grupo debería seguir como texto apilado en el PDF"
+
+
+# --------------------------------------------------------------------------- #
+# 4) Backward compatibility — default layout stacks, fitting table unchanged.
+# --------------------------------------------------------------------------- #
+def test_group_default_layout_is_stack():
+    g = model.Group(blocks=[_narrow_table()])
+    assert g.layout == "stack", "el layout por defecto debe ser 'stack'"
+
+
+# --------------------------------------------------------------------------- #
+# 5) Clickable cover index ("Índice") → chapter first page/slide.
+# --------------------------------------------------------------------------- #
+def _doc_with_index():
+    portada = model.Chapter(id="portada", title="Portada", version="1.0.0",
+                            blocks=[model.Heading(text="Índice", level=2),
+                                    model.TocEntry(label="Distribuciones",
+                                                   target_id="Distribuciones")])
+    cap = model.Chapter(id="num", title="Distribuciones", version="1.0.0",
+                        blocks=[model.Markdown(text="contenido del capítulo")])
+    return [portada, cap]
+
+
+def test_cover_index_is_clickable_pdf(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "idx.pdf")
+    res = render_pdf(_doc_with_index(), out, {"title": "T"})
+    assert res["path"] == out
+    doc = fitz.open(out)
+    try:
+        # The cover (page 0) must carry a GOTO link jumping to a later page.
+        goto = [lk for lk in doc[0].get_links()
+                if lk.get("kind") == fitz.LINK_GOTO and lk.get("page", 0) > 0]
+    finally:
+        doc.close()
+    assert goto, "el índice de la portada no produjo enlaces clicables en el PDF"
+
+
+def test_cover_index_shows_heading_pdf(tmp_path):
+    fitz = pytest.importorskip("fitz")
+    out = str(tmp_path / "idxh.pdf")
+    render_pdf(_doc_with_index(), out, {"title": "T"})
+    doc = fitz.open(out)
+    try:
+        text = "".join(page.get_text() for page in doc)
+    finally:
+        doc.close()
+    assert "Índice" in text, "la portada no muestra el encabezado 'Índice'"
+    assert "Este informe incluye" not in text, \
+        "la portada aún muestra el texto antiguo 'Este informe incluye'"
+
+
+def test_cover_index_is_clickable_pptx(tmp_path):
+    pptx = pytest.importorskip("pptx")
+    out = str(tmp_path / "idx.pptx")
+    render_pptx(_doc_with_index(), out, {"title": "T"})
+    prs = pptx.Presentation(out)
+    cover_xml = prs.slides[0]._element.xml
+    assert "hlinksldjump" in cover_xml, \
+        "el índice de la portada no produjo un salto de slide nativo en el PPTX"
+
+
+def test_default_group_renders_like_before_pptx(tmp_path):
+    pptx = pytest.importorskip("pptx")
+    from pptx.enum.shapes import MSO_SHAPE_TYPE
+    out = str(tmp_path / "stack.pptx")
+    grp = model.Group(blocks=[model.Heading(text="Y", level=2),
+                              _narrow_table(),
+                              model.Figure(make=_simple_fig, caption="g")])
+    render_pptx(_chapter([grp]), out, {"title": "T"})
+    prs = pptx.Presentation(out)
+    # Stacked group: the narrow table is a NATIVE table (selectable), and there is
+    # exactly one picture (the figure) — not the two-image side-by-side layout.
+    n_tables = sum(1 for s in prs.slides for sh in s.shapes if sh.has_table)
+    n_pics = sum(1 for s in prs.slides for sh in s.shapes
+                 if sh.shape_type == MSO_SHAPE_TYPE.PICTURE)
+    assert n_tables >= 1, "el grupo apilado debería usar una tabla nativa"
+    assert n_pics == 1, "el grupo apilado no debería duplicar imágenes"

python/functions/datascience/build_boxplots_figure.md

@@ -0,0 +1,125 @@
+---
+id: build_boxplots_figure_py_datascience
+name: build_boxplots_figure
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def build_boxplots_figure(boxes: list, title: str = \"\", max_boxes: int = 12) -> \"matplotlib.figure.Figure\""
+description: "Construye una unica figura matplotlib con boxplots de Tukey HORIZONTALES (uno por columna) usando ax.bxp: caja Q1-Q3, bigotes hasta 1.5*IQR, linea de mediana y puntos atipicos. Consume la salida de build_boxplot_stats (un dict box por columna, leido con .get) mas una lista opcional de outliers crudos por columna; si vienen los dibuja como puntos (showfliers), si no marca solo box[min]/box[max] cuando hay outliers de cola (igual que num_distr). Dibuja como mucho max_boxes cajas (las primeras, ya ordenadas por contaminacion por el caller) y avisa de la truncacion con (mostrando N de M). Backend Agg sin pyplot global; alto adaptativo al nº de cajas. Defensiva: omite entradas invalidas y NUNCA lanza — sin cajas validas devuelve una figura placeholder (sin boxplots). Es la version small-multiples del capitulo num_distr para responder que columnas tienen mas outliers de un vistazo."
+tags: [eda, outliers, boxplot, tukey, iqr, bxp, matplotlib, figure, visualization, small-multiples, datascience, impure]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: [matplotlib]
+example: |
+  from datascience.build_boxplot_stats import build_boxplot_stats
+  from datascience.build_boxplots_figure import build_boxplots_figure
+  boxes = [
+      {"name": "ingresos", "box": build_boxplot_stats({"min": 1.0, "max": 9e3,
+          "p25": 1e3, "median": 2e3, "p75": 3e3, "n_outliers": 7}), "fliers": None},
+      {"name": "edad", "box": build_boxplot_stats({"min": 0.0, "max": 99.0,
+          "p25": 25.0, "median": 38.0, "p75": 52.0}), "fliers": None},
+  ]
+  fig = build_boxplots_figure(boxes, title="Outliers por columna", max_boxes=12)
+tested: true
+tests:
+  - "test_returns_figure_with_axes"
+  - "test_empty_list_returns_placeholder_figure"
+  - "test_invalid_box_is_skipped_not_raised"
+  - "test_all_invalid_returns_placeholder"
+  - "test_raw_fliers_are_drawn"
+  - "test_max_boxes_truncates_and_does_not_raise"
+test_file_path: "python/functions/datascience/build_boxplots_figure_test.py"
+file_path: "python/functions/datascience/build_boxplots_figure.py"
+params:
+  - name: boxes
+    desc: "Lista de dicts, cada uno {\"name\": str, \"box\": dict, \"fliers\": list|None}. box es EXACTAMENTE la salida de build_boxplot_stats (claves leidas con .get: q1, median, q3, whisker_lo, whisker_hi, min, max, has_low_outliers, has_high_outliers, lower_fence, upper_fence, n_outliers). fliers es la lista opcional de outliers crudos: si viene se dibuja como puntos; si es None/ausente solo se marcan los extremos box[min]/box[max] cuando hay outliers de cola. Entradas que no son dict, sin box dict, o sin q1/median/q3 se omiten. El caller las pasa ya ordenadas por contaminacion (la mayor primera)."
+  - name: title
+    desc: "Titulo de la figura (fig.suptitle, alineado a la izquierda). Vacio => sin titulo. Si len(boxes) > max_boxes se le anade una nota \"(mostrando N de M)\" para que la truncacion no sea silenciosa. Default \"\"."
+  - name: max_boxes
+    desc: "Numero maximo de cajas a dibujar (las primeras de la lista). Default 12. Un valor no entero o <= 0 cae a 12. Si la lista trae mas entradas, las sobrantes se descartan pero se reporta en el titulo con (mostrando N de M)."
+output: "Un matplotlib.figure.Figure (figsize 7.0 x alto adaptativo = max(2.0, 0.5*n + 1.0), dpi 150) con un unico Axes que apila boxplots horizontales de Tukey (ax.bxp, orientation=horizontal con fallback vert=False), uno por columna valida, de arriba a abajo en el orden recibido. Cada caja: relleno #9ec6df, borde/bigotes/caps #5b8aa6, mediana #2e8b57, atipicos #c0392b. Etiquetas del eje Y = nombres de columna; eje X etiquetado \"valor\". Outliers dibujados desde fliers crudos (showfliers) o, si faltan, marcados en box[min]/box[max] segun has_low/high_outliers. Si no queda ninguna caja valida (lista vacia o todas invalidas) devuelve una Figure placeholder con texto centrado \"(sin boxplots)\"; cualquier error inesperado se captura y devuelve una Figure con el mensaje de error. NUNCA lanza. El caller rasteriza/cierra la figura; la funcion no la muestra ni la guarda."
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience.build_boxplot_stats import build_boxplot_stats
+from datascience.build_boxplots_figure import build_boxplots_figure
+
+# Un `box` por columna numérica, derivado del sub-bloque `numeric` del profile
+# (salida de describe_numeric). El caller los pasa ya ordenados por outlier_pct.
+boxes = [
+    {
+        "name": "ingresos",
+        "box": build_boxplot_stats({
+            "min": 1.0, "max": 9000.0,
+            "p25": 1000.0, "median": 2000.0, "p75": 3000.0,
+            "n_outliers": 7,
+        }),
+        "fliers": None,  # valores crudos desconocidos -> se marca solo el extremo.
+    },
+    {
+        "name": "edad",
+        "box": build_boxplot_stats({
+            "min": 0.0, "max": 99.0,
+            "p25": 25.0, "median": 38.0, "p75": 52.0,
+        }),
+        "fliers": [88.0, 95.0, 99.0],  # outliers crudos -> se dibujan como puntos.
+    },
+]
+
+fig = build_boxplots_figure(boxes, title="Outliers por columna", max_boxes=12)
+
+# El renderer del informe lo rasteriza; aquí solo persistimos para inspección.
+fig.savefig("/tmp/boxplots.png")
+```
+
+## Cuando usarla
+
+Úsala en el capítulo de outliers de un informe EDA cuando quieras comparar de un
+vistazo *qué columnas están más contaminadas por valores atípicos*: a diferencia
+de `num_distr` (que dibuja un histograma+boxplot por columna en figuras
+separadas), aquí apilas todos los boxplots horizontales en **una sola figura**
+(small multiples). Primero deriva el `box` de cada columna con
+`build_boxplot_stats`, ordénalas por `outlier_pct` descendente, envuélvelas como
+`{"name", "box", "fliers"}` y pásaselas. Si tienes los valores crudos fuera de
+las vallas, métele la lista `fliers` y se dibujarán como puntos; si no, la
+función marca solo los extremos `min`/`max` cuando hay cola.
+
+## Gotchas
+
+- **Impura por matplotlib.** Toca la maquinaria de render. Usa el backend `Agg`
+  y la API orientada a objetos `Figure`/`add_subplot` — NUNCA `pyplot.*` aquí,
+  para no tocar el estado global ni filtrar figuras entre llamadas. `pyplot` NO
+  es thread-safe; esta función construye el `Figure` directamente, así que es
+  segura de llamar en bucle desde el renderer.
+- **El caller cierra la figura.** Devuelve el `Figure` pero no lo muestra ni lo
+  guarda. Quien la consume debe rasterizarla y luego liberarla
+  (`matplotlib.pyplot.close(fig)`) para no acumular memoria en lotes grandes.
+- **`fliers` opcional, semántica distinta.** Si pasas la lista de outliers
+  crudos se dibujan todos como puntos (`showfliers=True`). Si es `None`/ausente
+  los valores son desconocidos y solo se marca un punto en `box["min"]` /
+  `box["max"]` cuando `has_low_outliers` / `has_high_outliers` — mismo criterio
+  que `num_distr`. No inventes fliers a partir del profile: el `box` no trae los
+  valores crudos, solo si los extremos superan las vallas.
+- **API de orientación de `ax.bxp`.** matplotlib reciente usa
+  `orientation="horizontal"`; las versiones antiguas usan `vert=False`. La
+  función prueba la primera y cae a la segunda en `except TypeError`, así que
+  funciona en ambas. Si `bxp` falla del todo, el Axes degrada a un texto
+  "(boxplot no disponible)" en vez de propagar.
+- **Truncación visible.** `max_boxes` (default 12) limita el nº de cajas para que
+  ninguna se solape; si la lista trae más, las sobrantes se descartan pero se
+  avisa en el título con "(mostrando N de M)". Pasa las columnas ya ordenadas por
+  contaminación para que las descartadas sean las menos relevantes.
+- **Defensiva, nunca lanza.** Lista vacía, entradas no-dict, sin `box`, o sin
+  `q1`/`median`/`q3` se omiten sin propagar; sin cajas válidas devuelve un
+  placeholder "(sin boxplots)" y cualquier error inesperado se captura en una
+  figura con el texto del error. No envuelvas la llamada en try/except por miedo
+  a un raise — no lo hay.

python/functions/datascience/build_boxplots_figure.py

@@ -0,0 +1,250 @@
+"""Impure EDA helper: a single figure of horizontal Tukey boxplots (`eda` group).
+
+Draws, in one ``matplotlib.figure.Figure``, a stack of horizontal Tukey boxplots
+(one per column) using ``ax.bxp``: each carries its box (Q1–Q3), whiskers (up to
+1.5·IQR), the median line and its outlier points. It consumes the output of the
+pure registry function ``build_boxplot_stats`` (one ``box`` dict per column) plus
+an optional list of raw outlier values per column; it never recomputes anything.
+
+It is the "small-multiples" companion of ``num_distr`` (which draws one
+histogram+boxplot per column): here every column shares a single figure so the
+caller can show, at a glance, *which* columns are the most contaminated by
+outliers (the caller passes them already ordered by contamination).
+
+Impure because it touches matplotlib's rendering machinery. It uses the headless
+Agg backend and the object-oriented ``Figure`` API (no ``pyplot``) so it leaks no
+global state and is safe to call repeatedly from a report renderer. It is fully
+defensive and NEVER raises: invalid entries are skipped and, if nothing valid
+remains, it returns a placeholder figure carrying a centered "(sin boxplots)".
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+from matplotlib.figure import Figure  # noqa: E402
+
+# Blue palette shared with the ``num_distr`` chapter so the report stays coherent.
+_BOX_FACE = "#9ec6df"   # box fill.
+_BOX_EDGE = "#5b8aa6"   # box / whisker / cap border.
+_MEDIAN = "#2e8b57"     # median line (sea green).
+_OUTLIER = "#c0392b"    # outlier points (soft red).
+# Muted gray for the placeholder / fallback message text.
+_MUTED_TEXT = "#5f6b7a"
+# Soft red for the error fallback message.
+_ERROR_TEXT = "#b00020"
+
+
+def _num(value):
+    """Coerce ``value`` to float defensively; None for None/bool/non-numeric/NaN."""
+    # bool is a subclass of int; a stat value is never a real bool, so treat
+    # True/False as missing instead of silently coercing to 1.0/0.0.
+    if value is None or isinstance(value, bool):
+        return None
+    try:
+        f = float(value)
+    except (TypeError, ValueError):
+        return None
+    if f != f:  # NaN guard.
+        return None
+    return f
+
+
+def _placeholder_figure(message: str, color: str = _MUTED_TEXT) -> "Figure":
+    """Return a fallback ``Figure`` carrying a single centered message."""
+    fig = Figure(figsize=(7.0, 2.4), dpi=150)
+    ax = fig.add_subplot(111)
+    ax.axis("off")
+    ax.text(
+        0.5,
+        0.5,
+        message,
+        ha="center",
+        va="center",
+        fontsize=12,
+        color=color,
+        wrap=True,
+        transform=ax.transAxes,
+    )
+    fig.tight_layout()
+    return fig
+
+
+def build_boxplots_figure(
+    boxes: list,
+    title: str = "",
+    max_boxes: int = 12,
+) -> "matplotlib.figure.Figure":
+    """Build one figure of stacked horizontal Tukey boxplots (one per column).
+
+    For each entry the function builds a ``bxp`` stats record (``med, q1, q3,
+    whislo, whishi, fliers, label``) from its ``box`` sub-dict (the output of
+    ``build_boxplot_stats``) and draws all of them as horizontal boxplots sharing
+    the X axis, top-to-bottom in the order received (the caller is expected to
+    pass them already sorted by contamination).
+
+    Outliers are shown two ways:
+
+    - If an entry carries a ``fliers`` list (the raw out-of-fence values), they
+      are drawn as red points via ``ax.bxp(..., showfliers=True)``.
+    - If ``fliers`` is ``None``/absent, the raw values are unknown, so only the
+      extremes are marked: a red point at ``box["min"]`` when
+      ``box["has_low_outliers"]`` and at ``box["max"]`` when
+      ``box["has_high_outliers"]`` (same convention as ``num_distr``).
+
+    The function is fully defensive and NEVER raises. Entries that are not dicts,
+    lack a ``box`` dict, or miss any of ``q1``/``median``/``q3`` are skipped. If
+    after filtering no valid box remains it returns a placeholder ``Figure`` with
+    a centered "(sin boxplots)"; any unexpected error is caught and turned into a
+    fallback figure carrying the error text. It always returns a ``Figure``.
+
+    Args:
+        boxes: List of dicts ``{"name": str, "box": dict, "fliers": list|None}``.
+            ``box`` is exactly the output of ``build_boxplot_stats`` (read with
+            ``.get``: ``q1, median, q3, whisker_lo, whisker_hi, min, max,
+            has_low_outliers, has_high_outliers, ...``). ``fliers`` is the
+            optional list of raw outlier values; when present they are plotted,
+            otherwise only the extremes are marked.
+        title: Figure title (``fig.suptitle``). Empty => no title. When the list
+            is longer than ``max_boxes`` a "(mostrando N de M)" note is appended.
+        max_boxes: Draw at most the first ``max_boxes`` entries (default 12). The
+            rest are dropped but their omission is surfaced in the title note, so
+            the truncation is never silent.
+
+    Returns:
+        A ``matplotlib.figure.Figure`` with a single Axes holding the horizontal
+        boxplots (height adaptive to the box count so none overlap). The caller is
+        responsible for rasterizing/closing it; this function never shows nor
+        saves it.
+    """
+    try:
+        if not isinstance(boxes, (list, tuple)) or len(boxes) == 0:
+            return _placeholder_figure("(sin boxplots)")
+
+        total = len(boxes)
+
+        # Cap the number of boxes; tolerate a non-int / non-positive max_boxes.
+        try:
+            cap = int(max_boxes)
+        except (TypeError, ValueError):
+            cap = 12
+        if cap <= 0:
+            cap = 12
+        candidates = list(boxes)[:cap]
+
+        stats_list = []        # bxp stats records, in draw order.
+        labels = []            # Y tick labels (column names).
+        manual_markers = []    # (position, box) for entries without raw fliers.
+        any_fliers = False     # whether to enable showfliers in the bxp call.
+
+        for entry in candidates:
+            if not isinstance(entry, dict):
+                continue
+            box = entry.get("box")
+            if not isinstance(box, dict):
+                continue
+
+            q1 = _num(box.get("q1"))
+            med = _num(box.get("median"))
+            q3 = _num(box.get("q3"))
+            # Without the three quartiles a boxplot cannot be drawn — skip it.
+            if q1 is None or med is None or q3 is None:
+                continue
+
+            # Whisker extremes fall back to the quartiles when missing.
+            whislo = _num(box.get("whisker_lo"))
+            whishi = _num(box.get("whisker_hi"))
+            if whislo is None:
+                whislo = q1
+            if whishi is None:
+                whishi = q3
+
+            name = entry.get("name")
+            label = "" if name is None else str(name)
+
+            position = len(stats_list) + 1  # bxp positions are 1-indexed.
+            fliers_raw = entry.get("fliers")
+            if isinstance(fliers_raw, (list, tuple)):
+                fliers = [v for v in (_num(x) for x in fliers_raw) if v is not None]
+                if fliers:
+                    any_fliers = True
+            else:
+                # Raw values unknown: draw no bxp fliers, mark min/max by hand.
+                fliers = []
+                manual_markers.append((position, box))
+
+            stats_list.append({
+                "med": med,
+                "q1": q1,
+                "q3": q3,
+                "whislo": whislo,
+                "whishi": whishi,
+                "fliers": fliers,
+                "label": label,
+            })
+            labels.append(label)
+
+        if not stats_list:
+            return _placeholder_figure("(sin boxplots)")
+
+        n = len(stats_list)
+        positions = list(range(1, n + 1))
+
+        # Height grows with the box count so none of them overlap.
+        height = max(2.0, 0.5 * n + 1.0)
+        fig = Figure(figsize=(7.0, height), dpi=150)
+        ax = fig.add_subplot(111)
+
+        bxp_kw = dict(
+            showfliers=any_fliers, widths=0.5, patch_artist=True,
+            boxprops={"facecolor": _BOX_FACE, "edgecolor": _BOX_EDGE},
+            medianprops={"color": _MEDIAN, "linewidth": 1.6},
+            whiskerprops={"color": _BOX_EDGE},
+            capprops={"color": _BOX_EDGE},
+            flierprops={"marker": "o", "markersize": 3.5,
+                        "markerfacecolor": _OUTLIER, "markeredgecolor": _OUTLIER,
+                        "linestyle": "none"})
+        try:
+            # ``orientation`` is the current API; older matplotlib uses ``vert``.
+            try:
+                ax.bxp(stats_list, positions=positions,
+                       orientation="horizontal", **bxp_kw)
+            except TypeError:
+                ax.bxp(stats_list, positions=positions, vert=False, **bxp_kw)
+        except Exception:  # noqa: BLE001 — never let bxp kill the whole figure.
+            ax.text(0.5, 0.5, "(boxplot no disponible)", ha="center",
+                    va="center", fontsize=10, color=_MUTED_TEXT,
+                    transform=ax.transAxes)
+
+        # For entries without raw fliers, mark only the out-of-fence extremes.
+        for position, box in manual_markers:
+            mn = _num(box.get("min"))
+            mx = _num(box.get("max"))
+            if box.get("has_low_outliers") and mn is not None:
+                ax.plot([mn], [position], marker="o", markersize=3.5,
+                        color=_OUTLIER, zorder=5)
+            if box.get("has_high_outliers") and mx is not None:
+                ax.plot([mx], [position], marker="o", markersize=3.5,
+                        color=_OUTLIER, zorder=5)
+
+        # Pin the Y tick labels explicitly so they work across matplotlib
+        # versions regardless of whether ``bxp`` consumed the ``label`` key.
+        ax.set_yticks(positions)
+        ax.set_yticklabels(labels, fontsize=8)
+        ax.set_xlabel("valor", fontsize=9)
+        ax.tick_params(labelsize=7)
+        ax.margins(y=0.15)
+        for spine in ("top", "right"):
+            ax.spines[spine].set_visible(False)
+
+        # Surface truncation in the title instead of silently dropping boxes.
+        note = f"(mostrando {n} de {total})" if total > cap else ""
+        heading = "  ".join(p for p in (title, note) if p)
+        if heading:
+            fig.suptitle(heading, fontsize=12, x=0.02, ha="left")
+
+        fig.tight_layout()
+        return fig
+    except Exception as exc:  # noqa: BLE001 — never raise from a figure builder.
+        return _placeholder_figure(
+            f"error al dibujar boxplots: {exc}", color=_ERROR_TEXT)

python/functions/datascience/build_boxplots_figure_test.py

@@ -0,0 +1,109 @@
+"""Tests para build_boxplots_figure (boxplots horizontales de Tukey, grupo eda).
+
+Usa el backend Agg sin display; no muestra ni guarda figuras. Cada test cierra
+explícitamente la Figure construida (matplotlib.pyplot.close) para no acumular
+estado entre tests.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.figure import Figure  # noqa: E402
+
+from build_boxplots_figure import build_boxplots_figure
+
+
+def _box(name, q1, median, q3, mn, mx, low=False, high=False, fliers=None):
+    """Construye una entrada {name, box, fliers} con un box estilo build_boxplot_stats."""
+    iqr = q3 - q1
+    return {
+        "name": name,
+        "box": {
+            "q1": q1,
+            "median": median,
+            "q3": q3,
+            "iqr": iqr,
+            "lower_fence": q1 - 1.5 * iqr,
+            "upper_fence": q3 + 1.5 * iqr,
+            "whisker_lo": max(mn, q1 - 1.5 * iqr),
+            "whisker_hi": min(mx, q3 + 1.5 * iqr),
+            "min": mn,
+            "max": mx,
+            "has_low_outliers": low,
+            "has_high_outliers": high,
+            "n_outliers": 0,
+        },
+        "fliers": fliers,
+    }
+
+
+def test_returns_figure_with_axes():
+    boxes = [
+        _box("edad", 10.0, 25.0, 40.0, 1.0, 100.0, high=True),
+        _box("ingresos", 100.0, 200.0, 300.0, 50.0, 400.0),
+        _box("score", -1.0, 0.0, 1.0, -5.0, 5.0, low=True, high=True),
+    ]
+    fig = build_boxplots_figure(boxes, title="Boxplots", max_boxes=12)
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    # Tres cajas -> tres etiquetas en el eje Y.
+    ax = fig.axes[0]
+    assert len(ax.get_yticks()) == 3
+    plt.close(fig)
+
+
+def test_empty_list_returns_placeholder_figure():
+    fig = build_boxplots_figure([], title="vacío")
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_invalid_box_is_skipped_not_raised():
+    boxes = [
+        {"name": "rota", "box": {"q1": None, "median": None, "q3": None}},
+        {"name": "sin_box"},                         # falta la clave box.
+        "no_es_dict",                                 # entrada no-dict.
+        _box("buena", 1.0, 2.0, 3.0, 0.0, 10.0, high=True),
+    ]
+    fig = build_boxplots_figure(boxes)
+    assert isinstance(fig, Figure)
+    ax = fig.axes[0]
+    # Solo la caja válida sobrevive al filtrado.
+    assert len(ax.get_yticks()) == 1
+    plt.close(fig)
+
+
+def test_all_invalid_returns_placeholder():
+    boxes = [
+        {"name": "a", "box": {"q1": None, "median": 1.0, "q3": 2.0}},
+        {"name": "b"},
+    ]
+    fig = build_boxplots_figure(boxes)
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_raw_fliers_are_drawn():
+    boxes = [
+        _box("con_fliers", 10.0, 20.0, 30.0, 5.0, 200.0,
+             high=True, fliers=[150.0, 180.0, 200.0]),
+    ]
+    fig = build_boxplots_figure(boxes)
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_max_boxes_truncates_and_does_not_raise():
+    boxes = [_box(f"c{i}", float(i), float(i + 1), float(i + 2),
+                  float(i - 5), float(i + 10)) for i in range(20)]
+    fig = build_boxplots_figure(boxes, title="muchos", max_boxes=5)
+    assert isinstance(fig, Figure)
+    ax = fig.axes[0]
+    # Solo se dibujan las primeras 5 cajas.
+    assert len(ax.get_yticks()) == 5
+    plt.close(fig)

python/functions/datascience/categorical_top_bar_figure.md

@@ -0,0 +1,111 @@
+---
+id: categorical_top_bar_figure_py_datascience
+name: categorical_top_bar_figure
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def categorical_top_bar_figure(top: list, n_distinct: int = 0, title: str = \"\", top_k: int = 6, n_rows=None) -> \"matplotlib.figure.Figure\""
+description: "Construye una figura matplotlib de barras horizontales de las top_k categorías más frecuentes de una columna categórica, con la mayor arriba y agregando el resto en una barra gris \"Otros (N categorías)\". Contrato de entrada idéntico a categorical_top_pie_figure (swap directo donut↔barras): consume el bloque `top` de summarize_categorical y devuelve un matplotlib.figure.Figure listo para rasterizar por el renderer del informe EDA. Backend Agg sin pyplot global; defensivo total ante top vacío/None, nunca lanza."
+tags: [eda, categorical, bar, barh, matplotlib, figure, visualization, datascience, impure]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: [matplotlib]
+example: |
+  from categorical_top_bar_figure import categorical_top_bar_figure
+  top = [
+      {"value": "rojo", "count": 40, "pct": 0.4},
+      {"value": "azul", "count": 30, "pct": 0.3},
+      {"value": "verde", "count": 20, "pct": 0.2},
+  ]
+  fig = categorical_top_bar_figure(top, n_distinct=12, title="color", top_k=6, n_rows=100)
+tested: true
+tests:
+  - "test_returns_figure"
+  - "test_ten_items_topk_six_yields_seven_bars"
+  - "test_empty_top_does_not_raise_and_returns_figure"
+  - "test_long_value_truncated"
+  - "test_none_value_and_none_count_are_handled"
+  - "test_n_rows_adds_exact_others_bar"
+test_file_path: "python/functions/datascience/categorical_top_bar_figure_test.py"
+file_path: "python/functions/datascience/categorical_top_bar_figure.py"
+params:
+  - name: top
+    desc: "Lista de dicts {value, count, pct} ordenada de mayor a menor por count (salida del bloque `top` de summarize_categorical). Puede venir vacía o con dicts incompletos: items no-dict, sin count, con count None o count <= 0 se descartan. value None se admite (etiqueta vacía)."
+  - name: n_distinct
+    desc: "Nº total de categorías distintas de la columna. Etiqueta la barra agregada como \"Otros (n_distinct - top_k)\" (mínimo 0). Si no supera el nº de barras mostradas, se usa el overflow real de `top` como nº de categorías agregadas. Default 0."
+  - name: title
+    desc: "Título de la figura (nombre de la columna). Se trunca a ~48 chars con elipsis si es muy largo. Default \"\" (sin título)."
+  - name: top_k
+    desc: "Nº máximo de barras explícitas. Default 6. La barra \"Otros\" no cuenta contra este límite. Con top_k <= 0 se muestra al menos la categoría mayor."
+  - name: n_rows
+    desc: "Opcional. Total de filas del dataset. Si se da y la suma de counts mostrados < n_rows, la barra \"Otros\" usa (n_rows - suma_mostrada) como count para que sea exacta respecto al total real. Si se omite, \"Otros\" usa la suma de counts fuera del top_k mostrado (solo cuando top trae más de top_k items). Default None."
+output: "Un matplotlib.figure.Figure (figsize 6.4 x altura escalada con el nº de barras, dpi 150) con un Axes de barras horizontales: la categoría más frecuente arriba, la barra gris \"Otros (N categorías)\" abajo, cada barra anotada con su conteo y porcentaje al final y etiquetas de categoría (yticklabels) truncadas a ~22 chars. Si no hay counts válidos devuelve igualmente una Figure con un texto centrado \"sin datos categóricos\" (nunca lanza); cualquier error inesperado cae a una Figure con el texto del error. El caller rasteriza/cierra la figura; la función no la muestra ni la guarda."
+---
+
+## Ejemplo
+
+```python
+from categorical_top_bar_figure import categorical_top_bar_figure
+
+# `top` es la salida del bloque "top" de summarize_categorical (ya ordenado desc).
+top = [
+    {"value": "rojo", "count": 40, "pct": 0.40},
+    {"value": "azul", "count": 30, "pct": 0.30},
+    {"value": "verde", "count": 20, "pct": 0.20},
+    {"value": "amarillo", "count": 5, "pct": 0.05},
+]
+
+fig = categorical_top_bar_figure(
+    top,
+    n_distinct=12,            # 12 categorías distintas en total
+    title="color_producto",
+    top_k=6,                  # hasta 6 barras explícitas
+    n_rows=100,               # "Otros" = 100 - 95 = 5, sobre 8 categorías agregadas
+)
+
+# El renderer del informe lo rasteriza; aquí solo persistimos para inspección.
+fig.savefig("/tmp/barras_color.png")
+```
+
+## Cuando usarla
+
+Úsala dentro de un informe EDA cuando quieras comparar **magnitudes** de las
+categorías dominantes de una columna categórica: qué categoría manda y por
+cuánto frente a las siguientes. Pásale directamente el bloque `top` de
+`summarize_categorical` (ya ordenado de mayor a menor) más `n_distinct` para que
+la barra "Otros" indique cuántas categorías quedan agrupadas. Es el clon "de
+barras" del donut `categorical_top_pie_figure` con **contrato de entrada
+idéntico**: puedes intercambiar una por otra sin tocar el caller. Elige barras
+cuando importe comparar tamaños exactos; el donut cuando importe la proporción
+del total.
+
+## Gotchas
+
+- **Impura por matplotlib.** Toca la maquinaria de render. Usa el backend `Agg`
+  y la API orientada a objetos `Figure`/`add_subplot` — NUNCA `pyplot.*` aquí,
+  para no tocar el estado global ni filtrar figuras entre llamadas. `pyplot` NO
+  es thread-safe; esta función evita ese riesgo construyendo el `Figure`
+  directamente, así que es segura de llamar en bucle desde el renderer.
+- **El caller cierra la figura.** La función devuelve el `Figure` pero no lo
+  muestra ni lo guarda. Quien la consume debe rasterizarla y luego liberarla
+  (`fig.clf()` / `matplotlib.pyplot.close(fig)` si se usó pyplot en el caller)
+  para no acumular memoria en lotes grandes de columnas.
+- **`barh` dibuja de abajo arriba.** La categoría más frecuente va arriba porque
+  el orden de display se invierte antes de plotear; la barra "Otros" queda
+  siempre al fondo. No reordenes `top` esperando otro layout: la función asume
+  que ya viene ordenado desc por count.
+- **Magnitud exacta de "Otros" solo con `n_rows`.** Sin `n_rows`, la barra
+  "Otros" se calcula con el overflow presente en `top`; si `top` ya viene
+  recortado a `top_k` por el productor, no habrá "Otros" aunque existan más
+  categorías. Pasa `n_rows` (total de filas del dataset) para una barra correcta
+  respecto al total real.
+- **Defensiva, nunca lanza.** `top=[]`, `value=None`, `count=None` o counts no
+  numéricos se manejan sin error: en el peor caso devuelve una `Figure` con
+  "sin datos categóricos", y cualquier excepción inesperada cae a una `Figure`
+  con el texto del error. No envuelvas la llamada en try/except por miedo a un
+  raise — no lo hay.

python/functions/datascience/categorical_top_bar_figure.py

@@ -0,0 +1,233 @@
+"""Impure EDA helper: horizontal bar figure of the most common categories (`eda` group).
+
+Builds a horizontal bar chart of the ``top_k`` most frequent categories of a
+categorical column, folding everything else into a single gray
+"Otros (N categorías)" bar. The most frequent category sits at the top, each bar
+labelled with its count (and percentage) at the end. Returns a ready-to-rasterize
+``matplotlib.figure.Figure``; it never shows nor saves it.
+
+This is the "magnitude" twin of ``categorical_top_pie_figure``: identical input
+contract (same ``top``/``n_distinct``/``title``/``top_k``/``n_rows`` signature) so
+it can be swapped in directly, but it communicates comparable magnitudes via bars
+instead of proportions via wedges.
+
+Impure because it touches matplotlib's rendering machinery. It uses the headless
+Agg backend and the object-oriented ``Figure`` API (no ``pyplot``) so it leaks no
+global state and is safe to call repeatedly from a report renderer.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+from matplotlib.figure import Figure  # noqa: E402
+
+
+# Gray reserved for the aggregated "Otros" bar.
+_OTHER_COLOR = "#9e9e9e"
+# Muted gray for secondary text (title fallback, no-data message).
+_MUTED_TEXT = "#5f6b7a"
+# Soft red for the error fallback message.
+_ERROR_TEXT = "#b00020"
+# Pleasant, colour-blind-friendly qualitative palette for the explicit bars.
+_PALETTE = [
+    "#4C72B0",
+    "#DD8452",
+    "#55A868",
+    "#C44E52",
+    "#8172B3",
+    "#937860",
+    "#DA8BC3",
+    "#8C8C8C",
+    "#CCB974",
+    "#64B5CD",
+]
+
+
+def _truncate(text, width: int = 22) -> str:
+    """Truncate ``text`` to ``width`` chars, appending an ellipsis if cut."""
+    s = "" if text is None else str(text)
+    if len(s) <= width:
+        return s
+    if width <= 1:
+        return s[:width]
+    return s[: width - 1] + "…"
+
+
+def _message_figure(message: str, color: str = _MUTED_TEXT, title: str = "") -> "Figure":
+    """Return a fallback ``Figure`` carrying a single centered message."""
+    fig = Figure(figsize=(6.4, 4.0), dpi=150)
+    ax = fig.add_subplot(111)
+    ax.axis("off")
+    ax.text(
+        0.5,
+        0.5,
+        message,
+        ha="center",
+        va="center",
+        fontsize=12,
+        color=color,
+        wrap=True,
+        transform=ax.transAxes,
+    )
+    if title:
+        ax.set_title(_truncate(title, 48), fontsize=12, loc="center", pad=8)
+    fig.tight_layout()
+    return fig
+
+
+def categorical_top_bar_figure(
+    top: list,
+    n_distinct: int = 0,
+    title: str = "",
+    top_k: int = 6,
+    n_rows=None,
+) -> "matplotlib.figure.Figure":
+    """Build a horizontal bar figure of the most common categories of a column.
+
+    Renders the ``top_k`` most frequent categories as explicit horizontal bars,
+    largest at the top, and aggregates every remaining category into a single
+    gray "Otros (N categorías)" bar at the bottom. Each bar is annotated with its
+    count and percentage of the total at the end of the bar; the category names
+    are truncated Y tick labels.
+
+    The function shares the exact input contract of
+    ``categorical_top_pie_figure`` (the donut twin) so it is a drop-in swap. It is
+    fully defensive: empty input, missing/``None`` values or counts never raise.
+    When there is nothing valid to draw it still returns a ``Figure`` carrying a
+    centered "sin datos categóricos" message, and any unexpected error is caught
+    and turned into a fallback ``Figure`` carrying the error text.
+
+    Args:
+        top: List of ``{value, count, pct}`` dicts, already sorted by ``count``
+            descending (the ``top`` block of ``summarize_categorical``). May be
+            empty or carry incomplete/``None`` entries; non-dict items, items
+            without a positive numeric ``count`` and ``None`` counts are skipped.
+        n_distinct: Total number of distinct categories in the column. Used to
+            label the aggregated bar as "Otros (n_distinct - top_k)" (floored at
+            0). Ignored when it does not exceed the number of shown bars.
+        title: Figure title (the column name). Truncated when too long.
+        top_k: Maximum number of explicit bars. Default 6. The "Otros" bar does
+            not count against this limit.
+        n_rows: Optional total row count of the dataset. When given and the sum of
+            shown counts is below ``n_rows``, the "Otros" bar uses
+            ``n_rows - sum_shown`` as its count so it is exact with respect to the
+            real total. When omitted, "Otros" uses the sum of the counts that fall
+            outside the shown ``top_k`` (only when ``top`` carries more than
+            ``top_k`` items).
+
+    Returns:
+        A ``matplotlib.figure.Figure`` with a single horizontal-bar Axes. The
+        caller is responsible for rasterizing/closing it.
+    """
+    try:
+        safe_title = _truncate(title, 48)
+
+        # --- Defensive parse: keep only well-formed {value, count} with count > 0.
+        cleaned = []
+        if isinstance(top, list):
+            for item in top:
+                if not isinstance(item, dict):
+                    continue
+                count = item.get("count")
+                if count is None:
+                    continue
+                try:
+                    count = float(count)
+                except (TypeError, ValueError):
+                    continue
+                if count <= 0:
+                    continue
+                cleaned.append((item.get("value"), count))
+
+        if not cleaned:
+            return _message_figure("sin datos categóricos", title=title)
+
+        # --- Split into shown bars and the aggregated remainder.
+        shown = cleaned[: max(int(top_k), 0)]
+        if not shown:  # top_k <= 0 — show at least the largest category.
+            shown = cleaned[:1]
+
+        sum_shown = sum(c for _, c in shown)
+        overflow_count = sum(c for _, c in cleaned[len(shown):])
+
+        # How many categories are folded into "Otros".
+        try:
+            nd = int(n_distinct)
+        except (TypeError, ValueError):
+            nd = 0
+        others_categories = max(nd - len(shown), 0)
+        # If n_distinct is unknown/too small, fall back to the overflow we
+        # actually have in `top` beyond the shown bars.
+        overflow_items = len(cleaned) - len(shown)
+        if others_categories == 0 and overflow_items > 0:
+            others_categories = overflow_items
+
+        # Count attributed to the "Otros" bar.
+        others_count = 0.0
+        if n_rows is not None:
+            try:
+                total_rows = float(n_rows)
+            except (TypeError, ValueError):
+                total_rows = None
+            if total_rows is not None and total_rows > sum_shown:
+                others_count = total_rows - sum_shown
+        if others_count <= 0:
+            others_count = overflow_count
+
+        # --- Build the display order (top to bottom): largest .. smallest, Otros.
+        display_labels = [_truncate(v, 22) for v, _ in shown]
+        display_values = [c for _, c in shown]
+        display_colors = [_PALETTE[i % len(_PALETTE)] for i in range(len(shown))]
+
+        has_others = others_count > 0 and others_categories > 0
+        if has_others:
+            display_labels.append(f"Otros ({others_categories} categorías)")
+            display_values.append(others_count)
+            display_colors.append(_OTHER_COLOR)
+
+        total = sum(display_values) or 1.0
+
+        # barh draws bottom-up, so reverse the display order before plotting to
+        # land the largest category on top and "Otros" at the bottom.
+        labels = list(reversed(display_labels))
+        values = list(reversed(display_values))
+        colors = list(reversed(display_colors))
+        y_pos = range(len(values))
+
+        # Height scales with the number of bars so dense reports stay readable.
+        n_bars = len(values)
+        height = max(2.4, min(0.4 * n_bars + 1.2, 14.0))
+        fig = Figure(figsize=(6.4, height), dpi=150)
+        ax = fig.add_subplot(111)
+
+        ax.barh(list(y_pos), values, color=colors, edgecolor="white")
+        ax.set_yticks(list(y_pos))
+        ax.set_yticklabels(labels, fontsize=8)
+        ax.set_xlabel("conteo", fontsize=9)
+
+        max_val = max(values) if values else 1.0
+        ax.set_xlim(0, max_val * 1.18 if max_val > 0 else 1.0)
+
+        # Annotate each bar with its count and percentage at the end of the bar.
+        for y, val in zip(y_pos, values):
+            pct = val / total * 100.0
+            ax.text(
+                val + max_val * 0.012,
+                y,
+                f"{int(round(val))} ({pct:.0f}%)",
+                va="center",
+                ha="left",
+                fontsize=7,
+                color="#202020",
+            )
+
+        if safe_title:
+            ax.set_title(safe_title, fontsize=13, loc="left", pad=10)
+
+        fig.tight_layout()
+        return fig
+    except Exception as exc:  # noqa: BLE001 — never raise from a figure builder.
+        return _message_figure(
+            f"error al dibujar barras: {exc}", color=_ERROR_TEXT
+        )

python/functions/datascience/categorical_top_bar_figure_test.py

@@ -0,0 +1,103 @@
+"""Tests para categorical_top_bar_figure (barras de categorías top, grupo eda).
+
+Usa el backend Agg sin pyplot; no muestra ni guarda figuras. Cada test cierra
+explícitamente la Figure construida (matplotlib.pyplot.close) para no acumular
+estado entre tests.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.figure import Figure  # noqa: E402
+
+from categorical_top_bar_figure import categorical_top_bar_figure
+
+
+def _make_top(n):
+    """n items {value, count, pct} ordenados desc por count."""
+    return [
+        {"value": f"cat_{i}", "count": n - i, "pct": (n - i) / sum(range(1, n + 1))}
+        for i in range(n)
+    ]
+
+
+def _bar_count(ax):
+    """Devuelve el nº de barras (longitud del primer BarContainer del Axes)."""
+    if ax.containers:
+        return len(ax.containers[0])
+    return 0
+
+
+def test_returns_figure():
+    fig = categorical_top_bar_figure(_make_top(3), n_distinct=3, title="col")
+    assert isinstance(fig, Figure)
+    plt.close(fig)
+
+
+def test_ten_items_topk_six_yields_seven_bars():
+    top = _make_top(10)
+    fig = categorical_top_bar_figure(top, n_distinct=10, title="muchas", top_k=6)
+    ax = fig.axes[0]
+    # 6 categorías explícitas + 1 barra "Otros".
+    assert _bar_count(ax) == 7
+    plt.close(fig)
+
+
+def test_empty_top_does_not_raise_and_returns_figure():
+    fig = categorical_top_bar_figure([], n_distinct=0, title="vacía")
+    assert isinstance(fig, Figure)
+    # Sin datos: no debe haber barras.
+    assert _bar_count(fig.axes[0]) == 0
+    plt.close(fig)
+
+
+def test_long_value_truncated():
+    long_value = "una_categoria_con_un_nombre_larguisimo_que_excede_el_limite"
+    top = [
+        {"value": long_value, "count": 10, "pct": 0.5},
+        {"value": "corta", "count": 10, "pct": 0.5},
+    ]
+    fig = categorical_top_bar_figure(top, n_distinct=2, title="col", top_k=6)
+    ax = fig.axes[0]
+    tick_texts = [t.get_text() for t in ax.get_yticklabels()]
+    # El valor largo aparece truncado con elipsis y NO en su forma completa.
+    assert any("…" in t for t in tick_texts)
+    assert long_value not in " ".join(tick_texts)
+    plt.close(fig)
+
+
+def test_none_value_and_none_count_are_handled():
+    top = [
+        {"value": None, "count": 5, "pct": 0.5},
+        {"value": "b", "count": None, "pct": 0.0},  # count None -> se descarta
+        {"value": "c", "count": 5, "pct": 0.5},
+    ]
+    fig = categorical_top_bar_figure(top, n_distinct=2, title="con nones", top_k=6)
+    assert isinstance(fig, Figure)
+    # Solo 2 items válidos, sin overflow -> 2 barras, sin "Otros".
+    assert _bar_count(fig.axes[0]) == 2
+    plt.close(fig)
+
+
+def test_n_rows_adds_exact_others_bar():
+    # 3 categorías mostradas suman 30, dataset real 100 -> "Otros" = 70.
+    top = [
+        {"value": "a", "count": 15, "pct": 0.15},
+        {"value": "b", "count": 10, "pct": 0.10},
+        {"value": "c", "count": 5, "pct": 0.05},
+    ]
+    fig = categorical_top_bar_figure(
+        top, n_distinct=20, title="col", top_k=3, n_rows=100
+    )
+    ax = fig.axes[0]
+    # 3 explícitas + Otros.
+    assert _bar_count(ax) == 4
+    tick_texts = [t.get_text() for t in ax.get_yticklabels()]
+    # La barra Otros refleja n_distinct - top_k = 17 categorías.
+    assert any("Otros (17 categorías)" in t for t in tick_texts)
+    # Su anotación lleva el count 70.
+    annotation_texts = [t.get_text() for t in ax.texts]
+    assert any("70" in t for t in annotation_texts)
+    plt.close(fig)

python/functions/datascience/classify_relationship_type.md

@@ -0,0 +1,68 @@
+---
+name: classify_relationship_type
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def classify_relationship_type(xs: list, ys: list) -> dict"
+description: "Clasifica el TIPO de relacion entre dos variables numericas pareadas por indice para el EDA automatico del grupo eda. Limpia los pares de forma defensiva (descarta None/bool/NaN/inf), reusa pearson y spearman_corr del registry y ajusta polinomios de grado 2 y 3 con numpy.polyfit (R^2 manual), y a partir de esas senales etiqueta la forma: 'lineal', 'polinomica (grado 2/3)', 'monotona no-lineal' o 'debil/sin forma'. Orden de decision: debil -> monotona -> polinomica -> lineal (la primera que matchea gana), con umbrales calibrados para datos reales discretos/ruidosos. Devuelve ademas los coeficientes del mejor modelo en orden de numpy.polyval para pintar la curva de ajuste sobre el scatter. Funcion pura no-throw: ante datos insuficientes (menos de 5 pares validos o varianza ~0) o cualquier fallo devuelve el dict canonico con tipo='debil/sin forma' y el resto a None."
+tags: [eda, correlation, relationship, classification, polyfit, datascience, pure]
+params:
+  - name: xs
+    desc: "Lista (o tupla) de valores numericos de la primera variable, pareada por indice con ys. Cada par xs[i],ys[i] se descarta si cualquiera de los dos es None, bool, NaN o inf. Lectura defensiva."
+  - name: ys
+    desc: "Lista (o tupla) de valores numericos de la segunda variable, pareada por indice con xs. Mismas reglas de limpieza que xs."
+output: "Dict con SIEMPRE las mismas 8 claves: tipo (str: 'lineal' | 'polinómica (grado 2)' | 'polinómica (grado 3)' | 'monótona no-lineal' | 'débil/sin forma'); pearson (float|None: coeficiente de Pearson r); r2_linear (float|None: r**2 del ajuste lineal); spearman (float|None: rho de Spearman); r2_poly2 (float|None: R^2 del ajuste polinomico de grado 2); r2_poly3 (float|None: R^2 del ajuste de grado 3); best_degree (int|None: grado del modelo elegido — 1 lineal, 2/3 polinomico, None si monotona/debil); coeffs (list|None: coeficientes del mejor modelo en orden de numpy.polyval para pintar la curva, o None). Ante datos insuficientes o error: tipo='débil/sin forma' y el resto de claves a None."
+uses_functions: [pearson_py_datascience, spearman_corr_py_datascience]
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [numpy]
+tested: true
+tests: ["test_lineal", "test_polinomica_cuadratica", "test_monotona_no_lineal", "test_monotona_exponencial", "test_debil_sin_forma", "test_lista_vacia_no_lanza", "test_longitudes_distintas_no_lanza", "test_todos_none_no_lanza", "test_entradas_none_no_lanza", "test_constante_no_lanza", "test_filtra_nan_inf_bool"]
+test_file_path: "python/functions/datascience/classify_relationship_type_test.py"
+file_path: "python/functions/datascience/classify_relationship_type.py"
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience.classify_relationship_type import classify_relationship_type
+import numpy as np
+
+# Relacion claramente cuadratica (forma de parabola) sobre dominio simetrico.
+x = list(np.linspace(-10, 10, 60))
+y = [v * v for v in x]
+
+res = classify_relationship_type(x, y)
+print(res["tipo"])         # 'polinómica (grado 2)'
+print(res["best_degree"])  # 2
+print(res["r2_linear"])    # 0.0   -> el Pearson lineal no ve la parabola
+print(res["r2_poly2"])     # 1.0
+print(res["coeffs"])       # [1.0, -0.0, -0.0]  -> numpy.polyval(coeffs, x) ~ x**2
+
+# El capitulo pinta la curva de ajuste cuando coeffs no es None:
+#   if res["coeffs"] is not None:
+#       xs_fit = np.linspace(min(x), max(x), 200)
+#       ys_fit = np.polyval(res["coeffs"], xs_fit)
+#       ax.plot(xs_fit, ys_fit)   # curva sobre el ax.scatter(x, y)
+```
+
+## Cuando usarla
+
+- Usala en el capitulo de relaciones/correlaciones del EDA automatico, despues de detectar dos columnas numericas con alguna asociacion, para decidir QUE curva de ajuste pintar sobre el scatter (recta, parabola, cubica o ninguna) y poner una etiqueta legible al tipo de relacion.
+- Cuando un Pearson bajo no signifique "sin relacion": esta funcion cruza Pearson con Spearman y con ajustes polinomicos para distinguir una relacion lineal debil de una monotona no-lineal (que el rango si capta) o de una curva polinomica.
+- Cuando necesites un punto de entrada determinista y no-throw que, con los mismos datos, devuelva siempre el mismo `tipo` y los mismos `coeffs` listos para `numpy.polyval` sin tener que ajustar modelos a mano en el capitulo.
+
+## Gotchas
+
+- Funcion pura, deterministica y no-throw: ante menos de 5 pares validos, varianza ~0 (xs o ys constante) o cualquier excepcion interna devuelve el dict canonico `tipo="débil/sin forma"` con el resto de claves a `None`. El dict SIEMPRE trae las 8 claves: nunca compruebes existencia, comprueba `None`.
+- El orden de decision importa: `débil -> monótona -> polinómica -> lineal` (la primera que matchee gana). La monotonia se evalua ANTES que el ajuste polinomico, asi que una curva monotona suave (exp, log, potencias) sale `monótona no-lineal` aunque un cubico tambien la ajuste — la dominancia del rango (Spearman >> Pearson) es la senal mas interpretable. Solo cae en `polinómica` una forma curva NO monotona (p.ej. una parabola, Spearman ~0 pero R^2 polinomico alto).
+- Umbrales fijos (calibrados para EDA con datos discretos/ruidosos, no para inferencia formal): `débil/sin forma` si las tres senales son bajas a la vez (`abs(pearson) < 0.3` y `abs(spearman) < 0.3` y `mejor_poly < 0.3`); `monótona no-lineal` si `abs(spearman) - abs(pearson) >= 0.1` y `abs(spearman) >= 0.4`; `polinómica (grado N)` si el mejor polinomico mejora `>= 0.1` sobre el lineal y su R^2 `>= 0.3`; en cualquier otro caso con senal (no debil) `lineal`. El suelo de 0.3 evita llamar "debil" a relaciones reales pero discretas (conteos, escalas ordinales) con R^2 bajo pero direccion clara.
+- `coeffs` va en orden de `numpy.polyval` (grado descendente). Para `lineal` es `[pendiente, intercepto]` (grado 1); para `polinómica` los del grado elegido; para `monótona no-lineal` y `débil/sin forma` es `None` (el scatter pintara una curva suavizada o nada — lo decide el capitulo, no esta funcion).
+- `best_degree` prefiere el grado 2 sobre el 3 cuando empatan dentro de 0.02 de R^2 (parsimonia): no esperes grado 3 salvo que mejore claramente.
+- Los pares con `None`, `bool`, `NaN` o `inf` se descartan por indice en silencio; `bool` cuenta como no-numerico (un `True` no es `1`). El dominio de los datos afecta al resultado: una parabola sobre un dominio simetrico da Pearson ~0 (sale `polinómica`), pero sobre un dominio asimetrico el Pearson sube y puede salir `lineal`.

python/functions/datascience/classify_relationship_type.py

@@ -0,0 +1,187 @@
+"""Clasifica el TIPO de relacion entre dos variables numericas pareadas.
+
+Funcion pura del grupo eda. Dadas dos listas numericas pareadas por indice,
+limpia los pares de forma defensiva, calcula correlaciones lineal (Pearson) y de
+rangos (Spearman) y ajustes polinomicos de grado 2 y 3, y a partir de esas
+senales etiqueta la forma de la relacion para el EDA automatico:
+
+    "lineal" | "polinómica (grado 2)" | "polinómica (grado 3)" |
+    "monótona no-lineal" | "débil/sin forma"
+
+Ademas devuelve los coeficientes del mejor modelo (en orden de numpy.polyval)
+para que el capitulo pinte la curva de ajuste sobre el scatter. Reusa las
+funciones del registry `pearson` y `spearman_corr` en vez de reimplementarlas.
+
+NUNCA lanza: ante cualquier fallo o dato insuficiente devuelve el dict canonico
+con tipo="débil/sin forma" y el resto de claves a None.
+"""
+
+import math
+import warnings
+
+import numpy as np
+
+from datascience.datascience import pearson
+from datascience.spearman_corr import spearman_corr
+
+# Forma canonica de la respuesta cuando no se puede clasificar (datos
+# insuficientes, varianza nula o error interno). Siempre las mismas claves.
+_WEAK = {
+    "tipo": "débil/sin forma",
+    "pearson": None,
+    "r2_linear": None,
+    "spearman": None,
+    "r2_poly2": None,
+    "r2_poly3": None,
+    "best_degree": None,
+    "coeffs": None,
+}
+
+
+def _is_num(v) -> bool:
+    """True si v es un numero real finito (int/float, no bool, no NaN, no inf)."""
+    return (
+        isinstance(v, (int, float))
+        and not isinstance(v, bool)
+        and not (isinstance(v, float) and (math.isnan(v) or math.isinf(v)))
+    )
+
+
+def _poly_r2(coeffs, x_arr, y_arr, ss_tot: float) -> float:
+    """R^2 de un ajuste polinomico: 1 - SS_res/SS_tot. 0 si SS_tot==0."""
+    if ss_tot == 0.0:
+        return 0.0
+    pred = np.polyval(coeffs, x_arr)
+    ss_res = float(np.sum((y_arr - pred) ** 2))
+    return 1.0 - ss_res / ss_tot
+
+
+def classify_relationship_type(xs: list, ys: list) -> dict:
+    """Clasifica el tipo de relacion entre dos variables numericas pareadas.
+
+    Empareja xs[i],ys[i] por indice y descarta el par si cualquiera de los dos
+    es None, bool, NaN o inf. Sobre los pares limpios calcula Pearson r
+    (r2_linear = r**2), Spearman rho y los R^2 de ajustes polinomicos de grado 2
+    y 3 (con numpy.polyfit + R^2 manual). Con esas senales decide la etiqueta.
+
+    Orden de evaluacion de la etiqueta (la primera que matchee gana). Los
+    umbrales estan calibrados para datos reales, a menudo discretos y ruidosos
+    (conteos, escalas ordinales): una relacion con |r| >= 0.3, |rho| >= 0.3 o un
+    polinomio con R^2 >= 0.3 ya tiene FORMA y no debe etiquetarse como "debil".
+        1. "débil/sin forma" — todas las senales bajas a la vez:
+           abs(pearson) < 0.3 y abs(spearman) < 0.3 y mejor_poly < 0.3.
+        2. "monótona no-lineal" — el rango (Spearman) capta una monotonia que el
+           Pearson lineal no: abs(spearman) - abs(pearson) >= 0.1 y
+           abs(spearman) >= 0.4. No se fuerza un polinomio (coeffs/best_degree =
+           None); el capitulo dibuja la tendencia ordenada sobre el scatter.
+        3. "polinómica (grado N)" — el mejor polinomico mejora claramente sobre
+           el lineal (mejor_poly - r2_linear >= 0.1) y mejor_poly >= 0.3. N es el
+           grado (2 o 3) con mejor R^2, prefiriendo el 2 si empatan dentro de 0.02
+           (parsimonia).
+        4. "lineal" — el resto: hay senal (no es debil) y la forma que existe es
+           esencialmente lineal. best_degree=1, coeffs del ajuste de grado 1.
+
+    Si hay menos de 5 pares validos, o la varianza de xs o de ys es ~0
+    (constante), devuelve directamente "débil/sin forma".
+
+    Args:
+        xs: lista (o tupla) de valores numericos de la primera variable,
+            pareada por indice con ys. Pares con None/bool/NaN/inf se descartan.
+        ys: lista (o tupla) de valores numericos de la segunda variable,
+            pareada por indice con xs.
+
+    Returns:
+        dict con SIEMPRE las mismas claves:
+            tipo (str), pearson (float|None), r2_linear (float|None),
+            spearman (float|None), r2_poly2 (float|None), r2_poly3 (float|None),
+            best_degree (int|None: 1, 2, 3 o None),
+            coeffs (list|None: coeficientes en orden de numpy.polyval, o None).
+        Nunca lanza: ante fallo o datos insuficientes devuelve el dict debil.
+    """
+    try:
+        if xs is None or ys is None:
+            return dict(_WEAK)
+
+        pairs = [
+            (float(x), float(y))
+            for x, y in zip(xs, ys)
+            if _is_num(x) and _is_num(y)
+        ]
+
+        # Datos insuficientes para hablar de forma de la relacion.
+        if len(pairs) < 5:
+            return dict(_WEAK)
+
+        clean_x = [p[0] for p in pairs]
+        clean_y = [p[1] for p in pairs]
+
+        # Varianza ~0 en cualquiera de las series => relacion indefinida.
+        if len(set(clean_x)) < 2 or len(set(clean_y)) < 2:
+            return dict(_WEAK)
+        x_arr = np.asarray(clean_x, dtype=float)
+        y_arr = np.asarray(clean_y, dtype=float)
+        if float(np.var(x_arr)) < 1e-15 or float(np.var(y_arr)) < 1e-15:
+            return dict(_WEAK)
+
+        # Correlaciones reutilizando las funciones del registry.
+        r = pearson(clean_x, clean_y)
+        spearman = spearman_corr(clean_x, clean_y)
+        r2_linear = r ** 2
+
+        # Ajustes polinomicos grado 2 y 3 con R^2 manual.
+        ss_tot = float(np.sum((y_arr - float(np.mean(y_arr))) ** 2))
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            c1 = np.polyfit(x_arr, y_arr, 1)
+            c2 = np.polyfit(x_arr, y_arr, 2)
+            c3 = np.polyfit(x_arr, y_arr, 3)
+        r2_poly2 = _poly_r2(c2, x_arr, y_arr, ss_tot)
+        r2_poly3 = _poly_r2(c3, x_arr, y_arr, ss_tot)
+
+        mejor_poly = max(r2_poly2, r2_poly3)
+        # Grado del mejor polinomico, con preferencia por la parsimonia: solo se
+        # elige el grado 3 si supera al grado 2 por mas de 0.02.
+        best_poly_degree = 3 if (r2_poly3 - r2_poly2) > 0.02 else 2
+
+        abs_s = abs(spearman)
+        abs_p = abs(r)
+
+        # Decision en orden: debil-temprano -> monotona -> polinomica -> lineal.
+        if abs_p < 0.3 and abs_s < 0.3 and mejor_poly < 0.3:
+            # Ninguna senal supera el suelo de forma: relacion debil/sin forma.
+            tipo = "débil/sin forma"
+            best_degree = None
+            coeffs = None
+        elif (abs_s - abs_p) >= 0.1 and abs_s >= 0.4:
+            # Spearman (rango) capta una monotonia que el Pearson lineal no:
+            # relacion monotona no-lineal. No se fuerza un polinomio que tal vez
+            # no ajusta bien; el capitulo dibuja la tendencia ordenada.
+            tipo = "monótona no-lineal"
+            best_degree = None
+            coeffs = None
+        elif (mejor_poly - r2_linear) >= 0.1 and mejor_poly >= 0.3:
+            tipo = "polinómica (grado {})".format(best_poly_degree)
+            best_degree = best_poly_degree
+            best_coeffs = c2 if best_poly_degree == 2 else c3
+            coeffs = [float(c) for c in best_coeffs]
+        else:
+            # Hay senal (no es debil) y no es ni monotona-pura ni polinomica:
+            # la correlacion que existe es esencialmente lineal.
+            tipo = "lineal"
+            best_degree = 1
+            coeffs = [float(c) for c in c1]
+
+        return {
+            "tipo": tipo,
+            "pearson": round(float(r), 6),
+            "r2_linear": round(float(r2_linear), 6),
+            "spearman": round(float(spearman), 6),
+            "r2_poly2": round(float(r2_poly2), 6),
+            "r2_poly3": round(float(r2_poly3), 6),
+            "best_degree": best_degree,
+            "coeffs": (
+                [round(c, 8) for c in coeffs] if coeffs is not None else None
+            ),
+        }
+    except Exception:
+        return dict(_WEAK)

python/functions/datascience/classify_relationship_type_test.py

@@ -0,0 +1,174 @@
+"""Tests para classify_relationship_type."""
+
+import os
+import sys
+
+import numpy as np
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from classify_relationship_type import classify_relationship_type
+
+# Claves que el dict de salida debe contener SIEMPRE.
+_EXPECTED_KEYS = {
+    "tipo", "pearson", "r2_linear", "spearman",
+    "r2_poly2", "r2_poly3", "best_degree", "coeffs",
+}
+
+
+def _assert_shape(r):
+    """Toda salida tiene exactamente las 8 claves canonicas."""
+    assert isinstance(r, dict)
+    assert set(r.keys()) == _EXPECTED_KEYS
+
+
+def test_lineal():
+    """Golden: y = 2x + 1 con ruido pequeno -> 'lineal', best_degree=1."""
+    rng = np.random.default_rng(42)
+    x = np.linspace(0.0, 10.0, 50)
+    y = 2.0 * x + 1.0 + rng.normal(0.0, 0.3, 50)
+
+    r = classify_relationship_type(list(x), list(y))
+    _assert_shape(r)
+
+    assert r["tipo"] == "lineal"
+    assert r["best_degree"] == 1
+    assert r["r2_linear"] >= 0.5
+    # coeffs ~ [pendiente, intercepto] del ajuste de grado 1.
+    assert r["coeffs"] is not None and len(r["coeffs"]) == 2
+    assert abs(r["coeffs"][0] - 2.0) < 0.1   # pendiente ~2
+    assert abs(r["coeffs"][1] - 1.0) < 0.3   # intercepto ~1
+
+
+def test_polinomica_cuadratica():
+    """Golden: y = x**2 sobre [-10, 10] -> 'polinómica', best_degree in (2, 3)."""
+    x = np.linspace(-10.0, 10.0, 60)
+    y = x ** 2
+
+    r = classify_relationship_type(list(x), list(y))
+    _assert_shape(r)
+
+    assert r["tipo"].startswith("polinómica")
+    assert r["best_degree"] in (2, 3)
+    # Una parabola perfecta queda capturada por el grado 2 (parsimonia).
+    assert r["best_degree"] == 2
+    assert r["r2_poly2"] > 0.99
+    assert r["coeffs"] is not None and len(r["coeffs"]) == r["best_degree"] + 1
+
+
+def test_monotona_no_lineal():
+    """Golden: monotona convexa de cola pesada -> 'monótona no-lineal'.
+
+    y = 1/(N+1-i)**2 es estrictamente creciente (Spearman ~ 1) pero su cola
+    explosiva hace que ni la recta ni un polinomio de grado 2/3 la ajusten
+    (R^2 polinomico < 0.5), de modo que el Pearson lineal NO capta la relacion
+    que el rango (Spearman) si ve. Construccion deterministica (sin azar).
+    """
+    n = 200
+    i = np.arange(n, dtype=float)
+    y = 1.0 / (n + 1 - i) ** 2
+
+    r = classify_relationship_type(list(i), list(y))
+    _assert_shape(r)
+
+    assert r["tipo"] == "monótona no-lineal"
+    assert r["best_degree"] is None
+    assert r["coeffs"] is None
+    # Spearman fuerte y claramente por encima del Pearson.
+    assert abs(r["spearman"]) >= 0.5
+    assert abs(r["spearman"]) - abs(r["pearson"]) >= 0.15
+
+
+def test_monotona_exponencial():
+    """DoD literal: y = exp(x) (monotona no-lineal) -> 'monótona no-lineal'.
+
+    exp es estrictamente creciente (Spearman = 1) pero el Pearson lineal queda
+    claramente por debajo (~0.86), así que la dominancia del rango la marca como
+    monótona no-lineal en vez de lineal o polinómica.
+    """
+    x = np.linspace(0.0, 5.0, 80)
+    y = np.exp(x)
+
+    r = classify_relationship_type(list(x), list(y))
+    _assert_shape(r)
+
+    assert r["tipo"] == "monótona no-lineal"
+    assert r["best_degree"] is None and r["coeffs"] is None
+    assert abs(r["spearman"]) >= 0.9
+    assert abs(r["spearman"]) - abs(r["pearson"]) >= 0.1
+
+
+def test_debil_sin_forma():
+    """Golden: x e y independientes (semilla fija) -> 'débil/sin forma'."""
+    rng = np.random.default_rng(0)
+    x = rng.normal(0.0, 1.0, 200)
+    y = rng.normal(0.0, 1.0, 200)
+
+    r = classify_relationship_type(list(x), list(y))
+    _assert_shape(r)
+
+    assert r["tipo"] == "débil/sin forma"
+    assert r["best_degree"] is None
+    assert r["coeffs"] is None
+    # Todas las senales son bajas.
+    assert abs(r["pearson"]) < 0.3
+    assert r["r2_linear"] < 0.1
+
+
+def test_lista_vacia_no_lanza():
+    """Edge: listas vacias -> dict debil canonico, sin lanzar."""
+    r = classify_relationship_type([], [])
+    _assert_shape(r)
+    assert r["tipo"] == "débil/sin forma"
+    assert r["pearson"] is None
+    assert r["r2_linear"] is None
+    assert r["spearman"] is None
+    assert r["r2_poly2"] is None
+    assert r["r2_poly3"] is None
+    assert r["best_degree"] is None
+    assert r["coeffs"] is None
+
+
+def test_longitudes_distintas_no_lanza():
+    """Edge: listas de distinta longitud -> empareja por indice, no lanza."""
+    # zip trunca a la longitud minima: solo 3 pares (< 5) -> debil.
+    r = classify_relationship_type([1, 2, 3, 4, 5, 6, 7, 8], [1.0, 2.0, 3.0])
+    _assert_shape(r)
+    assert r["tipo"] == "débil/sin forma"
+    assert r["best_degree"] is None
+
+
+def test_todos_none_no_lanza():
+    """Edge: todos los valores None -> ningun par valido -> debil, no lanza."""
+    r = classify_relationship_type([None, None, None, None, None, None],
+                                   [None, None, None, None, None, None])
+    _assert_shape(r)
+    assert r["tipo"] == "débil/sin forma"
+    assert r["coeffs"] is None
+
+
+def test_entradas_none_no_lanza():
+    """Edge: xs/ys None directamente -> debil, no lanza."""
+    assert classify_relationship_type(None, None)["tipo"] == "débil/sin forma"
+    assert classify_relationship_type([1.0, 2.0], None)["tipo"] == "débil/sin forma"
+
+
+def test_constante_no_lanza():
+    """Edge: ys constante (varianza ~0) -> debil, no lanza."""
+    r = classify_relationship_type([1, 2, 3, 4, 5, 6, 7], [5, 5, 5, 5, 5, 5, 5])
+    _assert_shape(r)
+    assert r["tipo"] == "débil/sin forma"
+
+
+def test_filtra_nan_inf_bool():
+    """Edge: pares con NaN/inf/bool/None se descartan por indice."""
+    nan = float("nan")
+    inf = float("inf")
+    # Solo i=0,1,2,3,4 quedan validos (5 pares) y forman una recta perfecta.
+    xs = [0.0, 1.0, 2.0, 3.0, 4.0, nan, inf, True, None]
+    ys = [1.0, 3.0, 5.0, 7.0, 9.0, 1.0, 2.0, 3.0, 4.0]
+    r = classify_relationship_type(xs, ys)
+    _assert_shape(r)
+    # Los 5 pares validos son y = 2x + 1 exacto -> lineal.
+    assert r["tipo"] == "lineal"
+    assert r["best_degree"] == 1

python/functions/datascience/compute_text_duplicates.md

@@ -0,0 +1,102 @@
+---
+id: compute_text_duplicates_py_datascience
+name: compute_text_duplicates
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def compute_text_duplicates(texts, near_threshold=0.85, sample_max=2000) -> dict"
+description: "Detecta documentos duplicados en un corpus de texto. Los duplicados EXACTOS se calculan siempre con la stdlib: cada documento se normaliza (colapsa espacios, strip, lower) y se hashea con SHA-1; n_exact_dup es cuántos docs repiten uno ya visto y exact_dup_pct su porcentaje. Los CASI-duplicados (near-dup) usan la dependencia OPCIONAL datasketch (MinHash + LSH sobre 3-shingles de palabras); si no está instalada, esa parte degrada a available:False sin afectar al resto. Estilo dict-no-throw del grupo eda — nunca lanza."
+tags: [eda, datascience, text, nlp, duplicates, minhash, pure, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [hashlib, re]
+example: |
+  from datascience.compute_text_duplicates import compute_text_duplicates
+  texts = ["El gato come pescado", "El gato come pescado", "Un perro ladra"]
+  result = compute_text_duplicates(texts)
+  # {"n_docs": 3, "n_exact_dup": 1, "exact_dup_pct": 33.33, "n_unique": 2,
+  #  "near_dup": {"available": False, "n_near_dup_docs": 0}}
+tested: true
+tests:
+  - "test_duplicados_exactos"
+  - "test_sin_duplicados"
+  - "test_vacio"
+  - "test_near_dup_degrada"
+test_file_path: "python/functions/datascience/compute_text_duplicates_test.py"
+file_path: "python/functions/datascience/compute_text_duplicates.py"
+params:
+  - name: texts
+    desc: "Lista de documentos de texto. Los elementos None o que no sean str se descartan silenciosamente; n_docs cuenta solo los documentos válidos. None como argumento se trata como lista vacía."
+  - name: near_threshold
+    desc: "Umbral de similitud Jaccard (0–1) para considerar dos documentos casi-duplicados en el cálculo near-dup vía MinHashLSH. Solo aplica si datasketch está instalada. Default 0.85."
+  - name: sample_max
+    desc: "Número máximo de documentos muestreados (los primeros) para el cálculo near-dup, que es O(n) en memoria de MinHashes. No afecta al conteo de duplicados exactos, que siempre recorre todo el corpus. Default 2000."
+output: "Dict con exactamente 5 claves, siempre presentes: n_docs (int, docs válidos), n_exact_dup (int, docs que repiten un texto normalizado ya visto = n_docs - n_unique), exact_dup_pct (float a 2 decimales = n_exact_dup/n_docs*100, o None si el corpus está vacío), n_unique (int, nº de textos normalizados distintos), y near_dup (sub-dict con available:bool y n_near_dup_docs:int; cuando available es True incluye además threshold con el near_threshold usado). La función nunca lanza: captura toda excepción y degrada."
+---
+
+## Ejemplo
+
+```python
+from datascience.compute_text_duplicates import compute_text_duplicates
+
+# Tres copias del mismo texto (con espacios/casing distintos) + dos únicos.
+texts = [
+    "El gato come pescado",
+    "El gato come pescado",
+    "el  GATO   come pescado",   # mismo tras normalizar
+    "Un perro ladra",
+    "La luna brilla",
+]
+
+compute_text_duplicates(texts)
+# {
+#   "n_docs": 5,
+#   "n_exact_dup": 2,          # 3 copias del primer texto => 2 repeticiones
+#   "exact_dup_pct": 40.0,     # 2 / 5 * 100
+#   "n_unique": 3,             # 3 textos normalizados distintos
+#   "near_dup": {"available": False, "n_near_dup_docs": 0},  # datasketch ausente
+# }
+
+# Corpus vacío: contrato estable, exact_dup_pct None, sin excepción.
+compute_text_duplicates([])
+# {"n_docs": 0, "n_exact_dup": 0, "exact_dup_pct": None, "n_unique": 0,
+#  "near_dup": {"available": False, "n_near_dup_docs": 0}}
+```
+
+## Cuando usarla
+
+Úsala en la fase de calidad de un EDA de texto, cuando quieras saber cuánto de
+tu corpus es ruido duplicado antes de entrenar, vectorizar o muestrear: te da
+el porcentaje de duplicados exactos (`exact_dup_pct`), el número de documentos
+únicos (`n_unique`) y, si tienes `datasketch` instalada, una estimación de
+casi-duplicados (paráfrasis, copias con pequeñas ediciones) vía MinHash + LSH.
+Pásale directamente la columna/lista de textos crudos; la función filtra None y
+no-str por ti y nunca lanza, así que es segura para encadenar en pipelines de
+perfilado.
+
+## Gotchas
+
+- **Near-dup requiere `datasketch` (opcional).** Si la librería no está
+  instalada, `near_dup` degrada a `{"available": False, "n_near_dup_docs": 0}`
+  (sin clave `threshold`) y el resto del resultado se calcula igual. Los
+  duplicados **exactos** funcionan siempre porque solo usan la stdlib (hash).
+- **Normalización de exactos.** Dos textos cuentan como el mismo duplicado
+  exacto si coinciden tras `" ".join(doc.split()).strip().lower()`: se colapsan
+  espacios/tabuladores/saltos, se recortan extremos y se ignora el caso. Cambios
+  de puntuación o acentos SÍ los distinguen (no se eliminan).
+- **`n_exact_dup` cuenta repeticiones, no grupos.** Con 3 copias de un mismo
+  texto, `n_exact_dup` es 2 (las dos copias extra), no 1. Equivale a
+  `n_docs - n_unique`.
+- **`exact_dup_pct` es `None` con corpus vacío** (no `ZeroDivisionError`); en
+  cualquier otro caso es un float redondeado a 2 decimales.
+- **`sample_max` solo limita el near-dup.** El conteo de duplicados exactos
+  recorre todo el corpus; el near-dup muestrea los primeros `sample_max`
+  documentos para acotar memoria. Si el corpus está ordenado, considera barajar
+  antes para que la muestra sea representativa.
+- **Elementos no-str se descartan.** `True`/`False` no cuentan como str y se
+  ignoran igual que `None`; `n_docs` refleja solo los documentos válidos.

python/functions/datascience/compute_text_duplicates.py

@@ -0,0 +1,128 @@
+"""Detección de documentos duplicados en un corpus de texto.
+
+Función pura, estilo dict-no-throw del grupo `eda`: nunca lanza, siempre
+devuelve el mismo contrato de claves. Los duplicados EXACTOS se calculan
+siempre con la stdlib (normalización + hash SHA-1). Los CASI-duplicados
+(near-dup) requieren la dependencia opcional `datasketch`; si no está
+instalada, esa parte degrada limpiamente a ``available: False`` sin afectar
+al resto del cálculo.
+"""
+
+import hashlib
+import re
+
+
+def _compute_near_dup(valid, near_threshold, sample_max):
+    """Cuenta documentos con al menos otro casi-duplicado vía MinHash + LSH.
+
+    Import perezoso de ``datasketch``. Si la librería no está disponible (o
+    cualquier paso falla), degrada a ``{"available": False, "n_near_dup_docs": 0}``
+    sin propagar la excepción.
+
+    Args:
+        valid: lista de str ya filtrada (sin None ni no-str).
+        near_threshold: umbral de similitud Jaccard para LSH.
+        sample_max: número máximo de documentos a muestrear.
+
+    Returns:
+        dict con ``available`` (bool) y ``n_near_dup_docs`` (int). Cuando
+        ``available`` es True, incluye además ``threshold``.
+    """
+    try:
+        from datasketch import MinHash, MinHashLSH
+    except Exception:
+        return {"available": False, "n_near_dup_docs": 0}
+
+    try:
+        docs = valid[:sample_max]
+        num_perm = 128
+        lsh = MinHashLSH(threshold=near_threshold, num_perm=num_perm)
+        minhashes = {}
+
+        for i, doc in enumerate(docs):
+            tokens = re.findall(r"\w+", doc.lower())
+            shingles = set()
+            for j in range(len(tokens) - 2):
+                shingles.add(" ".join(tokens[j:j + 3]))
+            # Documentos con menos de 3 tokens no generan 3-shingles: caemos a
+            # los tokens sueltos para no perderlos del todo.
+            if not shingles:
+                shingles = set(tokens)
+            if not shingles:
+                # Documento sin tokens (cadena vacía / solo símbolos): se omite.
+                continue
+            m = MinHash(num_perm=num_perm)
+            for sh in shingles:
+                m.update(sh.encode("utf-8"))
+            key = "d{}".format(i)
+            minhashes[key] = m
+            lsh.insert(key, m)
+
+        n_near = 0
+        for key, m in minhashes.items():
+            matches = lsh.query(m)
+            if len(matches) > 1:
+                n_near += 1
+
+        return {
+            "available": True,
+            "n_near_dup_docs": int(n_near),
+            "threshold": near_threshold,
+        }
+    except Exception:
+        return {"available": False, "n_near_dup_docs": 0}
+
+
+def compute_text_duplicates(texts, near_threshold=0.85, sample_max=2000) -> dict:
+    """Detecta duplicados exactos y casi-duplicados en un corpus de texto.
+
+    Args:
+        texts: lista de documentos. Los elementos None o que no sean str se
+            descartan; ``n_docs`` cuenta solo los válidos.
+        near_threshold: umbral de similitud Jaccard para considerar dos
+            documentos casi-duplicados (solo near-dup, requiere datasketch).
+        sample_max: tope de documentos muestreados para el cálculo near-dup.
+
+    Returns:
+        dict con las claves ``n_docs``, ``n_exact_dup``, ``exact_dup_pct``
+        (float redondeado a 2 decimales, o None si el corpus está vacío),
+        ``n_unique`` y ``near_dup`` (sub-dict con ``available`` y
+        ``n_near_dup_docs``, más ``threshold`` cuando está disponible).
+        Nunca lanza: captura toda excepción y degrada.
+    """
+    # Filtrado defensivo de documentos válidos.
+    try:
+        valid = [t for t in texts if isinstance(t, str)] if texts is not None else []
+    except Exception:
+        valid = []
+
+    n_docs = len(valid)
+
+    # Duplicados exactos: normalizar + hash SHA-1 (stdlib, siempre disponible).
+    try:
+        seen = set()
+        n_exact_dup = 0
+        for doc in valid:
+            norm = " ".join(doc.split()).strip().lower()
+            digest = hashlib.sha1(norm.encode("utf-8")).hexdigest()
+            if digest in seen:
+                n_exact_dup += 1
+            else:
+                seen.add(digest)
+        n_unique = len(seen)
+    except Exception:
+        n_exact_dup = 0
+        n_unique = 0
+
+    exact_dup_pct = round(n_exact_dup / n_docs * 100, 2) if n_docs > 0 else None
+
+    # Casi-duplicados: opcional vía datasketch, degrada solo.
+    near_dup = _compute_near_dup(valid, near_threshold, sample_max)
+
+    return {
+        "n_docs": n_docs,
+        "n_exact_dup": n_exact_dup,
+        "exact_dup_pct": exact_dup_pct,
+        "n_unique": n_unique,
+        "near_dup": near_dup,
+    }

python/functions/datascience/compute_text_duplicates_test.py

@@ -0,0 +1,77 @@
+"""Tests para compute_text_duplicates.
+
+Importa el modulo hoja directamente (`datascience.compute_text_duplicates`)
+para no depender de que el paquete reexporte la funcion en su __init__.
+datasketch normalmente NO esta instalada en el venv, asi que near_dup
+degrada a available=False; los tests no requieren la libreria.
+"""
+
+from datascience.compute_text_duplicates import compute_text_duplicates
+
+
+EXPECTED_KEYS = {"n_docs", "n_exact_dup", "exact_dup_pct", "n_unique", "near_dup"}
+
+
+def test_duplicados_exactos():
+    """3 copias del mismo texto + 2 únicos: n_exact_dup=2, pct>0."""
+    texts = [
+        "El gato come pescado",
+        "El gato come pescado",
+        "el  GATO   come pescado",  # mismo tras normalizar (espacios + case)
+        "Un perro ladra",
+        "La luna brilla",
+    ]
+    result = compute_text_duplicates(texts)
+
+    assert set(result.keys()) == EXPECTED_KEYS
+    assert result["n_docs"] == 5
+    # 3 copias del primer texto (2 son repeticion) + 2 textos unicos.
+    assert result["n_exact_dup"] == 2
+    assert result["n_unique"] == 3
+    assert result["exact_dup_pct"] is not None
+    assert result["exact_dup_pct"] > 0
+    # 2 / 5 * 100 = 40.0
+    assert abs(result["exact_dup_pct"] - 40.0) < 1e-9
+
+
+def test_sin_duplicados():
+    """Corpus sin repeticiones: n_exact_dup=0, n_unique==n_docs."""
+    texts = [
+        "primero documento distinto",
+        "segundo documento distinto",
+        "tercero documento distinto",
+    ]
+    result = compute_text_duplicates(texts)
+
+    assert result["n_docs"] == 3
+    assert result["n_exact_dup"] == 0
+    assert result["n_unique"] == 3
+    assert abs(result["exact_dup_pct"] - 0.0) < 1e-9
+
+
+def test_vacio():
+    """Corpus vacio: n_docs 0, exact_dup_pct None, no lanza."""
+    result = compute_text_duplicates([])
+
+    assert set(result.keys()) == EXPECTED_KEYS
+    assert result["n_docs"] == 0
+    assert result["n_exact_dup"] == 0
+    assert result["exact_dup_pct"] is None
+    assert result["n_unique"] == 0
+    assert result["near_dup"]["n_near_dup_docs"] == 0
+
+
+def test_near_dup_degrada():
+    """near_dup expone 'available' (bool) y no lanza aunque falte datasketch."""
+    texts = ["uno dos tres cuatro", "uno dos tres cuatro cinco", "algo distinto"]
+    result = compute_text_duplicates(texts)
+
+    near = result["near_dup"]
+    assert "available" in near
+    assert isinstance(near["available"], bool)
+    assert "n_near_dup_docs" in near
+    assert isinstance(near["n_near_dup_docs"], int)
+    # Tambien tolera None y entradas no-str sin lanzar.
+    mixed = compute_text_duplicates(["hola", None, 123, "hola"])
+    assert mixed["n_docs"] == 2
+    assert mixed["n_exact_dup"] == 1

python/functions/datascience/compute_text_length_stats.md

@@ -0,0 +1,86 @@
+---
+id: compute_text_length_stats_py_datascience
+name: compute_text_length_stats
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def compute_text_length_stats(texts, n_bins=20) -> dict"
+description: "Profiles the length distribution of a corpus of text documents for EDA: per-document characters, words (unicode \\w+ tokens) and sentences (segments split on .!?… with a minimum of 1 per non-empty doc), each summarized with mean/p50/p90/p99/min/max (nearest-rank percentiles), plus an equal-width histogram of per-document word counts. None and non-str items are discarded. Dict-no-throw: never raises. Stdlib only (re)."
+tags: [eda, datascience, text, nlp, length, statistics, pure, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [re, math]
+example: |
+  from datascience.compute_text_length_stats import compute_text_length_stats
+  result = compute_text_length_stats(["Hola mundo.", "Una frase mas larga aqui."], n_bins=5)
+tested: true
+tests:
+  - "test_basico"
+  - "test_vacio"
+  - "test_descarta_none"
+  - "test_un_documento"
+test_file_path: "python/functions/datascience/compute_text_length_stats_test.py"
+file_path: "python/functions/datascience/compute_text_length_stats.py"
+params:
+  - name: texts
+    desc: "List of text documents (str). None entries and any non-str items (ints, floats, etc.) are discarded before any computation. An empty string \"\" is kept (chars 0, words 0, sentences 0)."
+  - name: n_bins
+    desc: "Number of equal-width bins for the per-document word-count histogram. Default 20. When all docs have the same word count, there are <2 docs, or n_bins < 1, a single covering bin is returned instead."
+output: "Dict with keys n_docs (int), chars, words, sentences and word_hist. Each of the three axis sub-dicts has the exact keys mean (float, 2 decimals), p50, p90, p99, min, max (ints). When there are no valid documents, n_docs is 0, every axis statistic is None and word_hist is []. word_hist is a list of {lo: float, hi: float, count: int} bins; the sum of all bin counts equals n_docs."
+---
+
+## Ejemplo
+
+```python
+from datascience.compute_text_length_stats import compute_text_length_stats
+
+compute_text_length_stats(
+    [
+        "Hola mundo.",
+        "Una frase mas larga con varias palabras aqui.",
+        "Esto. Tiene. Tres frases distintas!",
+    ],
+    n_bins=5,
+)
+# {
+#   "n_docs": 3,
+#   "chars":     {"mean": 30.33, "p50": 35, "p90": 45, "p99": 45, "min": 11, "max": 45},
+#   "words":     {"mean": 5.0,   "p50": 5,  "p90": 8,  "p99": 8,  "min": 2,  "max": 8},
+#   "sentences": {"mean": 1.67,  "p50": 1,  "p90": 3,  "p99": 3,  "min": 1,  "max": 3},
+#   "word_hist": [
+#     {"lo": 2.0, "hi": 3.2, "count": 1},
+#     {"lo": 3.2, "hi": 4.4, "count": 0},
+#     {"lo": 4.4, "hi": 5.6, "count": 1},
+#     {"lo": 5.6, "hi": 6.8, "count": 0},
+#     {"lo": 6.8, "hi": 8.0, "count": 1},
+#   ],
+# }
+```
+
+## Cuando usarla
+
+Úsala al perfilar una columna o corpus de texto libre en un EDA: cuando
+necesites saber lo largos que son los documentos (en caracteres, palabras y
+frases) y cómo se reparte esa longitud antes de tokenizar, vectorizar o decidir
+truncados/ventanas para un modelo. Pásale la lista de strings crudos de la
+columna; `None` y valores no-texto se descartan solos. Encaja en el grupo `eda`
+como bloque de longitud junto a `summarize_categorical`.
+
+## Gotchas
+
+- Función pura, solo stdlib (`re`). No usa numpy, pandas ni sklearn.
+- Percentiles por método **nearest-rank** (devuelven un valor real de la lista,
+  no interpolan); por eso p50/p90/p99/min/max son enteros y `mean` es el único
+  float (redondeado a 2 decimales).
+- El conteo de frases es una **aproximación** por puntuación (`.!?…`): un texto
+  sin esa puntuación cuenta como 1 frase si no está vacío; abreviaturas o
+  ellipsis pueden inflar o reducir el conteo.
+- `word_hist` es equal-width entre min y max de palabras: con todos los docs
+  del mismo tamaño, menos de 2 docs, o `n_bins < 1`, devuelve un único bin.
+- Dict-no-throw: ante input inesperado devuelve la forma vacía
+  (`n_docs` 0, ejes `None`, `word_hist` []) en vez de lanzar.

python/functions/datascience/compute_text_length_stats.py

@@ -0,0 +1,168 @@
+"""Pure EDA helper: document length distribution for the `eda` group.
+
+Given a list of text documents, computes the length distribution along three
+axes (characters, words and sentences) plus an equal-width histogram of the
+per-document word counts. Stdlib only (``re`` + ``statistics`` semantics via a
+hand-rolled nearest-rank percentile). No numpy, no sklearn.
+
+The function is dict-no-throw: it never raises. On any unexpected input it
+degrades to the empty-shape result.
+"""
+
+import math
+import re
+
+_WORD_RE = re.compile(r"\w+", re.UNICODE)
+_SENT_RE = re.compile(r"[.!?…]+")
+
+
+def _empty_axis() -> dict:
+    """Return an axis sub-dict with every statistic set to ``None``."""
+    return {"mean": None, "p50": None, "p90": None, "p99": None, "min": None, "max": None}
+
+
+def _pct(sorted_vals, q):
+    """Nearest-rank percentile of an already-sorted list.
+
+    Args:
+        sorted_vals: List of numbers sorted ascending.
+        q: Percentile in the 0..100 range.
+
+    Returns:
+        The value at the nearest rank, or ``None`` for an empty list.
+    """
+    n = len(sorted_vals)
+    if n == 0:
+        return None
+    if q <= 0:
+        return sorted_vals[0]
+    rank = math.ceil(q / 100.0 * n)
+    if rank < 1:
+        rank = 1
+    if rank > n:
+        rank = n
+    return sorted_vals[rank - 1]
+
+
+def _axis_stats(values) -> dict:
+    """Compute mean/p50/p90/p99/min/max over a list of integer counts.
+
+    ``mean`` is rounded to 2 decimals; every other statistic is an integer
+    (they are counts). Returns an all-``None`` axis for an empty list.
+    """
+    if not values:
+        return _empty_axis()
+    sv = sorted(values)
+    return {
+        "mean": round(sum(sv) / len(sv), 2),
+        "p50": int(_pct(sv, 50)),
+        "p90": int(_pct(sv, 90)),
+        "p99": int(_pct(sv, 99)),
+        "min": int(sv[0]),
+        "max": int(sv[-1]),
+    }
+
+
+def _word_hist(word_counts, n_bins) -> list:
+    """Equal-width histogram of per-document word counts.
+
+    Builds ``n_bins`` bins between ``min`` and ``max`` of the word counts. When
+    every document has the same number of words, there are fewer than 2
+    documents, or ``n_bins`` is not at least 1, a single covering bin is
+    returned. With no documents the result is ``[]``. The sum of bin ``count``
+    always equals ``len(word_counts)``.
+    """
+    if not word_counts:
+        return []
+    wmin = min(word_counts)
+    wmax = max(word_counts)
+    if wmax == wmin or len(word_counts) < 2 or n_bins < 1:
+        return [{"lo": float(wmin), "hi": float(wmax), "count": len(word_counts)}]
+
+    width = (wmax - wmin) / n_bins
+    bins = []
+    for i in range(n_bins):
+        lo = wmin + i * width
+        hi = wmin + (i + 1) * width
+        bins.append({"lo": float(lo), "hi": float(hi), "count": 0})
+    # Pin the last upper edge to the real maximum to avoid float drift.
+    bins[-1]["hi"] = float(wmax)
+
+    for wc in word_counts:
+        if wc >= wmax:
+            idx = n_bins - 1
+        else:
+            idx = int((wc - wmin) / width)
+            if idx < 0:
+                idx = 0
+            elif idx >= n_bins:
+                idx = n_bins - 1
+        bins[idx]["count"] += 1
+    return bins
+
+
+def compute_text_length_stats(texts, n_bins=20) -> dict:
+    """Summarize the length distribution of a corpus of text documents.
+
+    For each document three lengths are measured: characters (``len(doc)``),
+    words (count of ``\\w+`` unicode tokens) and sentences (non-empty segments
+    after splitting on ``.!?…``, with a minimum of 1 for any non-empty
+    document). For each axis the mean, p50, p90, p99, min and max are reported,
+    plus an equal-width histogram of the per-document word counts.
+
+    ``None`` entries and any non-``str`` items in ``texts`` are discarded.
+    The function never raises: on empty/``None`` input or any internal error it
+    returns the empty-shape result (``n_docs`` 0, all-``None`` axes, ``[]``
+    histogram).
+
+    Args:
+        texts: List of text documents (``str``). ``None`` and non-``str``
+            items are dropped.
+        n_bins: Number of equal-width bins for the word-count histogram.
+            Default 20.
+
+    Returns:
+        Dict with keys ``n_docs``, ``chars``, ``words``, ``sentences`` and
+        ``word_hist``. Each of the three axes is a sub-dict with ``mean``
+        (float, 2 decimals), ``p50``, ``p90``, ``p99``, ``min`` and ``max``
+        (ints), all ``None`` when there are no documents. ``word_hist`` is a
+        list of ``{lo, hi, count}`` bins whose ``count`` sums to ``n_docs``.
+    """
+    empty_axis = _empty_axis()
+    fallback = {
+        "n_docs": 0,
+        "chars": dict(empty_axis),
+        "words": dict(empty_axis),
+        "sentences": dict(empty_axis),
+        "word_hist": [],
+    }
+    try:
+        if not texts:
+            return fallback
+
+        docs = [t for t in texts if isinstance(t, str)]
+        n_docs = len(docs)
+        if n_docs == 0:
+            return fallback
+
+        char_counts = [len(d) for d in docs]
+        word_counts = [len(_WORD_RE.findall(d)) for d in docs]
+
+        sent_counts = []
+        for d in docs:
+            segments = [s for s in _SENT_RE.split(d) if s.strip()]
+            n = len(segments)
+            if d and n == 0:
+                # Non-empty document with no detectable sentence: count as 1.
+                n = 1
+            sent_counts.append(n)
+
+        return {
+            "n_docs": n_docs,
+            "chars": _axis_stats(char_counts),
+            "words": _axis_stats(word_counts),
+            "sentences": _axis_stats(sent_counts),
+            "word_hist": _word_hist(word_counts, n_bins),
+        }
+    except Exception:
+        return fallback

python/functions/datascience/compute_text_length_stats_test.py

@@ -0,0 +1,70 @@
+"""Tests para compute_text_length_stats.
+
+Inserta `python/functions` en sys.path (relativo a este archivo) para importar
+el modulo hoja por su paquete `datascience`, sin depender de que el paquete lo
+reexporte en su __init__.
+"""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from datascience.compute_text_length_stats import compute_text_length_stats
+
+
+def test_basico():
+    """Varios textos de longitudes distintas: stats y histograma coherentes."""
+    texts = [
+        "Hola mundo.",                      # 2 words, 1 sentence
+        "Una frase mas larga con varias palabras aqui.",  # 8 words, 1 sentence
+        "Corto.",                           # 1 word, 1 sentence
+        "Esto. Tiene. Tres frases distintas!",            # 5 words, 3 sentences
+    ]
+    result = compute_text_length_stats(texts)
+
+    assert result["n_docs"] == 4
+    # Diferentes longitudes en palabras -> max estrictamente mayor que min.
+    assert result["words"]["max"] > result["words"]["min"]
+    # El histograma de palabras no esta vacio.
+    assert result["word_hist"] != []
+    # La suma de counts del histograma cubre todos los documentos.
+    assert sum(b["count"] for b in result["word_hist"]) == result["n_docs"]
+    # mean es float redondeado; min/max son enteros.
+    assert isinstance(result["words"]["mean"], float)
+    assert isinstance(result["words"]["min"], int)
+    assert isinstance(result["words"]["max"], int)
+    # El documento con 3 frases empuja el max de sentences a >= 3.
+    assert result["sentences"]["max"] >= 3
+
+
+def test_vacio():
+    """Lista vacia: n_docs 0, subdicts None, word_hist []."""
+    result = compute_text_length_stats([])
+    assert result["n_docs"] == 0
+    for axis in ("chars", "words", "sentences"):
+        for key in ("mean", "p50", "p90", "p99", "min", "max"):
+            assert result[axis][key] is None
+    assert result["word_hist"] == []
+
+
+def test_descarta_none():
+    """None y valores no-str se descartan del computo."""
+    result = compute_text_length_stats(["hello world", None, 123, 4.5, "foo bar baz"])
+    # Solo dos strings validos.
+    assert result["n_docs"] == 2
+    assert result["words"]["min"] == 2  # "hello world"
+    assert result["words"]["max"] == 3  # "foo bar baz"
+    assert sum(b["count"] for b in result["word_hist"]) == 2
+
+
+def test_un_documento():
+    """Un solo documento: word_hist tiene exactamente un bin con count 1."""
+    result = compute_text_length_stats(["solo un documento aqui"])
+    assert result["n_docs"] == 1
+    assert len(result["word_hist"]) == 1
+    assert result["word_hist"][0]["count"] == 1
+    # Con un unico documento, p50 == min == max == su numero de palabras (4).
+    assert result["words"]["min"] == 4
+    assert result["words"]["max"] == 4
+    assert result["words"]["p50"] == 4

python/functions/datascience/compute_text_readability.md

@@ -0,0 +1,88 @@
+---
+id: compute_text_readability_py_datascience
+name: compute_text_readability
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def compute_text_readability(texts, sample_max=500) -> dict"
+description: "Calcula la legibilidad Flesch Reading Ease de un corpus de texto usando textstat con import perezoso y degradación. Filtra None/no-str/vacíos, muestrea hasta sample_max documentos (los primeros) y agrega los scores Flesch en {mean, p50, min, max}. Si textstat no está instalada devuelve available=False sin lanzar. Estilo dict-no-throw del grupo eda — nunca lanza."
+tags: [eda, datascience, text, nlp, readability, flesch, textstat, pure, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [math, textstat]
+example: |
+  from datascience.compute_text_readability import compute_text_readability
+  out = compute_text_readability(["The cat sat on the mat. It was warm and sunny."])
+  # {"available": True, "n_scored": 1, "flesch": {"mean": 109.0, "p50": 109.0, "min": 108.96..., "max": 108.96...}}
+tested: true
+tests:
+  - "test_prosa_ingles"
+  - "test_vacio"
+  - "test_degradacion"
+test_file_path: "python/functions/datascience/compute_text_readability_test.py"
+file_path: "python/functions/datascience/compute_text_readability.py"
+params:
+  - name: texts
+    desc: "Lista de str (documentos del corpus). Los elementos None, no-str o vacíos tras strip() se descartan silenciosamente. El orden se respeta: el muestreo toma los primeros documentos válidos."
+  - name: sample_max
+    desc: "Número máximo de documentos válidos a puntuar (los primeros). Default 500. Acota el coste en corpus grandes. Valores no convertibles a int caen a 500; negativos se tratan como 0."
+output: "Dict con exactamente 3 claves siempre presentes: available (bool: True si textstat se pudo importar), n_scored (int: nº de documentos efectivamente puntuados), flesch (dict con mean, p50, min, max). mean y p50 redondeados a 1 decimal; p50 por nearest-rank sobre los scores ordenados; min/max son los scores extremos sin redondear. Todos los valores de flesch son None cuando n_scored es 0. La función nunca lanza: cualquier excepción global (incluida ImportError de textstat) degrada a available=False, n_scored=0 y flesch todo None."
+---
+
+## Ejemplo
+
+```python
+from datascience.compute_text_readability import compute_text_readability
+
+textos = [
+    "The cat sat on the mat. It was a warm and sunny day in the park.",
+    "Reading is a wonderful habit. Books open doors to new worlds and ideas.",
+    "He ran quickly to the store to buy some fresh bread and a bottle of milk.",
+]
+
+compute_text_readability(textos)
+# {
+#   "available": True,
+#   "n_scored": 3,
+#   "flesch": {"mean": 91.4, "p50": 95.4, "min": 70.08..., "max": 108.83...}
+# }
+
+# Corpus vacío (textstat presente): available True pero nada que puntuar.
+compute_text_readability([])
+# {"available": True, "n_scored": 0,
+#  "flesch": {"mean": None, "p50": None, "min": None, "max": None}}
+```
+
+## Cuando usarla
+
+Úsala en un EDA de texto cuando necesites una métrica única y comparable de
+**lo fácil que es de leer** un corpus de documentos (descripciones, reviews,
+artículos, tickets). Devuelve el resumen Flesch Reading Ease agregado
+(`mean`/`p50`/`min`/`max`) listo para un report o un bloque del notebook, sin
+tener que iterar `textstat` a mano. Pásale la lista de textos crudos y, si el
+corpus es grande, limita el coste con `sample_max`. El estilo dict-no-throw
+permite incrustarla en pipelines del grupo `eda` sin envolver en try/except.
+
+## Gotchas
+
+- **`textstat` es una dependencia opcional.** Si no está instalada (o falla al
+  importar) la función NO lanza: devuelve `available=False`, `n_scored=0` y
+  `flesch` todo `None`. Comprueba `available` antes de interpretar los números.
+- **Flesch Reading Ease está pensado para prosa en inglés.** Aplicado a otros
+  idiomas o a texto no-prosa (código, listas, tablas, cadenas muy cortas) los
+  scores no son interpretables, aunque se calculen sin error.
+- **Escala Flesch:** valores **altos** = más fácil de leer (≈90–100 muy fácil),
+  valores **bajos** = más difícil (puede ser negativo en texto muy denso). No
+  se recortan a ningún rango: se reportan tal cual los devuelve `textstat`.
+- **`available=True` con `n_scored=0`** significa que `textstat` está presente
+  pero el corpus no aportó documentos puntuables (vacío, solo None/no-str, o
+  todos los docs fallaron al puntuar). Es distinto de `available=False`.
+- **Muestreo = los primeros `sample_max`**, no aleatorio. Si el orden del corpus
+  está sesgado, el resumen reflejará ese sesgo.
+- **`mean` y `p50` redondean a 1 decimal**; `min`/`max` se devuelven sin
+  redondear (los scores extremos reales).

python/functions/datascience/compute_text_readability.py

@@ -0,0 +1,121 @@
+"""Legibilidad Flesch Reading Ease de un corpus de texto.
+
+Función pura del grupo `eda`, estilo dict-no-throw: nunca lanza. Usa la
+librería `textstat` con import perezoso y degradación: si `textstat` no está
+instalada (o falla al importar), devuelve un resultado con `available=False`
+en lugar de propagar el error.
+"""
+
+
+def _percentile_nearest_rank(sorted_values, pct):
+    """Percentil por nearest-rank sobre una lista ya ordenada ascendente.
+
+    rank = ceil(pct/100 * n); índice 1-based recortado a [1, n].
+    Devuelve None si la lista está vacía.
+    """
+    n = len(sorted_values)
+    if n == 0:
+        return None
+    import math
+
+    rank = math.ceil((pct / 100.0) * n)
+    if rank < 1:
+        rank = 1
+    if rank > n:
+        rank = n
+    return sorted_values[rank - 1]
+
+
+def compute_text_readability(texts, sample_max=500) -> dict:
+    """Calcula la legibilidad Flesch Reading Ease de un corpus.
+
+    Args:
+        texts: lista de str. Los elementos None, no-str o vacíos (tras strip)
+            se descartan. Se muestrean los primeros `sample_max` documentos
+            válidos.
+        sample_max: número máximo de documentos a puntuar (los primeros).
+
+    Returns:
+        Dict con la forma exacta::
+
+            {"available": bool, "n_scored": int,
+             "flesch": {"mean": float|None, "p50": float|None,
+                        "min": float|None, "max": float|None}}
+
+        `available` es True si `textstat` se pudo importar. La función nunca
+        lanza: cualquier excepción global degrada a `available=False`.
+    """
+    empty = {
+        "available": False,
+        "n_scored": 0,
+        "flesch": {"mean": None, "p50": None, "min": None, "max": None},
+    }
+    try:
+        # Import perezoso con degradación: textstat es una dependencia opcional.
+        try:
+            import textstat
+        except Exception:
+            return {
+                "available": False,
+                "n_scored": 0,
+                "flesch": {"mean": None, "p50": None, "min": None, "max": None},
+            }
+
+        # Filtrar y muestrear documentos válidos (los primeros sample_max).
+        docs = []
+        if texts is not None:
+            try:
+                limit = int(sample_max)
+            except Exception:
+                limit = 500
+            if limit < 0:
+                limit = 0
+            for item in texts:
+                if not isinstance(item, str):
+                    continue
+                if item.strip() == "":
+                    continue
+                docs.append(item)
+                if len(docs) >= limit:
+                    break
+
+        scores = []
+        for doc in docs:
+            try:
+                score = textstat.flesch_reading_ease(doc)
+            except Exception:
+                continue
+            try:
+                score = float(score)
+            except Exception:
+                continue
+            scores.append(score)
+
+        n_scored = len(scores)
+        if n_scored == 0:
+            # textstat presente pero corpus vacío / sin puntuar.
+            return {
+                "available": True,
+                "n_scored": 0,
+                "flesch": {"mean": None, "p50": None, "min": None, "max": None},
+            }
+
+        mean_val = round(sum(scores) / n_scored, 1)
+        sorted_scores = sorted(scores)
+        p50_raw = _percentile_nearest_rank(sorted_scores, 50)
+        p50_val = round(p50_raw, 1) if p50_raw is not None else None
+        min_val = sorted_scores[0]
+        max_val = sorted_scores[-1]
+
+        return {
+            "available": True,
+            "n_scored": n_scored,
+            "flesch": {
+                "mean": mean_val,
+                "p50": p50_val,
+                "min": min_val,
+                "max": max_val,
+            },
+        }
+    except Exception:
+        return empty

python/functions/datascience/compute_text_readability_test.py

@@ -0,0 +1,74 @@
+"""Tests para compute_text_readability."""
+
+import sys
+import os
+import builtins
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from datascience.compute_text_readability import compute_text_readability
+
+
+EXPECTED_KEYS = {"available", "n_scored", "flesch"}
+FLESCH_KEYS = {"mean", "p50", "min", "max"}
+
+
+def test_prosa_ingles():
+    """Varios textos en prosa inglesa: available True, n_scored>0, mean no None."""
+    texts = [
+        "The cat sat on the mat. It was a warm and sunny day in the park.",
+        "She sells sea shells by the sea shore. The shells she sells are surely sea shells.",
+        "Reading is a wonderful habit. Books open doors to new worlds and ideas.",
+        "He ran quickly to the store to buy some fresh bread and a bottle of milk.",
+    ]
+    out = compute_text_readability(texts)
+
+    assert set(out.keys()) == EXPECTED_KEYS
+    assert out["available"] is True
+    assert out["n_scored"] > 0
+    assert set(out["flesch"].keys()) == FLESCH_KEYS
+    assert out["flesch"]["mean"] is not None
+    assert out["flesch"]["p50"] is not None
+    assert out["flesch"]["min"] is not None
+    assert out["flesch"]["max"] is not None
+    # min <= mean/p50 <= max coherente.
+    assert out["flesch"]["min"] <= out["flesch"]["max"]
+
+
+def test_vacio():
+    """Corpus vacío con textstat presente: available True, n_scored 0, flesch None."""
+    out = compute_text_readability([])
+
+    assert set(out.keys()) == EXPECTED_KEYS
+    assert out["available"] is True
+    assert out["n_scored"] == 0
+    assert out["flesch"]["mean"] is None
+    assert out["flesch"]["p50"] is None
+    assert out["flesch"]["min"] is None
+    assert out["flesch"]["max"] is None
+
+    # Elementos no-str / vacíos también se descartan -> n_scored 0.
+    out2 = compute_text_readability([None, "", "   ", 123])
+    assert out2["available"] is True
+    assert out2["n_scored"] == 0
+
+
+def test_degradacion(monkeypatch):
+    """Sin textstat (ImportError forzado): degrada a available False sin lanzar."""
+    import datascience.compute_text_readability as m
+
+    real = builtins.__import__
+
+    def fake(name, *a, **k):
+        if name == "textstat" or name.startswith("textstat."):
+            raise ImportError("simulado")
+        return real(name, *a, **k)
+
+    monkeypatch.setattr(builtins, "__import__", fake)
+    out = m.compute_text_readability(["The cat sat on the mat. It was happy and warm."])
+    assert out["available"] is False
+    assert out["n_scored"] == 0
+    assert out["flesch"]["mean"] is None
+    assert out["flesch"]["p50"] is None
+    assert out["flesch"]["min"] is None
+    assert out["flesch"]["max"] is None

python/functions/datascience/compute_top_ngrams.md

@@ -0,0 +1,103 @@
+---
+id: compute_top_ngrams_py_datascience
+name: compute_top_ngrams
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def compute_top_ngrams(texts, n=2, top_k=15, remove_stopwords=True) -> dict"
+description: "Calcula los n-gramas de palabras más frecuentes de un corpus de texto (n=1 unigramas, 2 bigramas, 3 trigramas...). Tokeniza a minúsculas con re.findall(r'\\w+', ...), descarta tokens numéricos y, si remove_stopwords=True, elimina stopwords ES+EN ANTES de formar los n-gramas (n-gramas contiguos sobre la secuencia de tokens de contenido, sin cruzar documentos). Pura y autocontenida con collections.Counter, sin sklearn. Estilo dict-no-throw del grupo eda: nunca lanza."
+tags: [eda, datascience, text, nlp, ngrams, bigrams, trigrams, pure, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [re, collections]
+example: |
+  from datascience.compute_top_ngrams import compute_top_ngrams
+  texts = ["machine learning rocks", "we love machine learning"]
+  compute_top_ngrams(texts, n=2, top_k=5)
+  # {"n": 2, "top": [{"ngram": "machine learning", "count": 2}, ...]}
+tested: true
+tests:
+  - "test_bigramas"
+  - "test_trigramas"
+  - "test_vacio"
+  - "test_stopwords"
+test_file_path: "python/functions/datascience/compute_top_ngrams_test.py"
+file_path: "python/functions/datascience/compute_top_ngrams.py"
+params:
+  - name: texts
+    desc: "Lista (o tupla) de cadenas. Los elementos None o que no sean str se descartan silenciosamente. Cada documento se tokeniza por separado; los n-gramas no cruzan la frontera entre documentos."
+  - name: n
+    desc: "Tamaño del n-grama: 1 unigramas, 2 bigramas, 3 trigramas, etc. Valores < 1 o no enteros producen top vacío (se conserva tal cual en la clave 'n' del retorno)."
+  - name: top_k
+    desc: "Número máximo de n-gramas a devolver, ordenados por frecuencia descendente con desempate alfabético determinista. Default 15. Valores negativos se tratan como 0."
+  - name: remove_stopwords
+    desc: "Si True (default) elimina las stopwords ES+EN de una lista inline (~130 términos de altísima frecuencia) ANTES de formar los n-gramas, de modo que los n-gramas se construyen sobre la secuencia de tokens de contenido."
+output: "Dict con exactamente 2 claves: n (el n recibido, sin normalizar) y top (lista de dicts {'ngram': str, 'count': int} ordenada por count descendente, longitud <= top_k). ngram es la unión de los tokens del n-grama por un espacio. Corpus vacío, tokens insuficientes para formar n-gramas o cualquier excepción interna degradan a {'n': n, 'top': []}. La función nunca lanza."
+---
+
+## Ejemplo
+
+```python
+from datascience.compute_top_ngrams import compute_top_ngrams
+
+texts = [
+    "machine learning rocks",
+    "machine learning is fun",
+    "we love machine learning",
+]
+
+# Bigramas (n=2): "machine learning" aparece en los 3 documentos.
+compute_top_ngrams(texts, n=2, top_k=5)
+# {
+#   "n": 2,
+#   "top": [
+#       {"ngram": "machine learning", "count": 3},
+#       {"ngram": "learning fun",     "count": 1},
+#       {"ngram": "learning rocks",   "count": 1},
+#       {"ngram": "love machine",     "count": 1},
+#   ],
+# }
+
+# Unigramas con stopwords fuera (default): solo palabras de contenido.
+compute_top_ngrams(["the cat sat on the mat"], n=1, top_k=3)
+# {"n": 1, "top": [{"ngram": "cat", "count": 1},
+#                  {"ngram": "mat", "count": 1},
+#                  {"ngram": "sat", "count": 1}]}
+```
+
+## Cuando usarla
+
+Úsala en la fase de EDA de texto cuando, además del vocabulario suelto, necesites
+ver qué **combinaciones de palabras contiguas** dominan un corpus: colocaciones,
+frases técnicas recurrentes ("machine learning", "data analyst"), o patrones de
+trigramas en titulares/descripciones. Es el complemento natural de un perfil de
+vocabulario: pasa de "qué palabras aparecen" a "qué secuencias aparecen". Llámala
+con `n=1` para unigramas, `n=2` para bigramas y `n=3` para trigramas, y ajusta
+`top_k` al tamaño de la tabla que vas a renderizar. Deja `remove_stopwords=True`
+para que los n-gramas reflejen contenido y no conectores gramaticales.
+
+## Gotchas
+
+- **Las stopwords se eliminan ANTES de formar los n-gramas.** Con
+  `remove_stopwords=True` la frase "data of analysis" produce el bigrama
+  "data analysis" (el "of" intermedio desaparece y los tokens de contenido se
+  vuelven contiguos), no "data of" ni "of analysis". Si quieres preservar la
+  adyacencia literal del texto original, pasa `remove_stopwords=False`.
+- **Los n-gramas NO cruzan documentos.** Cada elemento de `texts` se tokeniza y
+  recorre por separado; el último token de un documento nunca se combina con el
+  primero del siguiente.
+- **Tokens puramente numéricos se descartan** (`tok.isdigit()`), pero los
+  alfanuméricos mixtos no: "3d" o "covid19" sí cuentan como tokens. Un decimal
+  como "3.5" se parte en "3" y "5" por `\w+` y ambos se descartan por numéricos.
+- **La lista de stopwords es inline ES+EN**, pensada para textos generales en
+  esos dos idiomas. Para otros idiomas o jerga específica de dominio puede dejar
+  pasar conectores; en ese caso filtra el corpus aguas arriba o usa
+  `remove_stopwords=False` y posfiltra.
+- **`top` puede tener menos de `top_k` elementos** si el corpus no tiene tantos
+  n-gramas distintos. El desempate por frecuencia es alfabético (determinista),
+  no por orden de aparición.

python/functions/datascience/compute_top_ngrams.py

@@ -0,0 +1,94 @@
+"""Top n-gramas de palabras más frecuentes de un corpus de texto.
+
+Función pura, autocontenida (solo stdlib: re + collections.Counter). No depende
+de scikit-learn ni de ninguna otra librería externa. Estilo dict-no-throw del
+grupo `eda`: ante cualquier entrada degenerada o excepción interna devuelve
+``{"n": n, "top": []}`` en vez de lanzar.
+"""
+
+import re
+from collections import Counter
+
+# Lista inline de stopwords ES + EN (~80 términos de altísima frecuencia).
+# Se eliminan ANTES de formar los n-gramas: los n-gramas se construyen sobre la
+# secuencia de tokens de contenido, no sobre el texto original.
+_STOPWORDS = frozenset({
+    # Español
+    "de", "la", "que", "el", "en", "y", "a", "los", "del", "se", "las", "por",
+    "un", "para", "con", "no", "una", "su", "al", "lo", "como", "más", "mas",
+    "pero", "sus", "le", "ya", "o", "este", "sí", "si", "porque", "esta",
+    "entre", "cuando", "muy", "sin", "sobre", "también", "tambien", "me",
+    "hasta", "hay", "donde", "quien", "desde", "todo", "nos", "durante",
+    "todos", "uno", "les", "ni", "contra", "otros", "ese", "eso", "ante",
+    "ellos", "e", "esto", "mí", "antes", "algunos", "qué", "unos", "yo",
+    "otro", "otras", "otra", "él", "tanto", "esa", "estos", "mucho", "quienes",
+    "nada", "muchos", "cual", "poco", "ella", "estar", "estas", "algunas",
+    "algo", "nosotros",
+    # Inglés
+    "the", "of", "and", "to", "in", "is", "it", "for", "on", "with", "as",
+    "are", "was", "be", "this", "that", "by", "an", "or", "at", "from", "but",
+    "not", "have", "has", "had", "they", "you", "we", "he", "she", "his",
+    "her", "their", "its", "i", "my", "me", "our", "us", "do", "does", "did",
+    "will", "would", "can", "could", "should", "there", "which", "who", "what",
+    "when", "where", "how", "all", "if", "so", "than", "then", "out", "up",
+})
+
+
+def compute_top_ngrams(texts, n=2, top_k=15, remove_stopwords=True) -> dict:
+    """Calcula los n-gramas de palabras más frecuentes de un corpus.
+
+    Args:
+        texts: lista de cadenas. Los elementos ``None`` o que no sean ``str`` se
+            descartan silenciosamente.
+        n: tamaño del n-grama (1 = unigramas, 2 = bigramas, 3 = trigramas...).
+            Valores < 1 o no enteros producen ``top`` vacío.
+        top_k: número máximo de n-gramas a devolver, ordenados por frecuencia
+            descendente (con desempate alfabético determinista).
+        remove_stopwords: si ``True`` elimina las stopwords ES+EN ANTES de
+            formar los n-gramas, de modo que los n-gramas se construyen sobre la
+            secuencia de tokens de contenido (no cruzando documentos).
+
+    Returns:
+        ``{"n": n, "top": [{"ngram": "w1 w2", "count": int}, ...]}``. Corpus
+        vacío, sin tokens suficientes o cualquier excepción interna degrada a
+        ``{"n": n, "top": []}``. Nunca lanza.
+    """
+    try:
+        if not isinstance(n, int) or n < 1:
+            return {"n": n, "top": []}
+
+        try:
+            limit = int(top_k)
+        except (TypeError, ValueError):
+            limit = 0
+        if limit < 0:
+            limit = 0
+
+        if not isinstance(texts, (list, tuple)):
+            return {"n": n, "top": []}
+
+        counter = Counter()
+        for doc in texts:
+            if not isinstance(doc, str):
+                continue
+            tokens = [
+                tok
+                for tok in re.findall(r"\w+", doc.lower(), re.UNICODE)
+                if not tok.isdigit()
+            ]
+            if remove_stopwords:
+                tokens = [tok for tok in tokens if tok not in _STOPWORDS]
+            if len(tokens) < n:
+                continue
+            for i in range(len(tokens) - n + 1):
+                ngram = " ".join(tokens[i:i + n])
+                counter[ngram] += 1
+
+        if not counter:
+            return {"n": n, "top": []}
+
+        ordered = sorted(counter.items(), key=lambda kv: (-kv[1], kv[0]))
+        top = [{"ngram": ngram, "count": count} for ngram, count in ordered[:limit]]
+        return {"n": n, "top": top}
+    except Exception:
+        return {"n": n, "top": []}

python/functions/datascience/compute_top_ngrams_test.py

@@ -0,0 +1,65 @@
+"""Tests para compute_top_ngrams."""
+
+import sys
+import os
+
+# sys.path estándar: añade `python/functions/` para importar por paquete raíz.
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from datascience.compute_top_ngrams import compute_top_ngrams
+
+
+def test_bigramas():
+    # "machine learning" se repite en cada documento -> bigrama más frecuente.
+    texts = [
+        "machine learning rocks",
+        "machine learning is fun",
+        "we love machine learning",
+    ]
+    result = compute_top_ngrams(texts, n=2, top_k=5)
+    assert result["n"] == 2
+    assert result["top"], "esperaba al menos un bigrama"
+    assert result["top"][0]["ngram"] == "machine learning"
+    assert result["top"][0]["count"] == 3
+    # Cada entrada respeta el contrato {"ngram": str, "count": int}.
+    for item in result["top"]:
+        assert isinstance(item["ngram"], str)
+        assert isinstance(item["count"], int)
+
+
+def test_trigramas():
+    texts = [
+        "alpha beta gamma delta",
+        "alpha beta gamma omega",
+    ]
+    # Con stopwords desactivadas para no descartar tokens de contenido.
+    result = compute_top_ngrams(texts, n=3, top_k=5, remove_stopwords=False)
+    assert result["n"] == 3
+    ngrams = {item["ngram"]: item["count"] for item in result["top"]}
+    # "alpha beta gamma" aparece en ambos documentos.
+    assert ngrams.get("alpha beta gamma") == 2
+    # Trigramas únicos de cada documento.
+    assert ngrams.get("beta gamma delta") == 1
+    assert ngrams.get("beta gamma omega") == 1
+
+
+def test_vacio():
+    assert compute_top_ngrams([], n=2) == {"n": 2, "top": []}
+    # Documentos no-str / None se descartan -> corpus efectivamente vacío.
+    assert compute_top_ngrams([None, 123, {"a": 1}], n=2) == {"n": 2, "top": []}
+
+
+def test_stopwords():
+    # "the cat" debería desaparecer al quitar stopwords ("the" es stopword EN).
+    texts = ["the cat the cat the cat"]
+    con = compute_top_ngrams(texts, n=2, top_k=10, remove_stopwords=True)
+    sin = compute_top_ngrams(texts, n=2, top_k=10, remove_stopwords=False)
+
+    con_ngrams = {item["ngram"] for item in con["top"]}
+    sin_ngrams = {item["ngram"] for item in sin["top"]}
+
+    # Sin filtrar, el bigrama dominante es "the cat".
+    assert "the cat" in sin_ngrams
+    # Al filtrar stopwords, ya no aparece "the cat" (queda solo "cat cat").
+    assert "the cat" not in con_ngrams
+    assert con_ngrams != sin_ngrams

python/functions/datascience/compute_vocabulary_stats.md

@@ -0,0 +1,91 @@
+---
+id: compute_vocabulary_stats_py_datascience
+name: compute_vocabulary_stats
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def compute_vocabulary_stats(texts: list, top_k: int = 20, remove_stopwords: bool = True) -> dict"
+description: "Profiles the vocabulary of a text corpus for EDA: tokenises a list of documents, counts term frequencies and derives lexical-richness measures — total tokens, unique types, type-token ratio (TTR), hapax legomena and the top-k most frequent terms. Pure, stdlib only (re + collections.Counter); no nltk, no sklearn. Inline ES+EN stopword list, opt-out via remove_stopwords. Never raises: empty/degenerate input returns the zeroed result."
+tags: [eda, datascience, text, nlp, vocabulary, ttr, hapax, pure, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [re, collections]
+example: |
+  from datascience.compute_vocabulary_stats import compute_vocabulary_stats
+  result = compute_vocabulary_stats(["el gato y el perro", "gato veloz"], top_k=5)
+tested: true
+tests:
+  - "test_basico"
+  - "test_vacio"
+  - "test_stopwords_quitadas"
+  - "test_stopwords_conservadas"
+test_file_path: "python/functions/datascience/compute_vocabulary_stats_test.py"
+file_path: "python/functions/datascience/compute_vocabulary_stats.py"
+params:
+  - name: texts
+    desc: "List of documents (strings) forming the corpus. Entries that are None or not a str are silently discarded. Tokens are extracted per document with re.findall(r'\\w+', doc.lower(), re.UNICODE); purely numeric tokens (tok.isdigit()) are dropped."
+  - name: top_k
+    desc: "Maximum number of most-frequent terms to return in top_terms. Default 20. Does not affect n_tokens/n_types/ttr/hapax — only the length of the top_terms list."
+  - name: remove_stopwords
+    desc: "When True (default) common Spanish+English stopwords from the inline _STOPWORDS set (~120 entries) are removed from the token stream before any counting. Set False to keep every word (raw lexical profile)."
+output: "Dict with the exact keys n_tokens (int), n_types (int), ttr (float|None, n_types/n_tokens rounded to 4 dp), n_hapax (int, terms occurring exactly once), hapax_pct (float|None, n_hapax/n_types*100 rounded to 2 dp) and top_terms (list of {term, count, pct} sorted by count descending, pct = count/n_tokens*100 rounded to 2 dp). For an empty corpus (no tokens after filtering): n_tokens=0, n_types=0, ttr=None, n_hapax=0, hapax_pct=None, top_terms=[]. Any exception degrades to that same empty result — the function never throws."
+---
+
+## Ejemplo
+
+```python
+from datascience.compute_vocabulary_stats import compute_vocabulary_stats
+
+compute_vocabulary_stats(
+    ["el gato y el perro", "gato veloz corre", "perro perro perro"],
+    top_k=5,
+)
+# {
+#   "n_tokens": 6,        # stopwords (el, y) eliminadas por defecto
+#   "n_types": 3,         # gato, perro, veloz, corre -> tras quitar stopwords
+#   "ttr": 0.5,           # n_types / n_tokens
+#   "n_hapax": 2,         # veloz, corre (1 aparicion cada uno)
+#   "hapax_pct": 50.0,    # n_hapax / n_types * 100
+#   "top_terms": [
+#     {"term": "perro", "count": 4, "pct": 44.44},
+#     {"term": "gato",  "count": 2, "pct": 22.22},
+#     ...
+#   ],
+# }
+
+# Perfil lexico crudo (sin filtrar stopwords):
+compute_vocabulary_stats(["the cat and the dog"], remove_stopwords=False)
+```
+
+## Cuando usarla
+
+Úsala al perfilar una columna o corpus de texto libre en un EDA del grupo `eda`:
+cuando necesites medir la riqueza léxica (cuántos tokens y cuántas palabras
+distintas, type-token ratio, porcentaje de palabras que solo aparecen una vez) y
+ver qué términos dominan el vocabulario (top-k frecuencias). Pásale la lista de
+documentos crudos (filas de la columna); `None` y valores no-string se ignoran
+solos. Es el equivalente para texto largo de `summarize_categorical`, que perfila
+categorías cortas.
+
+## Gotchas
+
+- Función pura y stdlib-only, pero el resultado depende del **idioma**: la lista
+  `_STOPWORDS` cubre español e inglés. Para otros idiomas pon
+  `remove_stopwords=False` o filtra fuera, o el perfil mezclará stopwords no
+  reconocidas en `top_terms`.
+- La tokenización es `\w+` con `re.UNICODE`: separa por puntuación y conserva
+  acentos/ñ, pero NO hace stemming ni lematización — "gato" y "gatos" cuentan
+  como tipos distintos. Tampoco hace stripping de acentos, así que "más" (con
+  tilde) y "mas" son tokens diferentes (ambos están en la stoplist).
+- Los tokens **puramente numéricos** (`"123"`) se descartan siempre; un token
+  alfanumérico mixto (`"covid19"`) se conserva.
+- `ttr` baja artificialmente en corpus grandes (más texto, más repetición): no
+  compares TTR entre corpus de tamaños muy distintos sin normalizar.
+- Nunca lanza: entrada vacía, `None`, o cualquier excepción interna devuelven el
+  resultado con ceros/`None`/`[]`. Comprueba `n_tokens == 0` para detectar el
+  caso degenerado.

python/functions/datascience/compute_vocabulary_stats.py

@@ -0,0 +1,99 @@
+"""Profile the vocabulary of a text corpus for EDA (pure, stdlib only).
+
+Tokenises a list of documents, counts term frequencies and derives lexical
+richness measures (type-token ratio, hapax legomena) plus the top-k terms.
+No external NLP dependencies (no nltk, no sklearn) — only ``re`` and
+``collections`` from the standard library.
+"""
+
+import re
+from collections import Counter
+
+# Common Spanish + English stopwords. Inline, lowercase, no accents stripped
+# beyond what already appears here. Filtering is opt-in via remove_stopwords.
+_STOPWORDS = {
+    # Spanish
+    "de", "la", "que", "el", "en", "y", "a", "los", "del", "se", "las", "por",
+    "un", "para", "con", "no", "una", "su", "al", "es", "lo", "como", "mas",
+    "más", "pero", "sus", "le", "ya", "o", "este", "si", "sí", "porque",
+    "esta", "entre", "cuando", "muy", "sin", "sobre", "tambien", "también",
+    "me", "hasta", "hay", "donde", "quien", "desde", "todo", "nos", "durante",
+    "todos", "uno", "les", "ni", "contra", "otros", "ese", "eso", "ante",
+    "ellos", "e", "esto", "antes", "algunos", "que", "unos", "yo", "otro",
+    "otras", "otra", "el", "tanto", "esa", "estos", "mucho", "nada", "muchos",
+    # English
+    "the", "of", "and", "to", "in", "is", "it", "for", "on", "with", "as",
+    "was", "but", "are", "this", "that", "an", "be", "by", "or", "not", "at",
+    "from", "my", "i", "you", "he", "she", "we", "they", "his", "her", "its",
+    "our", "their", "what", "which", "who", "whom", "has", "have", "had", "do",
+    "does", "did", "will", "would", "can", "could", "should", "may", "might",
+    "must", "if", "then", "than", "so", "too", "very", "just", "also", "were",
+    "been", "being", "there", "here", "all", "any", "some", "more", "most",
+    "out", "up", "down", "into", "over", "such", "only", "own", "same",
+}
+
+
+def compute_vocabulary_stats(texts, top_k=20, remove_stopwords=True) -> dict:
+    """Profile the vocabulary of a corpus of documents.
+
+    Args:
+        texts: List of strings (the corpus). Entries that are None or not a
+            string are discarded silently.
+        top_k: Maximum number of most-frequent terms to include in
+            ``top_terms``. Default 20. Does not affect the other measures.
+        remove_stopwords: When True (default) common ES+EN stopwords are
+            dropped from the token stream before any counting.
+
+    Returns:
+        A dict with the exact keys ``n_tokens``, ``n_types``, ``ttr``,
+        ``n_hapax``, ``hapax_pct`` and ``top_terms``. For an empty corpus (no
+        tokens after filtering): n_tokens=0, n_types=0, ttr=None, n_hapax=0,
+        hapax_pct=None, top_terms=[]. Never raises — any exception degrades to
+        the empty-corpus result.
+    """
+    empty = {
+        "n_tokens": 0,
+        "n_types": 0,
+        "ttr": None,
+        "n_hapax": 0,
+        "hapax_pct": None,
+        "top_terms": [],
+    }
+    try:
+        tokens = []
+        for doc in texts or []:
+            if not isinstance(doc, str):
+                continue
+            for tok in re.findall(r"\w+", doc.lower(), re.UNICODE):
+                if tok.isdigit():
+                    continue
+                if remove_stopwords and tok in _STOPWORDS:
+                    continue
+                tokens.append(tok)
+
+        n_tokens = len(tokens)
+        if n_tokens == 0:
+            return dict(empty)
+
+        counts = Counter(tokens)
+        n_types = len(counts)
+        ttr = round(n_types / n_tokens, 4)
+
+        n_hapax = sum(1 for c in counts.values() if c == 1)
+        hapax_pct = round(n_hapax / n_types * 100, 2)
+
+        top_terms = [
+            {"term": term, "count": count, "pct": round(count / n_tokens * 100, 2)}
+            for term, count in counts.most_common(top_k)
+        ]
+
+        return {
+            "n_tokens": n_tokens,
+            "n_types": n_types,
+            "ttr": ttr,
+            "n_hapax": n_hapax,
+            "hapax_pct": hapax_pct,
+            "top_terms": top_terms,
+        }
+    except Exception:
+        return dict(empty)

python/functions/datascience/compute_vocabulary_stats_test.py

@@ -0,0 +1,74 @@
+"""Tests para compute_vocabulary_stats."""
+
+import os
+import sys
+
+sys.path.insert(
+    0, os.path.join(os.path.dirname(__file__), "..", "..", "functions")
+)
+
+from datascience.compute_vocabulary_stats import compute_vocabulary_stats
+
+
+def test_basico():
+    # Corpus con repeticiones y hapax. Stopwords desactivadas para controlar
+    # exactamente que tokens entran.
+    texts = ["gato gato perro", "perro perro raton", "elefante"]
+    r = compute_vocabulary_stats(texts, top_k=10, remove_stopwords=False)
+
+    # n_types < n_tokens cuando hay repeticiones.
+    assert r["n_types"] < r["n_tokens"]
+    assert r["n_tokens"] == 7
+    assert r["n_types"] == 4  # gato, perro, raton, elefante
+
+    # ttr en (0, 1].
+    assert 0 < r["ttr"] <= 1
+    assert r["ttr"] == round(4 / 7, 4)
+
+    # top_terms ordenado por count descendente.
+    counts = [t["count"] for t in r["top_terms"]]
+    assert counts == sorted(counts, reverse=True)
+    assert r["top_terms"][0]["term"] == "perro"
+    assert r["top_terms"][0]["count"] == 3
+
+    # hapax: raton y elefante aparecen exactamente una vez.
+    assert r["n_hapax"] == 2
+    assert r["hapax_pct"] == round(2 / 4 * 100, 2)
+
+    # pct coherente con count/n_tokens.
+    assert r["top_terms"][0]["pct"] == round(3 / 7 * 100, 2)
+
+
+def test_vacio():
+    # Sin documentos validos -> ceros / None / [].
+    for arg in ([], None, [None, 123, ""], ["123 456"]):
+        r = compute_vocabulary_stats(arg)
+        assert r["n_tokens"] == 0
+        assert r["n_types"] == 0
+        assert r["ttr"] is None
+        assert r["n_hapax"] == 0
+        assert r["hapax_pct"] is None
+        assert r["top_terms"] == []
+
+
+def test_stopwords_quitadas():
+    texts = ["the gato the perro", "de la casa azul"]
+    r = compute_vocabulary_stats(texts, remove_stopwords=True)
+    terms = {t["term"] for t in r["top_terms"]}
+    # Stopwords ES+EN no deben aparecer.
+    assert "the" not in terms
+    assert "de" not in terms
+    assert "la" not in terms
+    # Palabras de contenido si.
+    assert "gato" in terms
+    assert "casa" in terms
+
+
+def test_stopwords_conservadas():
+    texts = ["the gato the perro", "de la casa azul"]
+    r = compute_vocabulary_stats(texts, remove_stopwords=False)
+    terms = {t["term"] for t in r["top_terms"]}
+    # Con el filtro desactivado, las stopwords se conservan.
+    assert "the" in terms
+    assert "de" in terms
+    assert "la" in terms

python/functions/datascience/confidence_interval_mean.md

@@ -0,0 +1,87 @@
+---
+name: confidence_interval_mean
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def confidence_interval_mean(data: list, other: list = None, confidence: float = 0.95) -> dict"
+description: "Intervalo de confianza (IC) de la media de una muestra con la t de Student, o de la DIFERENCIA de medias de dos muestras independientes con el metodo de Welch (sin asumir varianzas iguales). Una muestra: df=n-1, se=sd_muestral/sqrt(n) (sd con ddof=1), tcrit=t.ppf((1+confidence)/2, df), ci=mean+/-tcrit*se. Dos muestras: IC de mean(data)-mean(other) con se=sqrt(se1^2+se2^2) y grados de libertad de Welch-Satterthwaite. Pura y robusta: nunca lanza; ante casos degenerados (muestra vacia, n<2) devuelve nan + clave note, y con varianza cero el IC colapsa al punto (no es error). Usa scipy.stats y numpy."
+tags: [papers, statistics, confidence-interval, welch, t-test, python]
+params:
+  - name: data
+    desc: "muestra de observaciones numericas (lista de numeros). Si other es None, el IC es el de la media de data."
+  - name: other
+    desc: "segunda muestra independiente (lista de numeros) o None (default). Si se da, el IC es el de la diferencia de medias mean(data)-mean(other) calculada con Welch (no asume varianzas iguales)."
+  - name: confidence
+    desc: "nivel de confianza en (0, 1); 0.95 = IC del 95% (default). El cuantil critico es t.ppf((1+confidence)/2, df)."
+output: "dict {mean, ci_low, ci_high, se, df, confidence, n}. mean = media de data (una muestra) o la diferencia mean(data)-mean(other) (dos muestras). En el caso de dos muestras se anaden ademas n1 y n2 (y n = n1+n2). df son los grados de libertad de la t (Welch-Satterthwaite si dos muestras). Casos degenerados (muestra vacia, n<2) anaden la clave note y dejan ci_low/ci_high/se (y a veces df) en nan; con varianza cero y n>=2 el IC colapsa a [mean, mean] con se=0 (con note, sin nan). Nunca None ni excepcion."
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [scipy, numpy]
+tested: true
+tests: ["test_one_sample_golden_contra_scipy", "test_one_sample_distinto_nivel_confianza", "test_welch_diferencia_golden_contra_scipy", "test_edge_un_solo_elemento_no_lanza_nan_note", "test_edge_lista_vacia_no_lanza_note", "test_edge_varianza_cero_colapsa_al_punto", "test_edge_welch_muestra_vacia_no_lanza_note", "test_edge_welch_n1_uno_no_lanza_note"]
+test_file_path: "python/functions/datascience/confidence_interval_mean_test.py"
+file_path: "python/functions/datascience/confidence_interval_mean.py"
+---
+
+## Ejemplo
+
+```python
+from datascience import confidence_interval_mean
+
+# IC del 95% de la media de una muestra (t de Student).
+data = [2, 4, 4, 4, 5, 5, 7, 9]
+ci = confidence_interval_mean(data, confidence=0.95)
+print(ci["mean"])     # -> 5.0
+print(ci["df"])       # -> 7.0  (n - 1)
+print(round(ci["ci_low"], 5), round(ci["ci_high"], 5))
+# -> 3.21251 6.78749   (se con sd muestral ddof=1 ~ 2.13809)
+
+# IC del 95% de la DIFERENCIA de medias (Welch, no asume varianzas iguales).
+control = [23.0, 21.0, 25.0, 22.0, 24.0, 26.0]
+tratado = [18.0, 20.0, 17.0, 19.0, 21.0]
+diff = confidence_interval_mean(control, tratado, confidence=0.95)
+print(diff["mean"])   # -> 4.5  (mean(control) - mean(tratado))
+print(round(diff["ci_low"], 4), round(diff["ci_high"], 4))
+# Si el intervalo no incluye 0, la diferencia es significativa al 5%.
+
+# Degenerados: nunca lanza.
+print(confidence_interval_mean([5])["note"])      # n < 2: ... indefinidos
+print(confidence_interval_mean([3, 3, 3])["se"])  # -> 0.0  (IC colapsa a [3, 3])
+```
+
+## Cuando usarla
+
+Cuando quieras cuantificar la **incertidumbre de una media estimada** a partir de
+una muestra: reporta `[ci_low, ci_high]` en vez de un punto suelto para mostrar
+el rango plausible del valor real al nivel de confianza pedido. Usala tambien
+para **comparar dos grupos** (A/B test, control vs tratamiento, antes vs
+despues con grupos independientes): pasa las dos muestras y, si el IC de la
+diferencia **no incluye el 0**, la diferencia es significativa al nivel
+`1 - confidence`. Es el complemento del p-valor: ademas de "hay efecto", te dice
+"de que tamano y con que margen". Para dos muestras usa Welch por defecto, asi
+que no necesitas comprobar antes si las varianzas son iguales.
+
+## Gotchas
+
+- Pura y determinista (no hace I/O, no muta las entradas), pero **no** es
+  stdlib-only: depende de `scipy.stats` y `numpy` (ambos en el venv del proyecto).
+- Con `other` usa **Welch** (df de Welch-Satterthwaite): NO asume varianzas
+  iguales ni tamanos de muestra iguales. Si necesitas el t-test clasico de
+  varianzas agrupadas (pooled), esta funcion no lo hace.
+- `sd` se calcula con **ddof=1** (sd muestral), que es lo correcto para el IC de
+  una media con la t. Atajos como `sd_poblacional/sqrt(n)` (ddof=0) dan un
+  intervalo demasiado estrecho.
+- En el caso de dos muestras, `mean` es la **diferencia** `mean(data) - mean(other)`
+  (no la media de data). El orden importa: el signo del IC depende de cual va
+  primero.
+- Nunca lanza. Casos degenerados devuelven `nan` en `ci_low`/`ci_high`/`se`
+  (y a veces `df`) mas una clave `note`: muestra vacia o `n < 2` en cualquiera de
+  las muestras. **Excepcion**: con varianza cero y `n >= 2` el IC colapsa al
+  punto `[mean, mean]` con `se = 0` (no es un error, no hay `nan`).
+- Comprueba `"note" in out` antes de usar `ci_low`/`ci_high` si la muestra puede
+  ser degenerada.

python/functions/datascience/confidence_interval_mean.py

@@ -0,0 +1,176 @@
+"""Intervalo de confianza de la media (una muestra) o de la diferencia de medias (Welch).
+
+Funcion pura del grupo papers. Calcula el intervalo de confianza (IC) de la media
+de una muestra usando la t de Student, o el IC de la diferencia de medias de dos
+muestras independientes con el metodo de Welch (sin asumir varianzas iguales).
+
+- Una muestra: ``df = n - 1``, ``se = sd / sqrt(n)`` (sd con ddof=1),
+  ``tcrit = t.ppf((1 + confidence) / 2, df)``, ``ci = mean +/- tcrit * se``.
+- Dos muestras (Welch): IC de ``mean(data) - mean(other)``, con
+  ``se = sqrt(se1^2 + se2^2)`` y grados de libertad de Welch-Satterthwaite.
+
+No lanza excepciones: ante casos degenerados (muestras vacias, ``n < 2``,
+varianza cero) devuelve un dict coherente con ``ci_low``/``ci_high``/``se`` en
+``nan`` (salvo el sub-caso de varianza cero, donde el IC colapsa al punto) y una
+clave ``note`` explicando el caso. Usa ``scipy.stats`` y ``numpy``.
+"""
+
+from __future__ import annotations
+
+import math
+
+import numpy as np
+from scipy import stats
+
+
+def confidence_interval_mean(
+    data: list, other: list = None, confidence: float = 0.95
+) -> dict:
+    """Intervalo de confianza de la media o de la diferencia de medias (Welch).
+
+    Si ``other`` es ``None``, calcula el IC de la media de ``data`` con la t de
+    Student. Si se proporciona ``other``, calcula el IC de la diferencia
+    ``mean(data) - mean(other)`` con el metodo de Welch (no asume varianzas
+    iguales) y grados de libertad de Welch-Satterthwaite.
+
+    Es una funcion pura y determinista: no hace I/O ni muta las entradas. No
+    lanza excepcion ante datos degenerados; en su lugar devuelve un dict con la
+    clave ``note`` y los campos numericos indefinidos a ``nan``.
+
+    Args:
+        data: muestra de observaciones numericas (lista de numeros).
+        other: segunda muestra independiente. Si se da, el IC es el de la
+            diferencia de medias ``mean(data) - mean(other)`` con Welch. Si es
+            ``None`` (default), el IC es el de la media de ``data``.
+        confidence: nivel de confianza en (0, 1), p.ej. 0.95 para el 95%.
+
+    Returns:
+        dict con las claves:
+            mean: media de ``data`` (una muestra) o la diferencia
+                ``mean(data) - mean(other)`` (dos muestras).
+            ci_low: extremo inferior del intervalo de confianza.
+            ci_high: extremo superior del intervalo de confianza.
+            se: error estandar de la media (o de la diferencia).
+            df: grados de libertad de la t (Welch-Satterthwaite si dos muestras).
+            confidence: nivel de confianza aplicado (float).
+            n: tamano de la muestra (una muestra) o tamano total ``n1 + n2``
+                (dos muestras; ademas se incluyen ``n1`` y ``n2``).
+
+        En el caso de dos muestras se incluyen ademas ``n1`` y ``n2``. Casos
+        degenerados (muestra vacia, ``n < 2``, etc.) anaden la clave ``note`` y
+        dejan ``ci_low``/``ci_high``/``se`` (y a veces ``df``) en ``nan``.
+    """
+    conf = float(confidence)
+
+    if other is None:
+        return _ci_one_sample(data, conf)
+    return _ci_welch(data, other, conf)
+
+
+def _ci_one_sample(data: list, conf: float) -> dict:
+    """IC de la media de una sola muestra con la t de Student."""
+    arr = np.asarray(list(data), dtype=float)
+    n = int(arr.size)
+
+    base = {
+        "mean": float("nan"),
+        "ci_low": float("nan"),
+        "ci_high": float("nan"),
+        "se": float("nan"),
+        "df": float("nan"),
+        "confidence": conf,
+        "n": n,
+    }
+
+    if n == 0:
+        base["note"] = "muestra vacia: media e intervalo indefinidos"
+        return base
+
+    mean = float(arr.mean())
+    base["mean"] = mean
+
+    if n < 2:
+        base["note"] = "n < 2: error estandar y grados de libertad indefinidos"
+        return base
+
+    df = n - 1
+    base["df"] = float(df)
+
+    sd = float(arr.std(ddof=1))
+    se = sd / math.sqrt(n)
+    base["se"] = se
+
+    # Varianza cero: el IC colapsa al punto (no es un error).
+    if se == 0.0:
+        base["ci_low"] = mean
+        base["ci_high"] = mean
+        base["note"] = "varianza cero: el intervalo colapsa a la media"
+        return base
+
+    tcrit = float(stats.t.ppf((1.0 + conf) / 2.0, df))
+    margin = tcrit * se
+    base["ci_low"] = mean - margin
+    base["ci_high"] = mean + margin
+    return base
+
+
+def _ci_welch(data: list, other: list, conf: float) -> dict:
+    """IC de la diferencia de medias de dos muestras con el metodo de Welch."""
+    a = np.asarray(list(data), dtype=float)
+    b = np.asarray(list(other), dtype=float)
+    n1 = int(a.size)
+    n2 = int(b.size)
+
+    base = {
+        "mean": float("nan"),
+        "ci_low": float("nan"),
+        "ci_high": float("nan"),
+        "se": float("nan"),
+        "df": float("nan"),
+        "confidence": conf,
+        "n": n1 + n2,
+        "n1": n1,
+        "n2": n2,
+    }
+
+    if n1 == 0 or n2 == 0:
+        base["note"] = "alguna muestra esta vacia: diferencia e intervalo indefinidos"
+        return base
+
+    mean1 = float(a.mean())
+    mean2 = float(b.mean())
+    diff = mean1 - mean2
+    base["mean"] = diff
+
+    if n1 < 2 or n2 < 2:
+        base["note"] = (
+            "n < 2 en alguna muestra: error estandar y grados de libertad indefinidos"
+        )
+        return base
+
+    sd1 = float(a.std(ddof=1))
+    sd2 = float(b.std(ddof=1))
+    se1 = sd1 / math.sqrt(n1)
+    se2 = sd2 / math.sqrt(n2)
+    se = math.sqrt(se1 * se1 + se2 * se2)
+    base["se"] = se
+
+    # Ambas varianzas cero: el IC de la diferencia colapsa al punto.
+    if se == 0.0:
+        base["ci_low"] = diff
+        base["ci_high"] = diff
+        base["df"] = float("nan")
+        base["note"] = "varianza cero en ambas muestras: el intervalo colapsa a la diferencia"
+        return base
+
+    # Grados de libertad de Welch-Satterthwaite.
+    df = (se1 * se1 + se2 * se2) ** 2 / (
+        (se1**4) / (n1 - 1) + (se2**4) / (n2 - 1)
+    )
+    base["df"] = float(df)
+
+    tcrit = float(stats.t.ppf((1.0 + conf) / 2.0, df))
+    margin = tcrit * se
+    base["ci_low"] = diff - margin
+    base["ci_high"] = diff + margin
+    return base

python/functions/datascience/confidence_interval_mean_test.py

@@ -0,0 +1,140 @@
+"""Tests para confidence_interval_mean (IC de la media / diferencia de medias Welch).
+
+Importa el modulo hoja directamente (`confidence_interval_mean`) para no depender
+de que el paquete reexporte la funcion en su __init__ (lo integra el orquestador
+al cerrar el grupo).
+
+Los golden se calculan con scipy dentro del propio test para que sean robustos:
+la funcion bajo prueba debe coincidir con la referencia de scipy a ~1e-9.
+"""
+
+import math
+
+import numpy as np
+from scipy import stats
+
+from confidence_interval_mean import confidence_interval_mean
+
+
+def test_one_sample_golden_contra_scipy():
+    # mean=5.0, n=8. Este dataset tiene sd POBLACIONAL (ddof=0) exactamente 2.0,
+    # pero la sd MUESTRAL (ddof=1, la que exige la spec y la que es correcta para
+    # el IC de una media con la t) es sqrt(32/7) ~ 2.13809. El golden robusto se
+    # calcula con scipy usando se con ddof=1, no con el atajo 2.0/sqrt(8).
+    data = [2, 4, 4, 4, 5, 5, 7, 9]
+    out = confidence_interval_mean(data, confidence=0.95)
+
+    n = len(data)
+    mean = float(np.mean(data))
+    sd = float(np.std(data, ddof=1))  # sample sd ~ 2.13809
+    se = sd / math.sqrt(n)
+    lo, hi = stats.t.interval(0.95, df=n - 1, loc=mean, scale=se)
+
+    assert abs(out["mean"] - 5.0) < 1e-9
+    assert abs(out["se"] - se) < 1e-12
+    assert out["df"] == 7.0
+    assert out["n"] == 8
+    assert out["confidence"] == 0.95
+    assert abs(out["ci_low"] - lo) < 1e-9
+    assert abs(out["ci_high"] - hi) < 1e-9
+    # Valores tabulados correctos para ddof=1 (no los 3.32793/6.67207 del
+    # enunciado, que asumian erroneamente sd=2.0 / ddof=0).
+    assert abs(out["ci_low"] - 3.21251) < 1e-3
+    assert abs(out["ci_high"] - 6.78749) < 1e-3
+    assert "note" not in out
+
+
+def test_one_sample_distinto_nivel_confianza():
+    data = [10.0, 12.0, 11.0, 13.0, 9.0, 14.0]
+    out = confidence_interval_mean(data, confidence=0.99)
+
+    n = len(data)
+    mean = float(np.mean(data))
+    se = float(np.std(data, ddof=1)) / math.sqrt(n)
+    lo, hi = stats.t.interval(0.99, df=n - 1, loc=mean, scale=se)
+
+    assert abs(out["mean"] - mean) < 1e-12
+    assert abs(out["ci_low"] - lo) < 1e-9
+    assert abs(out["ci_high"] - hi) < 1e-9
+    assert out["df"] == float(n - 1)
+
+
+def test_welch_diferencia_golden_contra_scipy():
+    data = [23.0, 21.0, 25.0, 22.0, 24.0, 26.0]
+    other = [18.0, 20.0, 17.0, 19.0, 21.0]
+    conf = 0.95
+    out = confidence_interval_mean(data, other, confidence=conf)
+
+    a = np.asarray(data, dtype=float)
+    b = np.asarray(other, dtype=float)
+    n1, n2 = a.size, b.size
+    mean1, mean2 = float(a.mean()), float(b.mean())
+    diff = mean1 - mean2
+    se1 = float(a.std(ddof=1)) / math.sqrt(n1)
+    se2 = float(b.std(ddof=1)) / math.sqrt(n2)
+    se = math.sqrt(se1**2 + se2**2)
+    df = (se1**2 + se2**2) ** 2 / (se1**4 / (n1 - 1) + se2**4 / (n2 - 1))
+    lo, hi = stats.t.interval(conf, df=df, loc=diff, scale=se)
+
+    assert abs(out["mean"] - diff) < 1e-9
+    assert abs(out["mean"] - (mean1 - mean2)) < 1e-9
+    assert abs(out["se"] - se) < 1e-12
+    assert abs(out["df"] - df) < 1e-9
+    assert abs(out["ci_low"] - lo) < 1e-9
+    assert abs(out["ci_high"] - hi) < 1e-9
+    assert out["n1"] == n1
+    assert out["n2"] == n2
+    assert out["n"] == n1 + n2
+    assert "note" not in out
+
+
+def test_edge_un_solo_elemento_no_lanza_nan_note():
+    out = confidence_interval_mean([5], confidence=0.95)
+    assert out["mean"] == 5.0  # la media si esta definida con n=1
+    assert math.isnan(out["se"])
+    assert math.isnan(out["ci_low"])
+    assert math.isnan(out["ci_high"])
+    assert math.isnan(out["df"])
+    assert out["n"] == 1
+    assert "note" in out
+
+
+def test_edge_lista_vacia_no_lanza_note():
+    out = confidence_interval_mean([], confidence=0.95)
+    assert math.isnan(out["mean"])
+    assert math.isnan(out["ci_low"])
+    assert math.isnan(out["ci_high"])
+    assert math.isnan(out["se"])
+    assert out["n"] == 0
+    assert "note" in out
+
+
+def test_edge_varianza_cero_colapsa_al_punto():
+    out = confidence_interval_mean([3, 3, 3], confidence=0.95)
+    assert out["mean"] == 3.0
+    assert out["se"] == 0.0
+    assert out["ci_low"] == 3.0
+    assert out["ci_high"] == 3.0
+    assert not math.isnan(out["ci_low"])
+    assert out["n"] == 3
+    assert "note" in out
+
+
+def test_edge_welch_muestra_vacia_no_lanza_note():
+    out = confidence_interval_mean([1.0, 2.0, 3.0], [], confidence=0.95)
+    assert math.isnan(out["mean"])
+    assert math.isnan(out["ci_low"])
+    assert math.isnan(out["se"])
+    assert out["n1"] == 3
+    assert out["n2"] == 0
+    assert "note" in out
+
+
+def test_edge_welch_n1_uno_no_lanza_note():
+    out = confidence_interval_mean([5.0], [1.0, 2.0, 3.0], confidence=0.95)
+    # La diferencia de medias si esta definida.
+    assert abs(out["mean"] - (5.0 - 2.0)) < 1e-9
+    assert math.isnan(out["se"])
+    assert math.isnan(out["ci_low"])
+    assert math.isnan(out["df"])
+    assert "note" in out

python/functions/datascience/detect_corpus_language.md

@@ -0,0 +1,80 @@
+---
+name: detect_corpus_language
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def detect_corpus_language(texts, top_k=10, sample_max=1000) -> dict"
+description: "Estima la distribucion de idiomas de un corpus de textos con la libreria langdetect (import perezoso). Funcion pura y defensiva del grupo eda: filtra documentos None/no-str/vacios, muestrea hasta sample_max docs, clasifica cada uno con detect() ignorando los que langdetect no puede resolver (LangDetectException), y devuelve la distribucion top_k por frecuencia mas el idioma dominante. Si langdetect no esta instalada o algo falla, degrada a {available: False, ...} y NUNCA lanza (dict-no-throw). Seed fija (DetectorFactory.seed=0) para deteccion determinista."
+tags: [eda, datascience, text, nlp, language-detection, langdetect, pure, python]
+params:
+  - name: texts
+    desc: "Lista de strings (documentos). Los elementos None, no-str o vacios tras strip se descartan antes de clasificar."
+  - name: top_k
+    desc: "Numero maximo de idiomas a devolver en distribution, ordenados por count descendente (desempate por codigo ISO ascendente). Default 10."
+  - name: sample_max
+    desc: "Numero maximo de documentos a clasificar (se toman los primeros del corpus) para acotar el coste. Default 1000."
+output: >
+  Dict con forma fija (dict-no-throw, nunca lanza):
+  {"available": bool, "n_detected": int,
+   "distribution": [{"lang": str, "count": int, "pct": float}, ...],
+   "dominant": str|None}.
+  available=True si langdetect es importable; lang son codigos ISO 639-1 ("es","en","fr",...);
+  pct = count/n_detected*100 redondeado a 2 decimales; n_detected = docs clasificados con exito;
+  dominant = idioma mas frecuente (None si no hubo detecciones). Corpus vacio con langdetect
+  presente -> available True, n_detected 0, distribution [], dominant None. Sin langdetect (o
+  fallo global) -> available False y el resto de campos a su valor vacio.
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [langdetect]
+tested: true
+tests: ["test_mixto_es_en", "test_vacio", "test_degradacion"]
+test_file_path: "python/functions/datascience/detect_corpus_language_test.py"
+file_path: "python/functions/datascience/detect_corpus_language.py"
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience.detect_corpus_language import detect_corpus_language
+
+corpus = [
+    "este es un texto bastante largo en español para detectar el idioma correctamente",
+    "la inteligencia artificial transforma la manera en que trabajamos cada dia",
+    "this is a fairly long english text to detect the language correctly without issues",
+]
+out = detect_corpus_language(corpus)
+# {"available": True, "n_detected": 3,
+#  "distribution": [{"lang": "es", "count": 2, "pct": 66.67},
+#                   {"lang": "en", "count": 1, "pct": 33.33}],
+#  "dominant": "es"}
+```
+
+## Cuando usarla
+
+Cuando perfiles una columna o corpus de texto en un EDA y necesites saber en
+que idioma(s) esta escrito antes de elegir tokenizadores, stopwords, modelos
+NLP o stemmers. Util tambien como check de calidad: detectar corpus mezclados
+o un idioma inesperado. Llamala con la lista de textos crudos; la funcion
+limpia, muestrea y resume sola.
+
+## Gotchas
+
+- `langdetect` es **opcional**: si no esta instalada, la funcion no lanza —
+  devuelve `{"available": False, "n_detected": 0, "distribution": [], "dominant": None}`.
+  Comprueba `out["available"]` antes de usar la distribucion.
+- **Textos cortos** (pocas palabras o sin features lingüisticas) pueden no
+  detectarse: langdetect lanza `LangDetectException`, que se ignora y el doc no
+  cuenta en `n_detected`. Pasa frases razonablemente largas para resultados fiables.
+- **Determinismo**: se fija `DetectorFactory.seed = 0` en cada llamada para que la
+  deteccion sea reproducible; sin esa semilla langdetect puede dar resultados
+  ligeramente distintos entre ejecuciones.
+- `distribution` esta truncada a `top_k`; si el corpus tiene mas idiomas que
+  `top_k`, la suma de los `count` mostrados puede ser menor que `n_detected`
+  (pero `dominant` siempre refleja el idioma mas frecuente del corpus completo).

python/functions/datascience/detect_corpus_language.py

@@ -0,0 +1,91 @@
+"""Detecta la distribucion de idiomas de un corpus de textos.
+
+Funcion pura y defensiva: el computo es determinista y local (sin I/O de red).
+La libreria opcional `langdetect` se importa de forma perezosa dentro de la
+funcion; si no esta instalada (o cualquier paso falla), la funcion degrada
+limpiamente a `available=False` y NUNCA lanza excepciones.
+"""
+
+
+def detect_corpus_language(texts, top_k=10, sample_max=1000) -> dict:
+    """Estima la distribucion de idiomas de un corpus con `langdetect`.
+
+    Args:
+        texts: lista de strings (documentos). Los elementos None, no-str o
+            vacios tras strip se descartan.
+        top_k: numero maximo de idiomas a devolver en `distribution`,
+            ordenados por frecuencia descendente.
+        sample_max: numero maximo de documentos a clasificar (se toman los
+            primeros) para acotar el coste.
+
+    Returns:
+        dict con la forma fija (dict-no-throw):
+        {
+            "available": bool,   # True si langdetect es importable
+            "n_detected": int,   # documentos clasificados con exito
+            "distribution": [{"lang": str, "count": int, "pct": float}, ...],
+            "dominant": str | None,
+        }
+    """
+    degraded = {
+        "available": False,
+        "n_detected": 0,
+        "distribution": [],
+        "dominant": None,
+    }
+    try:
+        # Import perezoso con degradacion: si langdetect no esta disponible,
+        # devolvemos el dict degradado sin lanzar.
+        try:
+            from langdetect import detect, DetectorFactory
+
+            # Semilla fija -> deteccion determinista entre ejecuciones.
+            DetectorFactory.seed = 0
+        except Exception:
+            return dict(degraded)
+
+        # Normaliza y filtra el corpus.
+        docs = []
+        if texts:
+            for t in texts:
+                if isinstance(t, str):
+                    s = t.strip()
+                    if s:
+                        docs.append(s)
+
+        # Muestreo de los primeros `sample_max` documentos.
+        if sample_max is not None and sample_max >= 0:
+            docs = docs[:sample_max]
+
+        # Conteo por idioma; langdetect lanza LangDetectException en textos
+        # sin features detectables -> se ignora y se sigue.
+        counts: dict = {}
+        for doc in docs:
+            try:
+                lang = detect(doc)
+            except Exception:
+                continue
+            counts[lang] = counts.get(lang, 0) + 1
+
+        n_detected = sum(counts.values())
+
+        # Orden estable: por count descendente, desempate por codigo de idioma.
+        ordered = sorted(counts.items(), key=lambda kv: (-kv[1], kv[0]))
+
+        k = top_k if (top_k is not None and top_k >= 0) else len(ordered)
+        distribution = []
+        for lang, count in ordered[:k]:
+            pct = round(count / n_detected * 100, 2) if n_detected else 0.0
+            distribution.append({"lang": lang, "count": count, "pct": pct})
+
+        dominant = ordered[0][0] if ordered else None
+
+        return {
+            "available": True,
+            "n_detected": n_detected,
+            "distribution": distribution,
+            "dominant": dominant,
+        }
+    except Exception:
+        # Cualquier fallo global degrada a available False sin lanzar.
+        return dict(degraded)

python/functions/datascience/detect_corpus_language_test.py

@@ -0,0 +1,58 @@
+"""Tests para detect_corpus_language."""
+
+import builtins
+import os
+import sys
+
+# Anade python/functions a sys.path para importar el paquete `datascience`.
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from datascience.detect_corpus_language import detect_corpus_language
+
+_ES = [
+    "este es un texto bastante largo en español para detectar el idioma correctamente sin problemas",
+    "la inteligencia artificial transforma la manera en que trabajamos cada dia en muchos sectores",
+]
+_EN = [
+    "this is a fairly long english text to detect the language correctly without any length issues",
+    "machine learning models can classify documents into many different categories quite reliably",
+]
+
+
+def test_mixto_es_en():
+    """Golden: corpus mixto ES+EN claro -> available True, >=2 idiomas, counts coherentes."""
+    out = detect_corpus_language(_ES + _EN)
+    assert out["available"] is True
+    assert out["dominant"] in {"es", "en"}
+    assert len(out["distribution"]) >= 2
+    total = sum(item["count"] for item in out["distribution"])
+    assert total == out["n_detected"]
+    assert out["n_detected"] == 4
+
+
+def test_vacio():
+    """Edge: lista vacia con langdetect presente -> available True, sin detecciones."""
+    out = detect_corpus_language([])
+    assert out["available"] is True
+    assert out["n_detected"] == 0
+    assert out["distribution"] == []
+    assert out["dominant"] is None
+
+
+def test_degradacion(monkeypatch):
+    """Error path: si langdetect no es importable -> degrada a available False sin lanzar."""
+    import datascience.detect_corpus_language as m
+
+    real_import = builtins.__import__
+
+    def fake_import(name, *a, **k):
+        if name == "langdetect" or name.startswith("langdetect."):
+            raise ImportError("simulado")
+        return real_import(name, *a, **k)
+
+    monkeypatch.setattr(builtins, "__import__", fake_import)
+    out = m.detect_corpus_language(["hola mundo", "hello world"])
+    assert out["available"] is False
+    assert out["n_detected"] == 0
+    assert out["distribution"] == []
+    assert out["dominant"] is None

python/functions/datascience/draw_join_graph_figure.md

@@ -0,0 +1,103 @@
+---
+id: draw_join_graph_figure_py_datascience
+name: draw_join_graph_figure
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def draw_join_graph_figure(join_graph: dict, title: str = None) -> \"matplotlib.figure.Figure\""
+description: "Rasteriza el join graph de una base (relaciones FK inter-tabla, salida de build_join_graph) a un matplotlib.figure.Figure: nodos circulares con el nombre de cada tabla (hubs en color de acento cálido, el resto neutro) y aristas dirigidas etiquetadas from_col→to_col (más la cardinalidad si viene). Es la contrapartida dibujada del string Mermaid para que el capítulo de relaciones del informe AutomaticEDA muestre un diagrama real. Layout networkx spring_layout determinista (seed=42), backend Agg sin abrir ventanas; defensivo: nunca lanza y nunca hace I/O."
+tags: [eda, plot, relations, graph, matplotlib, figure, networkx, datascience, impure]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: [matplotlib, networkx]
+example: |
+  from draw_join_graph_figure import draw_join_graph_figure
+  join_graph = {
+      "nodes": [
+          {"table": "customers", "out_degree": 0, "in_degree": 1, "role": "dimension"},
+          {"table": "orders", "out_degree": 1, "in_degree": 0, "role": "fact"},
+      ],
+      "edges": [
+          {"from_table": "orders", "from_col": "customer_id",
+           "to_table": "customers", "to_col": "id", "cardinality": "N:1"},
+      ],
+      "hubs": ["orders"],
+  }
+  fig = draw_join_graph_figure(join_graph, title="Relaciones FK")
+  fig.savefig("/tmp/join_graph.png")
+tested: true
+tests:
+  - "test_returns_figure_with_axis"
+  - "test_savefig_produces_nonempty_png"
+  - "test_empty_dict_does_not_raise_and_savefig_png"
+  - "test_none_does_not_raise_and_savefig_png"
+test_file_path: "python/functions/datascience/draw_join_graph_figure_test.py"
+file_path: "python/functions/datascience/draw_join_graph_figure.py"
+params:
+  - name: join_graph
+    desc: "Dict producido por build_join_graph. Claves: `nodes` (list[dict] con table, out_degree, in_degree, role), `edges` (list[dict] con from_table, from_col, to_table, to_col y opcional cardinality/inclusion) y `hubs` (list[str] de tablas hub a destacar en color cálido). Claves ausentes, items no-dict, None o {} se toleran (devuelve Figure con texto, sin lanzar). Los nombres de nodo se derivan también de las aristas, así que un grafo con edges pero sin nodes explícitos igual se dibuja."
+  - name: title
+    desc: "Título dibujado sobre el diagrama. Si se omite (None) se usa \"Join graph\". Default None."
+output: "Un matplotlib.figure.Figure (figsize 7x5) con un único Axes que contiene el diagrama node-link dirigido: tablas como nodos circulares etiquetados (hubs en acento cálido #DD8452, resto en azul neutro #4C72B0) y FKs como flechas dirigidas con etiqueta from_col→to_col (+ cardinalidad). Si join_graph no tiene nodos ni aristas (o es None/{}), devuelve igualmente una Figure con el texto centrado \"Sin relaciones FK detectadas.\"; ante cualquier fallo interno devuelve una Figure con un mensaje genérico (nunca lanza). El caller rasteriza/cierra la figura; la función no la muestra ni la guarda."
+---
+
+## Ejemplo
+
+```python
+from draw_join_graph_figure import draw_join_graph_figure
+
+# `join_graph` es la salida de build_join_graph (nodes + edges + hubs).
+join_graph = {
+    "nodes": [
+        {"table": "customers", "out_degree": 0, "in_degree": 1, "role": "dimension"},
+        {"table": "orders", "out_degree": 2, "in_degree": 0, "role": "fact"},
+        {"table": "products", "out_degree": 0, "in_degree": 1, "role": "dimension"},
+    ],
+    "edges": [
+        {"from_table": "orders", "from_col": "customer_id",
+         "to_table": "customers", "to_col": "id", "cardinality": "N:1"},
+        {"from_table": "orders", "from_col": "product_id",
+         "to_table": "products", "to_col": "id", "cardinality": "N:1"},
+    ],
+    "hubs": ["orders"],  # `orders` se pinta en color de acento (tabla de hechos)
+}
+
+fig = draw_join_graph_figure(join_graph, title="Relaciones FK")
+
+# El renderer del informe lo rasteriza; aquí solo persistimos para inspección.
+fig.savefig("/tmp/join_graph.png")
+```
+
+## Cuando usarla
+
+Úsala en el capítulo de relaciones de un informe AutomaticEDA cuando quieras un
+diagrama **dibujado** del esquema relacional, no solo el bloque Mermaid pegable.
+Pásale directamente la salida de `build_join_graph` (`nodes` + `edges` + `hubs`)
+y obtienes una `matplotlib.figure.Figure` lista para que el renderer perezoso la
+rasterice. Es la pareja visual del string Mermaid: Mermaid sirve para pegar en
+Markdown/docs que lo soporten; esta función produce la imagen real (PNG/PDF) que
+va embebida en informes que no renderizan Mermaid.
+
+## Gotchas
+
+- **Impura por matplotlib.** Fija el backend `Agg` al importar — no abre
+  ventanas ni depende de un display. Segura de llamar en lotes desde el
+  renderer.
+- **Layout determinista (`seed=42`).** Usa `nx.spring_layout(G, seed=42)`, así
+  que la misma entrada produce el mismo diagrama (test reproducible). Para
+  grafos de 0/1 nodos usa una posición fija centrada en vez del spring layout.
+- **No hace I/O.** No llama `plt.show()` ni guarda a disco — solo devuelve la
+  `Figure`. Quien la consume la rasteriza y la libera (`plt.close(fig)`) para no
+  acumular memoria en informes con muchas tablas.
+- **Devuelve una Figure, NO un dict.** A diferencia de `build_join_graph` (que
+  devuelve el dict del grafo), esta función devuelve el objeto de figura ya
+  dibujado.
+- **Defensiva, nunca lanza.** `None`, `{}`, claves ausentes o items malformados
+  se manejan sin error: en el peor caso devuelve una `Figure` con
+  "Sin relaciones FK detectadas." (vacío) o un mensaje genérico (fallo interno).
+  No la envuelvas en try/except por miedo a un raise — no lo hay.

python/functions/datascience/draw_join_graph_figure.py

@@ -0,0 +1,214 @@
+"""Impure EDA helper: rasterize a join graph to a matplotlib Figure (`eda` group).
+
+Takes the join graph produced by ``build_join_graph`` (inter-table FK relations)
+and draws it as a directed node-link diagram on a ready-to-rasterize
+``matplotlib.figure.Figure``. Hub tables (the ones with the highest out-degree,
+candidate fact tables of a star schema) are highlighted in a warm accent colour;
+the rest use a neutral colour. Directed edges carry a ``from_col→to_col`` label
+(plus the cardinality when present).
+
+This is the *drawn* counterpart of the Mermaid string that ``build_join_graph``
+also emits: the relations chapter of an AutomaticEDA report can show a real
+picture instead of only the pasteable Mermaid block.
+
+Impure because it touches matplotlib's rendering machinery. It pins the headless
+Agg backend and a deterministic ``spring_layout`` seed so the output is
+reproducible. It never raises: on any internal failure (or empty input) it
+returns a ``Figure`` carrying a centered message, so the lazy render of the
+document is never broken.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+import matplotlib.pyplot as plt  # noqa: E402
+import networkx as nx  # noqa: E402
+
+# Warm accent reserved for hub tables (candidate fact tables / star-schema cores).
+_HUB_COLOR = "#DD8452"
+# Neutral blue for every other table.
+_NODE_COLOR = "#4C72B0"
+# Muted gray for the empty/error message text.
+_MUTED_TEXT = "#5f6b7a"
+# Edge colour and label colour.
+_EDGE_COLOR = "#7a7a7a"
+_EDGE_LABEL_COLOR = "#34495e"
+# Constant node size; shared with the edge drawing so arrowheads stop at the
+# node boundary instead of being hidden under the marker.
+_NODE_SIZE = 2200
+
+
+def _text_figure(message: str) -> "matplotlib.figure.Figure":
+    """Return a blank Figure carrying a single centered message.
+
+    Used both for the "no relations" case and as the never-raise fallback.
+    """
+    fig, ax = plt.subplots(figsize=(7, 5))
+    ax.axis("off")
+    ax.text(
+        0.5,
+        0.5,
+        message,
+        ha="center",
+        va="center",
+        fontsize=12,
+        color=_MUTED_TEXT,
+        transform=ax.transAxes,
+    )
+    fig.tight_layout()
+    return fig
+
+
+def _edge_label(edge: dict) -> str:
+    """Build the ``from_col→to_col`` label of an edge, appending cardinality."""
+    fc = edge.get("from_col")
+    tc = edge.get("to_col")
+    if fc is not None and tc is not None:
+        label = f"{fc}→{tc}"
+    elif fc is not None:
+        label = str(fc)
+    elif tc is not None:
+        label = str(tc)
+    else:
+        label = ""
+    card = edge.get("cardinality")
+    if card:
+        label = f"{label} ({card})" if label else str(card)
+    return label
+
+
+def draw_join_graph_figure(join_graph: dict, title: str = None):
+    """Rasterize a join graph to a matplotlib Figure.
+
+    Builds a ``networkx.DiGraph`` from the graph's nodes and edges, lays it out
+    with a deterministic ``spring_layout`` (``seed=42``) and draws it on a
+    ``matplotlib.figure.Figure``: tables as labelled circular nodes (hubs in a
+    warm accent, the rest neutral) and FK relations as directed arrows labelled
+    ``from_col→to_col`` (plus cardinality when available).
+
+    The function never raises. On empty/``None`` input it returns a Figure with
+    a centered "Sin relaciones FK detectadas." message; on any internal failure
+    it returns a Figure with a generic centered message. It never shows the
+    figure nor writes it to disk — the document renderer rasterizes it.
+
+    Args:
+        join_graph: Dict produced by ``build_join_graph`` with keys ``nodes``
+            (list of ``{table, out_degree, in_degree, role}``), ``edges`` (list
+            of ``{from_table, from_col, to_table, to_col, cardinality?,
+            inclusion?}``) and ``hubs`` (list of hub table names to highlight).
+            Missing keys, non-dict items, ``None`` or ``{}`` are all tolerated.
+        title: Optional title drawn above the diagram. When omitted, the title
+            defaults to "Join graph".
+
+    Returns:
+        A ``matplotlib.figure.Figure`` (figsize 7x5) with a single Axes holding
+        the node-link diagram. The caller rasterizes/closes it.
+    """
+    try:
+        jg = join_graph if isinstance(join_graph, dict) else {}
+        nodes = jg.get("nodes") or []
+        edges = jg.get("edges") or []
+        hubs = {h for h in (jg.get("hubs") or []) if h is not None}
+
+        # Collect node names from the declared nodes and, defensively, from the
+        # edges (so a graph with edges but no explicit nodes still draws).
+        node_names: list = []
+        seen: set = set()
+
+        def _register(name) -> None:
+            if name is not None and name not in seen:
+                seen.add(name)
+                node_names.append(name)
+
+        for n in nodes:
+            if isinstance(n, dict):
+                _register(n.get("table"))
+        for e in edges:
+            if isinstance(e, dict):
+                _register(e.get("from_table"))
+                _register(e.get("to_table"))
+
+        if not node_names:
+            return _text_figure("Sin relaciones FK detectadas.")
+
+        graph = nx.DiGraph()
+        for name in node_names:
+            graph.add_node(name)
+
+        edge_labels: dict = {}
+        for e in edges:
+            if not isinstance(e, dict):
+                continue
+            ft = e.get("from_table")
+            tt = e.get("to_table")
+            if ft is None or tt is None:
+                continue
+            graph.add_edge(ft, tt)
+            edge_labels[(ft, tt)] = _edge_label(e)
+
+        fig, ax = plt.subplots(figsize=(7, 5))
+
+        # Deterministic layout. Fixed positions for trivial graphs so a single
+        # node sits centered instead of at an arbitrary spring-layout point.
+        if graph.number_of_nodes() <= 1:
+            pos = {name: (0.5, 0.5) for name in graph.nodes()}
+        else:
+            pos = nx.spring_layout(graph, seed=42)
+
+        node_colors = [
+            _HUB_COLOR if name in hubs else _NODE_COLOR for name in graph.nodes()
+        ]
+        nx.draw_networkx_nodes(
+            graph,
+            pos,
+            ax=ax,
+            node_color=node_colors,
+            node_size=_NODE_SIZE,
+            node_shape="o",
+            edgecolors="white",
+            linewidths=1.5,
+        )
+        nx.draw_networkx_labels(
+            graph,
+            pos,
+            ax=ax,
+            font_size=9,
+            font_color="white",
+            font_weight="bold",
+        )
+        nx.draw_networkx_edges(
+            graph,
+            pos,
+            ax=ax,
+            arrows=True,
+            arrowstyle="-|>",
+            arrowsize=18,
+            edge_color=_EDGE_COLOR,
+            width=1.4,
+            connectionstyle="arc3,rad=0.06",
+            node_size=_NODE_SIZE,
+        )
+        if any(lbl for lbl in edge_labels.values()):
+            nx.draw_networkx_edge_labels(
+                graph,
+                pos,
+                edge_labels=edge_labels,
+                ax=ax,
+                font_size=7,
+                font_color=_EDGE_LABEL_COLOR,
+                bbox={
+                    "boxstyle": "round,pad=0.2",
+                    "fc": "white",
+                    "ec": "none",
+                    "alpha": 0.7,
+                },
+            )
+
+        ax.set_title(title if title else "Join graph", fontsize=13)
+        ax.axis("off")
+        fig.tight_layout()
+        return fig
+    except Exception:
+        # Never raise — the document render is lazy and must not be broken.
+        return _text_figure("No se pudo dibujar el join graph.")

python/functions/datascience/draw_join_graph_figure_test.py

@@ -0,0 +1,84 @@
+"""Tests para draw_join_graph_figure (rasteriza el join graph, grupo eda).
+
+Usa el backend Agg sin abrir ventanas; cada test cierra la Figure construida
+(matplotlib.pyplot.close) para no acumular estado entre tests. Las aserciones de
+guardado escriben a tmp_path (fixture de pytest) y comprueban que el PNG no está
+vacío.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.figure import Figure  # noqa: E402
+
+from draw_join_graph_figure import draw_join_graph_figure
+
+
+def _make_join_graph():
+    """Join graph mínimo: 3 nodos (customers/orders/products) y 2 aristas.
+
+    orders -> customers y orders -> products. `orders` es el hub (out_degree 2).
+    """
+    return {
+        "nodes": [
+            {"table": "customers", "out_degree": 0, "in_degree": 1, "role": "dimension"},
+            {"table": "orders", "out_degree": 2, "in_degree": 0, "role": "fact"},
+            {"table": "products", "out_degree": 0, "in_degree": 1, "role": "dimension"},
+        ],
+        "edges": [
+            {
+                "from_table": "orders",
+                "from_col": "customer_id",
+                "to_table": "customers",
+                "to_col": "id",
+                "cardinality": "N:1",
+                "inclusion": 1.0,
+            },
+            {
+                "from_table": "orders",
+                "from_col": "product_id",
+                "to_table": "products",
+                "to_col": "id",
+                "cardinality": "N:1",
+                "inclusion": 0.98,
+            },
+        ],
+        "hubs": ["orders"],
+    }
+
+
+def test_returns_figure_with_axis():
+    fig = draw_join_graph_figure(_make_join_graph(), title="Relaciones FK")
+    assert isinstance(fig, Figure)
+    # Al menos un eje con el diagrama.
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_savefig_produces_nonempty_png(tmp_path):
+    fig = draw_join_graph_figure(_make_join_graph())
+    out = tmp_path / "g.png"
+    fig.savefig(out)
+    assert out.exists()
+    assert out.stat().st_size > 0
+    plt.close(fig)
+
+
+def test_empty_dict_does_not_raise_and_savefig_png(tmp_path):
+    fig = draw_join_graph_figure({})
+    assert isinstance(fig, Figure)
+    out = tmp_path / "empty.png"
+    fig.savefig(out)
+    assert out.stat().st_size > 0
+    plt.close(fig)
+
+
+def test_none_does_not_raise_and_savefig_png(tmp_path):
+    fig = draw_join_graph_figure(None)
+    assert isinstance(fig, Figure)
+    out = tmp_path / "none.png"
+    fig.savefig(out)
+    assert out.stat().st_size > 0
+    plt.close(fig)

python/functions/datascience/effect_size_cohens_d.md

@@ -0,0 +1,80 @@
+---
+name: effect_size_cohens_d
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def effect_size_cohens_d(group_a: list, group_b: list) -> dict"
+description: "Tamano del efecto (effect size) entre dos grupos numericos: Cohen's d (diferencia de medias estandarizada por la desviacion tipica combinada, varianzas muestrales ddof=1), Hedges' g (d corregido por el sesgo al alza con muestras pequenas via el factor J) e interpretacion cualitativa de la magnitud segun los umbrales clasicos de Cohen (negligible/small/medium/large). El p-valor dice si hay diferencia; el effect size dice como de grande, de forma adimensional e independiente del N. Pura, sin dependencias externas; nunca lanza: los casos degenerados (varianza cero, N<2, listas vacias) devuelven NaN + una clave note."
+tags: [papers, statistics, effect-size, cohens-d, hedges-g, python]
+params:
+  - name: group_a
+    desc: "primera muestra (lista de numeros). Necesita >=2 observaciones para que exista la varianza muestral (ddof=1)."
+  - name: group_b
+    desc: "segunda muestra (lista de numeros). Necesita >=2 observaciones. El signo de cohens_d es positivo cuando mean_a > mean_b."
+output: "dict {cohens_d: float (diferencia de medias estandarizada, puede ser NaN), hedges_g: float (cohens_d * factor de correccion J, puede ser NaN), interpretation: str ('negligible'|'small'|'medium'|'large', o 'undefined' en casos degenerados), n_a: int, n_b: int, mean_a: float, mean_b: float, pooled_sd: float (desviacion tipica combinada)}. Casos degenerados (varianza cero en ambos grupos, N<2 en algun grupo, o listas vacias) anaden clave note. Nunca None ni excepcion."
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: [math]
+tested: true
+tests: ["test_golden_large_effect", "test_hedges_g_menor_en_magnitud_que_cohens_d", "test_interpretation_thresholds", "test_signo_positivo_cuando_a_mayor_que_b", "test_varianza_cero_no_lanza", "test_n_insuficiente_no_lanza", "test_listas_vacias_no_lanza", "test_un_grupo_vacio_no_lanza"]
+test_file_path: "python/functions/datascience/effect_size_cohens_d_test.py"
+file_path: "python/functions/datascience/effect_size_cohens_d.py"
+---
+
+## Ejemplo
+
+```python
+from datascience import effect_size_cohens_d
+
+# Dos grupos desplazados 2 unidades, misma dispersion.
+a = [1, 2, 3, 4, 5]   # media 3, varianza muestral 2.5
+b = [3, 4, 5, 6, 7]   # media 5, varianza muestral 2.5
+
+out = effect_size_cohens_d(a, b)
+print(out["cohens_d"])        # -> -1.264911...  (a esta 1.26 SD por debajo de b)
+print(out["hedges_g"])        # -> -1.142500...  (|g| < |d|: correccion N pequeno)
+print(out["interpretation"])  # -> "large"       (|d| >= 0.8)
+print(out["pooled_sd"])       # -> 1.581138...
+
+# Caso degenerado: varianza cero -> no lanza, NaN + note.
+deg = effect_size_cohens_d([5, 5, 5], [5, 5, 5])
+print(deg["interpretation"])  # -> "undefined"
+print(deg["note"])            # -> "varianza cero, effect size indefinido"
+```
+
+## Cuando usarla
+
+Cuando ya sepas que dos grupos difieren (o quieras cuantificar su diferencia)
+y necesites una medida **de magnitud, no de significancia**: comparar el antes
+y el despues de una intervencion, el grupo control frente al tratamiento, o dos
+cohortes. Reportala junto al p-valor para responder "¿como de grande es la
+diferencia?" — un p-valor minusculo con N enorme puede esconder un efecto
+trivial. Es adimensional (en unidades de desviaciones tipicas), asi que hace
+comparables resultados entre estudios y alimenta meta-analisis. Usa **Hedges' g**
+en lugar de Cohen's d cuando los grupos sean pequenos (decenas o menos): d
+sobreestima el efecto y g lo corrige.
+
+## Gotchas
+
+- Pura y sin dependencias externas (solo `math` de la stdlib).
+- Usa **varianza muestral** (ddof=1), no poblacional. Por eso cada grupo
+  necesita al menos 2 observaciones; con N=1 la varianza muestral no existe y la
+  funcion devuelve NaN + `note`.
+- **Nunca lanza excepcion**. Los casos degenerados devuelven `cohens_d` y
+  `hedges_g` a `float('nan')`, `interpretation="undefined"` y una clave `note`:
+  varianza cero en ambos grupos (`pooled_sd == 0`), N<2 en algun grupo, o listas
+  vacias. Comprueba con `math.isnan(out["cohens_d"])` o la presencia de `note`
+  antes de usar el resultado.
+- El **signo** de `cohens_d` depende del orden de los argumentos: positivo si
+  `mean_a > mean_b`, negativo en caso contrario. La `interpretation` usa `|d|`,
+  asi que no depende del orden.
+- `pooled_sd` asume varianzas comparables entre grupos (homogeneidad). Si las
+  dispersiones son muy distintas, Cohen's d clasico pierde precision; considera
+  variantes (Glass's delta) fuera del alcance de esta funcion.
+- Los umbrales de Cohen (0.2 / 0.5 / 0.8) son convencion, no ley: interpretalos
+  segun el dominio.

python/functions/datascience/effect_size_cohens_d.py

@@ -0,0 +1,156 @@
+"""Effect size de dos grupos: Cohen's d, Hedges' g e interpretacion cualitativa.
+
+Funcion pura del grupo papers. El p-valor responde a "¿hay diferencia?" pero no
+a "¿como de grande es?". El tamano del efecto (effect size) cuantifica la
+magnitud de la diferencia entre dos grupos de forma adimensional, independiente
+del N, y es lo que hace comparables resultados entre estudios (meta-analisis).
+
+- Cohen's d: diferencia de medias estandarizada por la desviacion tipica
+  combinada (pooled SD), con varianzas muestrales (ddof=1).
+- Hedges' g: Cohen's d corregido por el sesgo al alza que sufre d con muestras
+  pequenas, multiplicando por el factor de correccion J.
+- interpretation: etiqueta cualitativa de |d| segun los umbrales clasicos de
+  Cohen (negligible / small / medium / large).
+
+No usa dependencias externas: aritmetica de la libreria estandar (``math``).
+"""
+
+from __future__ import annotations
+
+import math
+
+
+def _mean(xs: list) -> float:
+    """Media aritmetica de una lista no vacia de numeros."""
+    return sum(float(x) for x in xs) / len(xs)
+
+
+def _sample_variance(xs: list, mean: float) -> float:
+    """Varianza muestral (ddof=1) de una lista con al menos 2 elementos."""
+    n = len(xs)
+    return sum((float(x) - mean) ** 2 for x in xs) / (n - 1)
+
+
+def _interpret(abs_d: float) -> str:
+    """Etiqueta cualitativa del tamano del efecto segun |d| (umbrales de Cohen)."""
+    if abs_d < 0.2:
+        return "negligible"
+    if abs_d < 0.5:
+        return "small"
+    if abs_d < 0.8:
+        return "medium"
+    return "large"
+
+
+def effect_size_cohens_d(group_a: list, group_b: list) -> dict:
+    """Calcula el tamano del efecto entre dos grupos numericos.
+
+    Devuelve Cohen's d (diferencia de medias estandarizada por la pooled SD),
+    Hedges' g (d corregido por sesgo de muestra pequena) y una etiqueta
+    cualitativa de la magnitud segun los umbrales de Cohen.
+
+    Es una funcion pura y determinista: no hace I/O, no muta la entrada. No lanza
+    excepcion ante datos degenerados; en su lugar devuelve un dict con
+    ``cohens_d`` / ``hedges_g`` a ``float('nan')``, ``interpretation`` a
+    ``"undefined"`` y una clave ``note`` explicando el caso.
+
+    Definiciones:
+        s_pooled = sqrt(((n1-1)*s1^2 + (n2-1)*s2^2) / (n1+n2-2)), con s1^2, s2^2
+            varianzas muestrales (ddof=1).
+        cohens_d = (mean_a - mean_b) / s_pooled.
+        J = 1 - 3 / (4*(n1+n2) - 9)  (factor de correccion de Hedges).
+        hedges_g = cohens_d * J.
+
+    Args:
+        group_a: primera muestra (lista de numeros). Necesita >=2 elementos para
+            que exista la varianza muestral.
+        group_b: segunda muestra (lista de numeros). Necesita >=2 elementos.
+
+    Returns:
+        dict con las claves:
+            cohens_d: float, diferencia de medias estandarizada (puede ser NaN).
+            hedges_g: float, Cohen's d corregido por sesgo (puede ser NaN).
+            interpretation: str, "negligible" | "small" | "medium" | "large", o
+                "undefined" en casos degenerados.
+            n_a: int, tamano de group_a.
+            n_b: int, tamano de group_b.
+            mean_a: float, media de group_a (NaN si vacio).
+            mean_b: float, media de group_b (NaN si vacio).
+            pooled_sd: float, desviacion tipica combinada (NaN si indefinida).
+
+        Casos degenerados (lista vacia, N<2 en algun grupo, o varianza cero en
+        ambos grupos -> pooled_sd == 0) anaden ademas una clave ``note``.
+    """
+    nan = float("nan")
+    n_a = len(group_a)
+    n_b = len(group_b)
+
+    # Listas vacias: ni media ni varianza definidas.
+    if n_a == 0 or n_b == 0:
+        return {
+            "cohens_d": nan,
+            "hedges_g": nan,
+            "interpretation": "undefined",
+            "n_a": n_a,
+            "n_b": n_b,
+            "mean_a": _mean(group_a) if n_a else nan,
+            "mean_b": _mean(group_b) if n_b else nan,
+            "pooled_sd": nan,
+            "note": "grupo vacio: media y varianza indefinidas, effect size indefinido",
+        }
+
+    mean_a = _mean(group_a)
+    mean_b = _mean(group_b)
+
+    # N insuficiente: la varianza muestral (ddof=1) no existe con un solo dato,
+    # y la correccion de Hedges no es fiable.
+    if n_a < 2 or n_b < 2:
+        return {
+            "cohens_d": nan,
+            "hedges_g": nan,
+            "interpretation": "undefined",
+            "n_a": n_a,
+            "n_b": n_b,
+            "mean_a": mean_a,
+            "mean_b": mean_b,
+            "pooled_sd": nan,
+            "note": (
+                "N insuficiente: cada grupo necesita >=2 observaciones para la "
+                "varianza muestral; effect size indefinido"
+            ),
+        }
+
+    var_a = _sample_variance(group_a, mean_a)
+    var_b = _sample_variance(group_b, mean_b)
+    pooled_sd = math.sqrt(
+        ((n_a - 1) * var_a + (n_b - 1) * var_b) / (n_a + n_b - 2)
+    )
+
+    # Varianza cero en ambos grupos: no se puede estandarizar (division por 0).
+    if pooled_sd == 0.0:
+        return {
+            "cohens_d": nan,
+            "hedges_g": nan,
+            "interpretation": "undefined",
+            "n_a": n_a,
+            "n_b": n_b,
+            "mean_a": mean_a,
+            "mean_b": mean_b,
+            "pooled_sd": 0.0,
+            "note": "varianza cero, effect size indefinido",
+        }
+
+    cohens_d = (mean_a - mean_b) / pooled_sd
+    j = 1.0 - 3.0 / (4.0 * (n_a + n_b) - 9.0)
+    hedges_g = cohens_d * j
+
+    return {
+        "cohens_d": cohens_d,
+        "hedges_g": hedges_g,
+        "interpretation": _interpret(abs(cohens_d)),
+        "n_a": n_a,
+        "n_b": n_b,
+        "mean_a": mean_a,
+        "mean_b": mean_b,
+        "pooled_sd": pooled_sd,
+    }

python/functions/datascience/effect_size_cohens_d_test.py

@@ -0,0 +1,96 @@
+"""Tests para effect_size_cohens_d (tamano del efecto de dos grupos).
+
+Importa el modulo hoja directamente (`effect_size_cohens_d`) para no depender de
+que el paquete reexporte la funcion en su __init__ (lo integra el orquestador al
+cerrar el grupo papers). El pytest del repo tiene pythonpath=["functions", ...],
+asi que el modulo hoja se resuelve por su nombre directo.
+"""
+
+import math
+
+from effect_size_cohens_d import effect_size_cohens_d
+
+
+def test_golden_large_effect():
+    # group_a: mean 3, var muestral 2.5; group_b: mean 5, var 2.5.
+    # pooled_sd = sqrt(2.5) ~= 1.5811388.
+    # cohens_d = (3-5)/1.5811388 ~= -1.264911.
+    # J = 1 - 3/(4*10-9) = 1 - 3/31 = 0.9032258.
+    # hedges_g = d * J = -1.2649111 * 0.9032258 ~= -1.142500.
+    out = effect_size_cohens_d([1, 2, 3, 4, 5], [3, 4, 5, 6, 7])
+    assert abs(out["cohens_d"] - (-1.26491)) < 1e-4
+    assert abs(out["hedges_g"] - (-1.14250)) < 1e-4
+    assert out["interpretation"] == "large"
+    assert out["n_a"] == 5
+    assert out["n_b"] == 5
+    assert abs(out["mean_a"] - 3.0) < 1e-12
+    assert abs(out["mean_b"] - 5.0) < 1e-12
+    assert abs(out["pooled_sd"] - math.sqrt(2.5)) < 1e-9
+    assert "note" not in out
+
+
+def test_hedges_g_menor_en_magnitud_que_cohens_d():
+    # La correccion J esta en (0, 1), asi que |g| < |d| siempre.
+    out = effect_size_cohens_d([1, 2, 3, 4, 5], [3, 4, 5, 6, 7])
+    assert abs(out["hedges_g"]) < abs(out["cohens_d"])
+
+
+def test_interpretation_thresholds():
+    # negligible: |d| < 0.2. Medias casi iguales con varianza grande.
+    neg = effect_size_cohens_d([0, 10, 20, 30], [1, 11, 21, 31])
+    assert neg["interpretation"] == "negligible"
+    assert abs(neg["cohens_d"]) < 0.2
+
+    # small: 0.2 <= |d| < 0.5.
+    small = effect_size_cohens_d([0, 10, 20, 30], [4, 14, 24, 34])
+    assert small["interpretation"] == "small"
+    assert 0.2 <= abs(small["cohens_d"]) < 0.5
+
+    # medium: 0.5 <= |d| < 0.8.
+    medium = effect_size_cohens_d([0, 10, 20, 30], [9, 19, 29, 39])
+    assert medium["interpretation"] == "medium"
+    assert 0.5 <= abs(medium["cohens_d"]) < 0.8
+
+
+def test_signo_positivo_cuando_a_mayor_que_b():
+    out = effect_size_cohens_d([10, 12, 14, 16], [1, 2, 3, 4])
+    assert out["cohens_d"] > 0
+    assert out["interpretation"] == "large"
+
+
+def test_varianza_cero_no_lanza():
+    out = effect_size_cohens_d([5, 5, 5], [5, 5, 5])
+    assert math.isnan(out["cohens_d"])
+    assert math.isnan(out["hedges_g"])
+    assert out["interpretation"] == "undefined"
+    assert out["pooled_sd"] == 0.0
+    assert "note" in out
+    assert "varianza cero" in out["note"]
+
+
+def test_n_insuficiente_no_lanza():
+    out = effect_size_cohens_d([3], [1, 2, 3])
+    assert math.isnan(out["cohens_d"])
+    assert math.isnan(out["hedges_g"])
+    assert out["interpretation"] == "undefined"
+    assert out["n_a"] == 1
+    assert out["n_b"] == 3
+    assert "note" in out
+
+
+def test_listas_vacias_no_lanza():
+    out = effect_size_cohens_d([], [])
+    assert math.isnan(out["cohens_d"])
+    assert math.isnan(out["hedges_g"])
+    assert out["interpretation"] == "undefined"
+    assert out["n_a"] == 0
+    assert out["n_b"] == 0
+    assert "note" in out
+
+
+def test_un_grupo_vacio_no_lanza():
+    out = effect_size_cohens_d([1, 2, 3], [])
+    assert math.isnan(out["cohens_d"])
+    assert out["interpretation"] == "undefined"
+    assert out["n_b"] == 0
+    assert "note" in out

python/functions/datascience/extract_null_mask.md

@@ -0,0 +1,97 @@
+---
+name: extract_null_mask
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def extract_null_mask(query_fn, table: str, columns: list, max_rows: int = 5000) -> dict"
+description: "Extrae la mascara de nulos (1=falta / 0=presente) de una muestra de filas de una tabla, una lista 0/1 por columna alineada por fila, para alimentar el capitulo de calidad / patron de nulos de AutomaticEDA sin que el capitulo toque la base de datos. Recibe un lector read-only inyectado `query_fn(sql) -> dict` (mismo contrato que duckdb_query_readonly / pg_query / el `_q` de profile_table) y NO abre ninguna conexion por su cuenta. Construye UNA sola query que proyecta por cada columna `CASE WHEN \"col\" IS NULL THEN 1 ELSE 0 END` con identificadores escapados y LIMIT. Devuelve dict dict-no-throw: columns (efectivamente leidas, en orden), mask (lista int 0/1 por columna, misma longitud todas) y n. Una celda None se cuenta defensivamente como 1 (falta)."
+tags: [eda, nulls, missing, datascience, automatic-eda, extraction, read-only, duckdb, postgres, python]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+params:
+  - name: query_fn
+    desc: "callable lector read-only del backend activo. Recibe un string SQL y devuelve un dict {'status':'ok','rows':[{col:val,...},...]} (mismo contrato que duckdb_query_readonly o el `_q` de profile_table). NO se abre ninguna conexion dentro de la funcion: toda la lectura pasa por query_fn. Si es None -> error."
+  - name: table
+    desc: "nombre de la tabla de la que muestrear la mascara de nulos. Se escapa con comillas dobles en la query. Vacio o None -> status error."
+  - name: columns
+    desc: "lista de nombres de columna a evaluar. Cada una produce una entrada en `mask` con una lista 0/1 paralela por fila (1=IS NULL, 0=presente). Cada nombre se escapa con comillas dobles. Vacia o None -> status error."
+  - name: max_rows
+    desc: "limite de filas a muestrear (clausula LIMIT). Default 5000. Protege frente a tablas enormes; con LIMIT obtienes el primer tramo, no un muestreo uniforme."
+output: "dict (nunca lanza). En exito: {'status':'ok','table':str,'columns':[str,...] (en orden),'mask':{col:[int 0/1,...],...} (1=falta/IS NULL, 0=presente; todas las listas con misma longitud = n),'n':int}. En error (sin lanzar): {'status':'error','error':str,'table':str,'columns':[],'mask':{},'n':0}. Errores: query_fn None, table vacia, columns vacia, o query_fn devuelve status!='ok' (se propaga su error)."
+tested: true
+tests: ["test_golden_mask_alineada", "test_celda_none_cuenta_como_falta", "test_columns_vacia_status_error", "test_query_fn_status_error_propaga", "test_query_fn_none_da_error_sin_reventar", "test_sql_contiene_case_y_limit"]
+test_file_path: "python/functions/datascience/extract_null_mask_test.py"
+file_path: "python/functions/datascience/extract_null_mask.py"
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience.extract_null_mask import extract_null_mask
+from infra import duckdb_query_readonly
+
+# El lector read-only se inyecta como closure (igual que el `_q` de profile_table).
+db = "data/clientes.duckdb"
+def _q(sql):
+    return duckdb_query_readonly(db, sql)
+
+res = extract_null_mask(_q, "clientes", ["email", "telefono", "edad"])
+# res == {
+#   "status": "ok",
+#   "table": "clientes",
+#   "columns": ["email", "telefono", "edad"],
+#   "mask": {
+#     "email":    [0, 0, 1, 0, ...],   # fila 2 sin email
+#     "telefono": [1, 0, 1, 0, ...],
+#     "edad":     [0, 0, 0, 1, ...],
+#   },
+#   "n": 5000,
+# }
+
+# % de nulos por columna a partir de la muestra:
+pct = {c: 100 * sum(bits) / max(res["n"], 1) for c, bits in res["mask"].items()}
+
+# Se entrega al capitulo de calidad sin que este toque la BD:
+ctx = {"null_mask": res}
+```
+
+## Cuando usarla
+
+Cuando el capitulo de calidad / patron de nulos de AutomaticEDA necesita saber
+DONDE faltan los valores (no solo cuantos) y NO debe abrir la base de datos por
+su cuenta: extraes aqui la mascara 0/1 por columna alineada por fila y se la pasas
+en `ctx['null_mask']`. Usala siempre que quieras detectar co-ocurrencia de nulos
+(filas que fallan en varias columnas a la vez), calcular el % de nulos sobre una
+muestra, o pintar un heatmap de missingness reutilizando un unico lector read-only
+inyectado, en vez de hacer N `COUNT(*) WHERE col IS NULL` por separado.
+
+## Gotchas
+
+- **Impura**: lee de la base de datos a traves de `query_fn`. No abre conexiones
+  por su cuenta — depende por completo del lector inyectado. Sigue el estilo
+  dict-no-throw del grupo `eda`: nunca lanza; ante cualquier fallo devuelve
+  `{"status":"error","error":...}` con `columns=[]`, `mask={}`, `n=0`.
+- **`error_type` en el frontmatter es `error_go_core` por convencion del registry**
+  (toda funcion impura debe declararlo y el indexer lo exige), pero el codigo
+  NO lanza esa excepcion: degrada al dict de error. Es metadata, no comportamiento.
+- **Muestra, no censo**: con `LIMIT max_rows` obtienes el primer tramo de filas que
+  devuelva el backend, no un muestreo uniforme ni la tabla entera. El % de nulos
+  derivado es una estimacion sobre esa muestra; para el conteo exacto usa un
+  agregado `COUNT(*)`/`COUNT(col)` aparte.
+- **Alineacion por fila**: `mask[col][i]` corresponde a la misma fila `i` que
+  `mask[otra_col][i]`. Todas las listas tienen longitud `n`, asi que puedes cruzar
+  columnas por indice (co-ocurrencia de nulos) sin re-alinear.
+- **Defensa None -> 1**: el SQL ya devuelve 0/1, pero si una celda llega como `None`
+  (CASE no aplicado, columna ausente en la fila, backend que nulifica) se cuenta
+  como 1 (falta). Un valor inesperado no convertible a int se trata como presente (0).
+- **No loguear los datos crudos**: aunque `mask` es solo 0/1, los nombres de columna
+  pueden revelar el esquema. En trazas usa `n` y el numero de columnas, no el dict
+  completo.

python/functions/datascience/extract_null_mask.py

@@ -0,0 +1,101 @@
+"""extract_null_mask — extrae la mascara de nulos (1=falta / 0=presente) de una tabla.
+
+Lector read-only inyectado: recibe `query_fn(sql) -> dict` con el mismo contrato
+que duckdb_query_readonly / pg_query (y que el `_q` de profile_table):
+`{"status": "ok", "rows": [{col: val, ...}, ...]}`. Esta funcion NO abre ninguna
+conexion por su cuenta — solo usa `query_fn`. Construye UNA sola query que, por
+cada columna pedida, evalua `CASE WHEN "col" IS NULL THEN 1 ELSE 0 END` y devuelve
+una muestra de filas con esos bits. El resultado es un dict `mask` con una lista
+0/1 por columna, alineada por fila (1 = el valor falta / IS NULL, 0 = presente),
+listo para alimentar el capitulo de calidad / patron de nulos de AutomaticEDA sin
+que el capitulo toque la base de datos.
+
+Estilo dict-no-throw del grupo `eda`: nunca lanza; captura cualquier excepcion y
+degrada a `{"status": "error", "error": str, ...}`.
+"""
+
+
+def _to_bit(value):
+    """Coacciona el valor 0/1 del CASE a int de forma defensiva.
+
+    El SQL ya devuelve 0 (presente) o 1 (falta). Por si una celda llega como None
+    (el CASE no se aplico o el backend la nulifico), se cuenta como 1 (falta). El
+    resto se reduce a int: un entero distinto de 0 cuenta como 1 (falta), 0 como
+    presente. Un valor no convertible se trata como presente (0) — nunca lanza.
+    """
+    if value is None:
+        return 1
+    try:
+        return 1 if int(value) != 0 else 0
+    except (TypeError, ValueError):
+        return 0
+
+
+def extract_null_mask(query_fn, table, columns, max_rows=5000):
+    """Extrae la mascara de nulos (1=falta / 0=presente) de una muestra de la tabla.
+
+    Args:
+        query_fn: callable lector read-only del backend activo. Recibe un string
+            SQL y devuelve un dict {"status": "ok", "rows": [{col: val, ...}]}
+            (mismo contrato que duckdb_query_readonly / el `_q` de profile_table).
+            No se abre ninguna conexion aqui: toda la lectura pasa por query_fn.
+        table: nombre de la tabla. Se escapa con comillas dobles en la query.
+        columns: lista de nombres de columna a evaluar. Cada una produce una
+            entrada en `mask` con una lista 0/1 paralela por fila. Vacia o None ->
+            status error.
+        max_rows: limite de filas a muestrear (clausula LIMIT). Default 5000.
+
+    Returns:
+        dict (nunca lanza):
+            {
+              "status": "ok" | "error",
+              "error": str,                 # solo si status == "error"
+              "table": str,
+              "columns": [str, ...],        # columnas efectivamente leidas, en orden
+              "mask": {col: [int 0/1, ...], ...},  # alineada por fila, 1=falta, 0=presente
+              "n": int                      # nº de filas muestreadas
+            }
+        Todas las listas de `mask` tienen la misma longitud (= n).
+    """
+    base = {"status": "ok", "table": table, "columns": [], "mask": {}, "n": 0}
+    try:
+        if query_fn is None:
+            return {**base, "status": "error", "error": "query_fn es None"}
+        if not table:
+            return {**base, "status": "error", "error": "table es obligatorio"}
+        if not columns:
+            return {**base, "status": "error", "error": "columns vacío"}
+
+        # Identificadores escapados con comillas dobles (como hace profile_table)
+        # para tolerar nombres con mayusculas/espacios/palabras reservadas. Cada
+        # columna se proyecta como su propio bit IS NULL conservando el alias.
+        select_sql = ", ".join(
+            f'(CASE WHEN "{c}" IS NULL THEN 1 ELSE 0 END) AS "{c}"' for c in columns
+        )
+        sql = f'SELECT {select_sql} FROM "{table}" LIMIT {int(max_rows)}'
+
+        q = query_fn(sql)
+        if not isinstance(q, dict) or q.get("status") != "ok":
+            err = (
+                q.get("error", "query_fn fallo")
+                if isinstance(q, dict)
+                else "query_fn no devolvio un dict"
+            )
+            return {**base, "status": "error", "error": err}
+
+        rows = q.get("rows", []) or []
+        mask = {c: [] for c in columns}
+        for row in rows:
+            for c in columns:
+                # row.get tolera filas que no traigan la columna (None -> falta).
+                mask[c].append(_to_bit(row.get(c) if isinstance(row, dict) else None))
+
+        return {
+            "status": "ok",
+            "table": table,
+            "columns": list(columns),
+            "mask": mask,
+            "n": len(rows),
+        }
+    except Exception as e:  # noqa: BLE001 - dict-no-throw: degradar, nunca lanzar
+        return {**base, "status": "error", "error": str(e)}

python/functions/datascience/extract_null_mask_test.py

@@ -0,0 +1,116 @@
+"""Tests para extract_null_mask.
+
+No usa DuckDB real: inyecta un query_fn FAKE (closure) que devuelve filas
+predefinidas (simulando el SELECT de bits 0/1) y, opcionalmente, captura el SQL
+recibido para verificar la query generada (CASE WHEN ... IS NULL + LIMIT). Asi el
+test es autocontenido y no depende de ningun backend.
+"""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from extract_null_mask import extract_null_mask
+
+
+def _fake_query(rows, captured=None, status="ok", error=None):
+    """Crea un query_fn FAKE.
+
+    `captured` (lista opcional) recibe el SQL ejecutado para poder inspeccionarlo.
+    `status`/`error` permiten simular un fallo del backend.
+    """
+
+    def _q(sql):
+        if captured is not None:
+            captured.append(sql)
+        if status != "ok":
+            return {"status": "error", "error": error or "boom"}
+        return {"status": "ok", "rows": rows}
+
+    return _q
+
+
+def test_golden_mask_alineada():
+    """Golden: mask 0/1 por columna alineada por fila, n correcto, status ok."""
+    # Cada fila simula el SELECT (CASE WHEN col IS NULL THEN 1 ELSE 0 END) AS col.
+    rows = [
+        {"email": 0, "telefono": 1, "edad": 0},
+        {"email": 0, "telefono": 0, "edad": 1},
+        {"email": 1, "telefono": 1, "edad": 0},
+    ]
+    res = extract_null_mask(_fake_query(rows), "clientes", ["email", "telefono", "edad"])
+    assert res["status"] == "ok"
+    assert res["table"] == "clientes"
+    assert res["columns"] == ["email", "telefono", "edad"]
+    assert res["n"] == 3
+    assert res["mask"]["email"] == [0, 0, 1]
+    assert res["mask"]["telefono"] == [1, 0, 1]
+    assert res["mask"]["edad"] == [0, 1, 0]
+    # Todas las listas con la misma longitud.
+    assert all(len(v) == res["n"] for v in res["mask"].values())
+
+
+def test_celda_none_cuenta_como_falta():
+    """Una celda None se cuenta defensivamente como 1 (falta)."""
+    rows = [
+        {"email": 0, "telefono": None},
+        {"email": None, "telefono": 1},
+        {"email": 1, "telefono": 0},
+    ]
+    res = extract_null_mask(_fake_query(rows), "clientes", ["email", "telefono"])
+    assert res["status"] == "ok"
+    assert res["mask"]["email"] == [0, 1, 1]
+    assert res["mask"]["telefono"] == [1, 1, 0]
+    assert res["n"] == 3
+
+
+def test_columns_vacia_status_error():
+    """columns vacia -> status error con columns/mask/n vacios."""
+    res = extract_null_mask(_fake_query([]), "clientes", [])
+    assert res["status"] == "error"
+    assert "columns" in res["error"]
+    assert res["table"] == "clientes"
+    assert res["columns"] == []
+    assert res["mask"] == {}
+    assert res["n"] == 0
+
+
+def test_query_fn_status_error_propaga():
+    """query_fn que devuelve status != ok -> se propaga como error, mask {}."""
+    res = extract_null_mask(
+        _fake_query([], status="error", error="db locked"),
+        "clientes",
+        ["email"],
+    )
+    assert res["status"] == "error"
+    assert "db locked" in res["error"]
+    assert res["mask"] == {}
+    assert res["n"] == 0
+
+
+def test_query_fn_none_da_error_sin_reventar():
+    """query_fn None -> error degradado, sin excepcion."""
+    res = extract_null_mask(None, "clientes", ["email"])
+    assert res["status"] == "error"
+    assert res["columns"] == []
+    assert res["mask"] == {}
+    assert res["n"] == 0
+
+
+def test_sql_contiene_case_y_limit():
+    """La query genera un CASE WHEN IS NULL por columna escapada + LIMIT sobre la tabla."""
+    captured = []
+    rows = [{"email": 0}]
+    extract_null_mask(
+        _fake_query(rows, captured),
+        "clientes_tbl",
+        ["email"],
+        max_rows=123,
+    )
+    assert len(captured) == 1
+    sql = captured[0]
+    assert 'CASE WHEN "email" IS NULL THEN 1 ELSE 0 END' in sql
+    assert 'AS "email"' in sql
+    assert 'FROM "clientes_tbl"' in sql
+    assert "LIMIT 123" in sql

python/functions/datascience/extract_text_sample.md

@@ -0,0 +1,102 @@
+---
+name: extract_text_sample
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def extract_text_sample(db_path: str, table: str, columns: list, backend: str = 'duckdb', sample: int = 2000) -> dict"
+description: "Muestrea columnas de texto de una tabla DuckDB/Postgres con push-down SQL (LIMIT sample), SIN traer la tabla entera a RAM. Funcion impura del grupo de capacidad `eda`: la usan los capitulos de texto/NLP del AutomaticEDA que necesitan valores crudos de texto (longitudes, tokens, ejemplos) sobre una muestra acotada. Construye el lector read-only query_fn(sql)->dict igual que build_eda_render_ctx (closure sobre duckdb_query_readonly / pg_query importados perezosamente desde infra). Escapa los identificadores con comillas dobles y lanza una sola query SELECT \"c1\", \"c2\" FROM \"table\" LIMIT n. Por columna, la lista de strings solo contiene valores NO None y NO vacios: cada celda no nula se convierte con str(...) y se descarta si queda cadena vacia. Estilo dict-no-throw del grupo eda: NUNCA lanza; ante cualquier fallo (query, conversion, backend desconocido) devuelve {status:'error', error:str, columns:{}, n:0}. La clave n reporta el numero de FILAS leidas por la query (antes de filtrar None/vacios)."
+tags: [eda, datascience, text, nlp, extraction, read-only, duckdb, postgres, python]
+uses_functions: [duckdb_query_readonly_py_infra, pg_query_py_infra]
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+params:
+  - name: db_path
+    desc: "ruta al archivo DuckDB, o DSN PostgreSQL si backend='postgres'. Se inyecta en el closure query_fn. No se valida aqui: si la base no existe o el DSN es invalido, la query devuelve status error y el resultado es {status:'error', ...} (no lanza)."
+  - name: table
+    desc: "nombre de la tabla. Se escapa con comillas dobles en la query (SELECT ... FROM \"table\")."
+  - name: columns
+    desc: "lista de nombres de columna de texto a muestrear. Se filtra a las entradas que sean str no vacio; cada nombre se escapa con comillas dobles. Si tras filtrar queda vacia -> {status:'ok', columns:{}, n:0} sin tocar la base."
+  - name: backend
+    desc: "'duckdb' (default) o 'postgres'. Selecciona el lector read-only del registry (duckdb_query_readonly / pg_query). Cualquier otro valor -> {status:'error', error:'backend desconocido: <valor>', columns:{}, n:0}."
+  - name: sample
+    desc: "maximo de filas a muestrear (clausula LIMIT). Default 2000. Acota memoria y tiempo: con tablas grandes obtienes el primer tramo por orden fisico (sin ORDER BY), no un muestreo uniforme."
+output: "dict dict-no-throw (NUNCA lanza): {status:'ok'|'error', columns:{col_name:[str,...]}, n:int, error:str}. En exito (status='ok') columns mapea cada columna pedida a la lista de sus valores de texto NO None y NO vacios (cada celda convertida con str(...)); n es el numero de FILAS leidas por la query (antes de filtrar None/vacios). columns vacio -> {status:'ok', columns:{}, n:0}. En error (backend desconocido, query con status!='ok', o cualquier excepcion) -> {status:'error', error:str, columns:{}, n:0}; la clave error solo aparece en este caso."
+tested: true
+tests: ["test_extract_basic", "test_backend_desconocido", "test_columns_vacio", "test_sample_limit"]
+test_file_path: "python/functions/datascience/extract_text_sample_test.py"
+file_path: "python/functions/datascience/extract_text_sample.py"
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+# Import directo del submodulo (no requiere export en datascience/__init__.py).
+from datascience.extract_text_sample import extract_text_sample
+
+# Muestrea hasta 2000 filas de dos columnas de texto de una tabla DuckDB.
+res = extract_text_sample(
+    "data/reviews.duckdb", "reviews", ["title", "body"],
+    backend="duckdb", sample=2000,
+)
+# res == {
+#   "status": "ok",
+#   "columns": {
+#     "title": ["Gran producto", "No funciona", ...],   # solo no-None, no-""
+#     "body":  ["Lo uso a diario...", ...],
+#   },
+#   "n": 2000,   # filas leidas por la query (antes de filtrar None/vacios)
+# }
+
+# Postgres: db_path es el DSN.
+res_pg = extract_text_sample(
+    "postgresql://user:pass@localhost:5433/trends", "comentarios", ["texto"],
+    backend="postgres", sample=500,
+)
+```
+
+## Cuando usarla
+
+Cuando necesites valores CRUDOS de texto de una o varias columnas para analisis
+NLP/texto (distribucion de longitudes, conteo de tokens, ejemplos representativos,
+deteccion de idioma) pero NO quieras cargar la tabla entera en memoria. Es el
+muestreador de texto del grupo `eda`: una sola llamada con push-down `LIMIT`
+devuelve listas de strings por columna, limpias de None y vacios, listas para
+alimentar un capitulo de texto del AutomaticEDA o cualquier rutina de tokenizado.
+Usala junto a `profile_table` / `build_eda_render_ctx` cuando el perfil agregado
+no basta y hace falta el texto real.
+
+## Gotchas
+
+- **Impura**: lee de la base de datos a traves de `query_fn` (closure sobre
+  `duckdb_query_readonly` / `pg_query`). No abre conexiones fuera de esos wrappers
+  del registry. Estilo dict-no-throw del grupo `eda`: NUNCA lanza; ante cualquier
+  fallo devuelve `{status:'error', error:str, columns:{}, n:0}`.
+- **`error_type` en el frontmatter es `error_go_core` por convencion del registry**
+  (toda funcion impura debe declararlo y el indexer lo exige), pero el codigo NO
+  lanza esa excepcion: degrada al dict de error. Es metadata, no comportamiento.
+- **Backend desconocido**: con un `backend` que no sea `duckdb` ni `postgres`
+  devuelve `{status:'error', error:'backend desconocido: <valor>', columns:{},
+  n:0}` sin tocar la base.
+- **Las listas NO incluyen None ni cadenas vacias**: cada celda no nula se pasa
+  por `str(...)` y se descarta si queda `""`. Por eso `len(columns[col])` puede ser
+  menor que `n` (que cuenta las filas leidas). Si necesitas alineacion por fila
+  (una entrada por fila aunque sea None), usa `build_eda_render_ctx` (raw_numeric),
+  no esta funcion.
+- **`LIMIT sample` sin `ORDER BY`**: con tablas grandes obtienes el primer tramo
+  por orden fisico del backend, no un muestreo uniforme ni reproducible. Sube
+  `sample` para mas cobertura, o pre-ordena/aleatoriza la tabla si necesitas
+  representatividad.
+- **DuckDB en sandbox por defecto**: `duckdb_query_readonly` abre la conexion con
+  `enable_external_access=False`, asi que la query solo puede leer la propia base
+  (no `read_csv`/`httpfs`/`ATTACH` a paths externos). Lee tablas ya existentes en
+  el archivo DuckDB sin problema.
+- **No loguear los datos crudos**: las listas de `columns` pueden contener texto
+  sensible (reviews, comentarios, PII). En trazas usa solo conteos (`n`,
+  `len(columns[col])`) y nombres de columna, no el dict completo.

python/functions/datascience/extract_text_sample.py

@@ -0,0 +1,112 @@
+"""extract_text_sample — muestrea columnas de texto de una tabla sin cargarla en RAM.
+
+Funcion impura (lee de la base de datos) del grupo de capacidad `eda`. Dado un
+``db_path`` + ``table`` (DuckDB o PostgreSQL) y una lista de ``columns`` de texto,
+trae una MUESTRA de esas columnas con push-down SQL (``LIMIT sample``), nunca la
+tabla entera. La usan los capitulos de texto/NLP del AutomaticEDA que necesitan
+valores crudos de texto (longitudes, tokens, ejemplos) sin materializar millones
+de filas en memoria.
+
+El lector read-only ``query_fn(sql) -> dict`` se construye igual que en
+``build_eda_render_ctx`` / ``profile_table``: un closure sobre el wrapper del
+registry (``duckdb_query_readonly`` / ``pg_query``), importado perezosamente
+dentro de la funcion para no crear ciclos al cargar el ``__init__`` del paquete
+``datascience``. Nunca abre conexiones fuera de esos wrappers.
+
+Estilo dict-no-throw del grupo `eda`: la funcion NUNCA lanza. Captura cualquier
+excepcion (query, conversion) y devuelve ``{"status":"error", "error":str(e),
+"columns":{}, "n":0}``. Si la query subyacente devuelve ``status != "ok"``, se
+propaga como error con el mensaje del wrapper.
+
+Por columna, la lista de strings solo contiene valores NO nulos y NO vacios:
+cada celda no-None se convierte con ``str(...)`` y se descarta si queda ``""``.
+La clave ``n`` reporta el numero de FILAS leidas por la query (antes de filtrar
+los None/vacios), util para saber cuanto se muestreo realmente.
+"""
+
+
+def extract_text_sample(db_path, table, columns, backend="duckdb", sample=2000):
+    """Muestrea columnas de texto de una tabla DuckDB/Postgres con push-down SQL.
+
+    Args:
+        db_path: ruta al archivo DuckDB, o DSN PostgreSQL si backend="postgres".
+            Se inyecta en el closure query_fn. No se valida aqui: si la base no
+            existe o el DSN es invalido, la query devuelve status error y el
+            resultado es {status:'error', ...} (no lanza).
+        table: nombre de la tabla. Se escapa con comillas dobles en la query.
+        columns: lista de nombres de columna de texto a muestrear. Se filtra a las
+            entradas que sean str no vacio; cada nombre se escapa con comillas
+            dobles. Si tras filtrar queda vacia -> {status:'ok', columns:{}, n:0}.
+        backend: "duckdb" (default) o "postgres". Selecciona el lector read-only
+            del registry (duckdb_query_readonly / pg_query). Cualquier otro valor
+            -> {status:'error', error:'backend desconocido: ...', columns:{}, n:0}.
+        sample: maximo de filas a muestrear (clausula LIMIT). Default 2000. Acota
+            memoria y tiempo: con tablas grandes obtienes el primer tramo por
+            orden fisico, no un muestreo uniforme.
+
+    Returns:
+        dict (dict-no-throw, NUNCA lanza):
+        {"status": "ok"|"error",
+         "columns": {col_name: [str, str, ...], ...},  # solo no-None, no-""
+         "n": int,        # nº de filas leidas por la query (antes de filtrar)
+         "error": str}    # solo presente si status == "error"
+    """
+    try:
+        # 1) Lector read-only del backend activo, construido como en
+        # build_eda_render_ctx (closure sobre el wrapper del registry). Imports
+        # perezosos: este modulo vive en el paquete `datascience`, importar a
+        # `infra` a nivel de modulo crearia un ciclo al cargar el __init__.
+        if backend == "duckdb":
+            from infra import duckdb_query_readonly
+
+            def query_fn(sql):
+                return duckdb_query_readonly(db_path, sql)
+
+        elif backend == "postgres":
+            from infra import pg_query
+
+            def query_fn(sql):
+                return pg_query(db_path, sql)
+
+        else:
+            return {
+                "status": "error",
+                "error": f"backend desconocido: {backend}",
+                "columns": {},
+                "n": 0,
+            }
+
+        # 2) Columnas validas (str no vacio). Si no queda ninguna, nada que
+        # muestrear: ok con columns vacio.
+        cols = []
+        if isinstance(columns, (list, tuple)):
+            cols = [c for c in columns if isinstance(c, str) and c != ""]
+        if not cols:
+            return {"status": "ok", "columns": {}, "n": 0}
+
+        # 3) Push-down: una sola query con LIMIT. Identificadores escapados con
+        # comillas dobles, igual que build_eda_render_ctx.
+        cols_sql = ", ".join(f'"{c}"' for c in cols)
+        sql = f'SELECT {cols_sql} FROM "{table}" LIMIT {int(sample)}'
+        q = query_fn(sql)
+        if not isinstance(q, dict) or q.get("status") != "ok":
+            err = q.get("error") if isinstance(q, dict) else "query sin resultado"
+            return {"status": "error", "error": str(err), "columns": {}, "n": 0}
+
+        rows = q.get("rows") or []
+        out = {c: [] for c in cols}
+        for row in rows:
+            if not isinstance(row, dict):
+                continue
+            for c in cols:
+                value = row.get(c)
+                if value is None:
+                    continue
+                s = str(value)
+                if s == "":
+                    continue
+                out[c].append(s)
+
+        return {"status": "ok", "columns": out, "n": len(rows)}
+    except Exception as exc:  # noqa: BLE001 - dict-no-throw del grupo eda
+        return {"status": "error", "error": str(exc), "columns": {}, "n": 0}

python/functions/datascience/extract_text_sample_test.py

@@ -0,0 +1,83 @@
+"""Tests para extract_text_sample.
+
+Self-contained: crea un DuckDB temporal pequeño con una columna de texto (algunas
+filas con NULL) y una numerica, y verifica que la muestra de texto trae solo los
+valores no nulos, que el backend desconocido y la lista de columnas vacia se
+manejan dict-no-throw, y que sample acota el numero de filas leidas.
+"""
+
+import os
+import sys
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_FUNCTIONS = os.path.abspath(os.path.join(_HERE, ".."))  # python/functions
+if _FUNCTIONS not in sys.path:
+    sys.path.insert(0, _FUNCTIONS)
+
+import duckdb  # noqa: E402
+
+from datascience.extract_text_sample import extract_text_sample  # noqa: E402
+
+_TABLE = "t"
+# 6 filas: txt VARCHAR con dos NULL, other INT siempre presente.
+_ROWS = [
+    ("alpha", 1),
+    ("beta", 2),
+    (None, 3),
+    ("gamma", 4),
+    (None, 5),
+    ("delta", 6),
+]
+_TXT_NON_NULL = {"alpha", "beta", "gamma", "delta"}
+
+
+def _make_db(tmp_path):
+    """Crea un DuckDB temporal con la tabla de prueba y devuelve su ruta."""
+    db_path = os.path.join(str(tmp_path), "text_sample.duckdb")
+    con = duckdb.connect(db_path)
+    try:
+        con.execute(f'CREATE TABLE "{_TABLE}" (txt VARCHAR, other INTEGER)')
+        con.executemany(f'INSERT INTO "{_TABLE}" VALUES (?, ?)', _ROWS)
+    finally:
+        con.close()
+    return db_path
+
+
+def test_extract_basic(tmp_path):
+    db_path = _make_db(tmp_path)
+    res = extract_text_sample(db_path, _TABLE, ["txt"])
+    assert res["status"] == "ok"
+    # n = filas leidas por la query (6), antes de filtrar None.
+    assert res["n"] == len(_ROWS)
+    # columns["txt"] trae solo los strings no nulos (los dos NULL fuera).
+    assert "txt" in res["columns"]
+    assert set(res["columns"]["txt"]) == _TXT_NON_NULL
+    assert len(res["columns"]["txt"]) == len(_TXT_NON_NULL)
+    # No se pidio "other", no debe aparecer.
+    assert "other" not in res["columns"]
+
+
+def test_backend_desconocido(tmp_path):
+    db_path = _make_db(tmp_path)
+    res = extract_text_sample(db_path, _TABLE, ["txt"], backend="mysql")
+    assert res["status"] == "error"
+    assert "backend desconocido" in res["error"]
+    assert res["columns"] == {}
+    assert res["n"] == 0
+
+
+def test_columns_vacio(tmp_path):
+    db_path = _make_db(tmp_path)
+    res = extract_text_sample(db_path, _TABLE, [])
+    assert res["status"] == "ok"
+    assert res["columns"] == {}
+    assert res["n"] == 0
+
+
+def test_sample_limit(tmp_path):
+    db_path = _make_db(tmp_path)
+    res = extract_text_sample(db_path, _TABLE, ["txt"], sample=2)
+    assert res["status"] == "ok"
+    # sample=2 -> la query lee como mucho 2 filas.
+    assert res["n"] == 2
+    assert len(res["columns"]["txt"]) <= 2

python/functions/datascience/fdr_correction.md

@@ -3,19 +3,19 @@ name: fdr_correction
 kind: function
 lang: py
 domain: datascience
-version: "1.0.0"
+version: "1.1.0"
 purity: pure
 signature: "def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = \"bh\") -> dict"
-description: "Correccion de comparaciones multiples (multiple-testing) sobre una lista de p-valores: Benjamini-Hochberg (FDR, 'bh') o Bonferroni (FWER, 'bonferroni'). Antidoto al sesgo de mineria de datos (data-mining bias): al evaluar muchas hipotesis a la vez (todos los pares de una matriz), el azar produce falsos positivos; esta funcion ajusta los p-valores y marca cuales siguen siendo significativos tras corregir. Pura, sin dependencias externas, alineada 1:1 con la entrada (admite None en posiciones sin test)."
+description: "Correccion de comparaciones multiples (multiple-testing) sobre una lista de p-valores: Benjamini-Hochberg (FDR, 'bh'), Bonferroni (FWER, 'bonferroni') o Holm-Bonferroni (FWER step-down, 'holm', mas potente que Bonferroni simple). Antidoto al sesgo de mineria de datos (data-mining bias): al evaluar muchas hipotesis a la vez (todos los pares de una matriz), el azar produce falsos positivos; esta funcion ajusta los p-valores y marca cuales siguen siendo significativos tras corregir. Pura, sin dependencias externas, alineada 1:1 con la entrada (admite None en posiciones sin test)."
-tags: [eda, statistics, multiple-testing, fdr, benjamini-hochberg, bonferroni, p-value, data-mining-bias, python]
+tags: [eda, statistics, multiple-testing, fdr, benjamini-hochberg, bonferroni, holm, holm-bonferroni, fwer, p-value, data-mining-bias, python]
 params:
   - name: pvalues
     desc: "lista de p-valores (floats en [0, 1]). Se admiten None u otros valores no validos en posiciones sin test disponible; se propagan como None en la salida y no cuentan como prueba (m)."
   - name: alpha
     desc: "nivel de significancia objetivo tras la correccion (default 0.05). Para BH es el umbral del FDR; para Bonferroni, del FWER (tasa de error por familia)."
   - name: method
-    desc: "'bh' = Benjamini-Hochberg (controla FDR, menos conservador, mas potencia); 'bonferroni' = controla FWER (mas conservador). Cualquier otro valor devuelve un dict con note."
+    desc: "'bh' = Benjamini-Hochberg (controla FDR, menos conservador, mas potencia); 'bonferroni' = controla FWER (mas conservador); 'holm' = Holm-Bonferroni (controla FWER, step-down, uniformemente mas potente que Bonferroni simple). Cualquier otro valor devuelve un dict con note."
-output: "dict {p_values_adjusted: lista alineada con pvalues (float ajustado o None), reject: lista de bool (True = significativo tras corregir), n_tests: nº de p-valores validos (m), n_rejected: nº de hipotesis rechazadas, alpha: float aplicado, method: str}. Casos degenerados (vacio, sin p validos, metodo desconocido) anaden clave note. Nunca None ni excepcion."
+output: "dict {p_values_adjusted: lista alineada con pvalues (float ajustado o None), reject: lista de bool (True = significativo tras corregir), n_tests: nº de p-valores validos (m), n_rejected: nº de hipotesis rechazadas, alpha: float aplicado, method: str ('bh' | 'bonferroni' | 'holm')}. Casos degenerados (vacio, sin p validos, metodo desconocido) anaden clave note. Nunca None ni excepcion."
 uses_functions: []
 uses_types: []
 returns: []
@@ -23,7 +23,7 @@ returns_optional: false
 error_type: ""
 imports: [math]
 tested: true
-tests: ["test_bh_golden_rechaza_dos_de_tres", "test_bonferroni_mas_conservador_que_bh", "test_p_values_adjusted_alineados_y_en_rango", "test_none_se_propaga_alineado", "test_lista_vacia_devuelve_note", "test_solo_none_devuelve_note", "test_metodo_desconocido_devuelve_note", "test_todos_significativos"]
+tests: ["test_bh_golden_rechaza_dos_de_tres", "test_bonferroni_mas_conservador_que_bh", "test_p_values_adjusted_alineados_y_en_rango", "test_none_se_propaga_alineado", "test_lista_vacia_devuelve_note", "test_solo_none_devuelve_note", "test_metodo_desconocido_devuelve_note", "test_todos_significativos", "test_holm_golden_rechaza_dos_de_cuatro", "test_holm_entre_bonferroni_y_bh", "test_none_se_propaga_alineado_holm", "test_lista_vacia_holm_devuelve_note"]
 test_file_path: "python/functions/datascience/fdr_correction_test.py"
 file_path: "python/functions/datascience/fdr_correction.py"
 ---
@@ -45,6 +45,13 @@ bon = fdr_correction(pvalues, alpha=0.05, method="bonferroni")
 print(bon["reject"])       # -> [True, False, False]
 print(bon["p_values_adjusted"])  # -> [0.03, 0.06, 1.0]
 
+# Holm-Bonferroni (step-down): controla el FWER como Bonferroni pero es mas
+# potente; rechaza al menos tanto como Bonferroni simple, nunca menos.
+holm = fdr_correction([0.01, 0.04, 0.03, 0.005], alpha=0.05, method="holm")
+print(holm["reject"])      # -> [True, False, False, True]
+print(holm["p_values_adjusted"])  # -> [0.03, 0.06, 0.06, 0.02]
+print(holm["n_rejected"])  # -> 2
+
 # Posiciones sin test (None) se propagan alineadas: el llamador puede pasar la
 # lista completa de pares y recuperar el mapeo 1:1.
 mix = fdr_correction([0.001, None, 0.9])
@@ -61,8 +68,11 @@ combinaciones y se quede con las que "pasan". Sin corregir, con N pruebas y
 alpha=0.05 esperas ~5% de falsos positivos *por azar*: cuantas mas pruebas, mas
 correlaciones espurias. Llama a `fdr_correction` con todos los p-valores de la
 familia y usa `reject` (no el umbral crudo) para decidir que es real. Usa `"bh"`
-por defecto (mejor potencia); `"bonferroni"` cuando un falso positivo sea muy
+por defecto (mejor potencia); `"holm"` (Holm-Bonferroni, FWER step-down) cuando
-costoso y prefieras maxima cautela.
+quieras controlar el FWER pero sin la perdida de potencia de Bonferroni simple
+(rechaza al menos tanto como `"bonferroni"`, nunca menos); `"bonferroni"` cuando
+un falso positivo sea muy costoso y prefieras la maxima cautela del metodo mas
+simple.
 
 ## Gotchas
 
@@ -76,8 +86,16 @@ costoso y prefieras maxima cautela.
   eso puedes pasar la lista completa de pares aunque algunos no tengan test.
 - `n_tests` es el numero de p-valores **validos** (m), que puede ser menor que
   `len(pvalues)` si hay `None`.
-- BH y Bonferroni controlan cosas distintas: BH la tasa de falsos
+- BH controla cosa distinta que Bonferroni/Holm: BH la tasa de falsos
-  descubrimientos (FDR), Bonferroni la probabilidad de *cualquier* falso
+  descubrimientos (FDR); Bonferroni y Holm la probabilidad de *cualquier* falso
   positivo (FWER). No son intercambiables; elige segun el coste de equivocarte.
+- `"holm"` y `"bonferroni"` controlan ambos el FWER, pero Holm es step-down y
+  uniformemente mas potente: rechaza al menos tantas hipotesis como Bonferroni
+  simple sobre el mismo set, nunca menos. Si controlas FWER, `"holm"` domina a
+  `"bonferroni"` salvo que necesites el ajuste mas simple por interpretabilidad.
 - Metodo desconocido o lista vacia/sin p validos no lanzan: devuelven un dict
-  con `note`.
+  con `note`. Los metodos validos son `"bh"`, `"bonferroni"` y `"holm"`.
+
+## Capability growth log
+
+- v1.1.0 (2026-06-30) — añade method="holm" (Holm-Bonferroni step-down, FWER, más potente que Bonferroni simple).

python/functions/datascience/fdr_correction.py

@@ -5,12 +5,15 @@ todos los pares de una matriz de asociacion), la probabilidad de obtener al meno
 un falso positivo por azar crece con el numero de pruebas: es el sesgo de mineria
 de datos (data-mining bias) descrito por Aronson en *Evidence-Based Technical
 Analysis* (cap. 6). Esta funcion ajusta los p-valores para controlar ese sesgo
-mediante dos metodos clasicos:
+mediante tres metodos clasicos:
 
 - Benjamini-Hochberg (``"bh"``): controla la tasa de falsos descubrimientos
   (False Discovery Rate, FDR). Menos conservador, mas potencia estadistica.
 - Bonferroni (``"bonferroni"``): controla la tasa de error por familia
   (Family-Wise Error Rate, FWER). Mas conservador.
+- Holm-Bonferroni (``"holm"``): controla el FWER como Bonferroni pero es un
+  procedimiento step-down uniformemente mas potente; rechaza al menos tantas
+  hipotesis como Bonferroni simple, nunca menos.
 
 No usa dependencias externas: aritmetica de la libreria estandar.
 """
@@ -35,8 +38,9 @@ def _is_valid_p(v) -> bool:
 def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> dict:
     """Corrige una lista de p-valores por comparaciones multiples.
 
-    Aplica Benjamini-Hochberg (FDR) o Bonferroni (FWER) sobre ``pvalues`` y
+    Aplica Benjamini-Hochberg (FDR), Bonferroni (FWER) o Holm-Bonferroni
-    devuelve, alineado posicion a posicion con la entrada, el p-valor ajustado y
+    (FWER, step-down) sobre ``pvalues`` y devuelve, alineado posicion a
+    posicion con la entrada, el p-valor ajustado y
     si cada hipotesis se rechaza al nivel ``alpha`` tras la correccion. Las
     posiciones cuyo valor no sea un p-valor valido (``None``, ``NaN``, fuera de
     ``[0, 1]`` o no numerico) se conservan en la salida como ``None`` /
@@ -53,8 +57,10 @@ def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> di
             otros valores no validos en posiciones sin test disponible; se
             propagan como ``None`` en la salida y no cuentan como prueba.
         alpha: nivel de significancia objetivo tras la correccion (default 0.05).
-            Para BH es el umbral del FDR; para Bonferroni, del FWER.
+            Para BH es el umbral del FDR; para Bonferroni y Holm, del FWER.
-        method: ``"bh"`` (Benjamini-Hochberg, FDR) o ``"bonferroni"`` (FWER).
+        method: ``"bh"`` (Benjamini-Hochberg, FDR), ``"bonferroni"`` (FWER) o
+            ``"holm"`` (Holm-Bonferroni, FWER step-down, mas potente que
+            Bonferroni simple).
 
     Returns:
         dict con las claves:
@@ -68,7 +74,7 @@ def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> di
             n_tests: numero de p-valores validos usados en la correccion (m).
             n_rejected: numero de hipotesis rechazadas (significativas).
             alpha: nivel de significancia aplicado (float).
-            method: metodo aplicado (``"bh"`` o ``"bonferroni"``).
+            method: metodo aplicado (``"bh"``, ``"bonferroni"`` o ``"holm"``).
 
         Casos degenerados (lista vacia, sin p-valores validos o metodo
         desconocido) anaden ademas una clave ``note`` y devuelven listas
@@ -76,7 +82,7 @@ def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> di
         en las posiciones invalidas).
     """
     method_norm = (method or "").strip().lower()
-    if method_norm not in {"bh", "bonferroni"}:
+    if method_norm not in {"bh", "bonferroni", "holm"}:
         n = len(pvalues)
         return {
             "p_values_adjusted": [None] * n,
@@ -86,8 +92,8 @@ def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> di
             "alpha": float(alpha),
             "method": method,
             "note": (
-                f"metodo desconocido '{method}'; usa 'bh' (Benjamini-Hochberg) "
+                f"metodo desconocido '{method}'; usa 'bh' (Benjamini-Hochberg), "
-                "o 'bonferroni'"
+                "'bonferroni' o 'holm' (Holm-Bonferroni)"
             ),
         }
 
@@ -129,6 +135,20 @@ def fdr_correction(pvalues: list, alpha: float = 0.05, method: str = "bh") -> di
             padj = min(1.0, p * m)
             adjusted[orig_idx] = padj
             reject[orig_idx] = padj <= a
+    elif method_norm == "holm":
+        # Holm-Bonferroni (step-down). Ordena p ascendente; para el rank k
+        # (1-indexed) el p ajustado crudo es (m - k + 1) * p_(k). Impon
+        # monotonicidad acumulada (no decreciente) recorriendo de menor a mayor:
+        # padj_(k) = max(padj_(k-1), min(1, (m-k+1)*p_(k))), con padj_(0)=0.
+        order = sorted(valid, key=lambda t: t[1])  # [(orig_idx, p), ...] por p asc
+        prev = 0.0
+        for k in range(1, m + 1):
+            orig_idx, p = order[k - 1]
+            raw = min(1.0, (m - k + 1) * p)
+            padj = max(prev, raw)
+            prev = padj
+            adjusted[orig_idx] = padj
+            reject[orig_idx] = padj <= a
     else:
         # Benjamini-Hochberg (step-up). Ordena p ascendente y calcula q-valores
         # con la monotonicidad acumulada de derecha a izquierda.

python/functions/datascience/fdr_correction_test.py

@@ -82,7 +82,8 @@ def test_solo_none_devuelve_note():
 
 
 def test_metodo_desconocido_devuelve_note():
-    out = fdr_correction([0.01, 0.02], method="holm")
+    # 'holm' ya es un metodo valido (v1.1.0); usamos uno realmente desconocido.
+    out = fdr_correction([0.01, 0.02], method="sidak")
     assert "note" in out
     assert out["n_rejected"] == 0
     assert out["reject"] == [False, False]
@@ -97,3 +98,66 @@ def test_todos_significativos():
     assert bon["n_rejected"] == 3
     assert all(bh["reject"])
     assert all(bon["reject"])
+
+
+def test_holm_golden_rechaza_dos_de_cuatro():
+    # Holm-Bonferroni (step-down) sobre [0.01, 0.04, 0.03, 0.005], m=4, alpha=0.05.
+    # Ordenado ascendente: 0.005, 0.01, 0.03, 0.04.
+    #   padj_(1) = 4*0.005 = 0.02
+    #   padj_(2) = max(0.02, 3*0.01=0.03) = 0.03
+    #   padj_(3) = max(0.03, 2*0.03=0.06) = 0.06
+    #   padj_(4) = max(0.06, 1*0.04=0.04) = 0.06
+    # Mapeado al orden de entrada [0.01, 0.04, 0.03, 0.005]:
+    #   0.01 -> 0.03, 0.04 -> 0.06, 0.03 -> 0.06, 0.005 -> 0.02
+    out = fdr_correction([0.01, 0.04, 0.03, 0.005], alpha=0.05, method="holm")
+    assert out["method"] == "holm"
+    assert out["n_tests"] == 4
+    adj = out["p_values_adjusted"]
+    assert abs(adj[0] - 0.03) < 1e-9
+    assert abs(adj[1] - 0.06) < 1e-9
+    assert abs(adj[2] - 0.06) < 1e-9
+    assert abs(adj[3] - 0.02) < 1e-9
+    assert out["reject"] == [True, False, False, True]
+    assert out["n_rejected"] == 2
+
+
+def test_holm_entre_bonferroni_y_bh():
+    # Holm controla FWER como Bonferroni pero es step-down: rechaza AL MENOS
+    # tanto como Bonferroni simple, y a lo sumo tanto como BH (FDR, menos
+    # conservador). Cadena de potencia: bonferroni <= holm <= bh.
+    pvalues = [0.01, 0.02, 0.04, 0.005]
+    bon = fdr_correction(pvalues, alpha=0.05, method="bonferroni")
+    holm = fdr_correction(pvalues, alpha=0.05, method="holm")
+    bh = fdr_correction(pvalues, alpha=0.05, method="bh")
+    assert holm["n_rejected"] >= bon["n_rejected"]
+    assert holm["n_rejected"] <= bh["n_rejected"]
+    # En este set Holm gana potencia frente a Bonferroni simple (estricto).
+    assert holm["n_rejected"] > bon["n_rejected"]
+
+    # Un set donde Holm es estrictamente mas conservador que BH.
+    pvals2 = [0.01, 0.02, 0.03, 0.04]
+    bon2 = fdr_correction(pvals2, alpha=0.05, method="bonferroni")
+    holm2 = fdr_correction(pvals2, alpha=0.05, method="holm")
+    bh2 = fdr_correction(pvals2, alpha=0.05, method="bh")
+    assert holm2["n_rejected"] >= bon2["n_rejected"]
+    assert holm2["n_rejected"] < bh2["n_rejected"]
+
+
+def test_none_se_propaga_alineado_holm():
+    # None se propaga alineado tambien con holm: la posicion central no cuenta
+    # como prueba (m=2) y se devuelve como None / False.
+    out = fdr_correction([0.001, None, 0.9], method="holm")
+    assert out["n_tests"] == 2
+    assert out["p_values_adjusted"][1] is None
+    assert out["reject"][1] is False
+    assert out["reject"][0] is True
+    assert len(out["reject"]) == 3
+
+
+def test_lista_vacia_holm_devuelve_note():
+    out = fdr_correction([], method="holm")
+    assert out["p_values_adjusted"] == []
+    assert out["reject"] == []
+    assert out["n_tests"] == 0
+    assert out["n_rejected"] == 0
+    assert "note" in out

python/functions/datascience/generate_synthetic_eda_folder.md

@@ -0,0 +1,77 @@
+---
+name: generate_synthetic_eda_folder
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def generate_synthetic_eda_folder(out_dir: str, n_rows: int = 2000, seed: int = 42) -> dict"
+description: "Genera una carpeta con 3 CSV RELACIONADOS (customers, orders, reviews) deterministas por seed (Faker + numpy) para ejercitar el motor AutomaticEDA multi-tabla / profile_database. orders.customer_id y reviews.customer_id estan contenidos al 100% en customers.customer_id (PK uuid), de modo que la deteccion FK por containment (min_inclusion=0.9) descubre ambas relaciones. customers es la tabla padre; reutiliza helpers de generate_synthetic_eda_table (texto multi-idioma, lat/lon validas, amount con outliers). Estilo dict-no-throw: nunca lanza."
+tags: [eda, synthetic, faker, testing, fixture, datascience]
+params:
+  - name: out_dir
+    desc: "Carpeta de salida. Se crea con mkdir -p si no existe. Recibe customers.csv, orders.csv y reviews.csv."
+  - name: n_rows
+    desc: "Numero de clientes (filas de customers). orders ~= 2*n_rows filas, reviews ~= n_rows filas. Default 2000."
+  - name: seed
+    desc: "Semilla para Faker (Faker.seed) y numpy (np.random.default_rng). Mismo seed -> CSVs identicos byte a byte. Default 42."
+output: "dict dict-no-throw. En exito {status:'ok', out_dir, files:{customers,orders,reviews}, n_customers, n_orders, n_reviews, expected_relations:[{from_table,from_col,to_table,to_col}, ...], seed}. En error (sin lanzar, p.ej. n_rows<=0) {status:'error', error:str}. expected_relations declara las 2 FK orders->customers y reviews->customers (ambas por customer_id)."
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+tested: true
+tests: ["test_genera_ok_y_archivos", "test_determinismo_mismo_seed", "test_seeds_distintos_difieren", "test_fk_containment", "test_review_text_mediana_palabras", "test_n_rows_invalido"]
+test_file_path: "python/functions/datascience/generate_synthetic_eda_folder_test.py"
+file_path: "python/functions/datascience/generate_synthetic_eda_folder.py"
+---
+
+## Ejemplo
+
+```bash
+# Genera /tmp/eda_folder/{customers,orders,reviews}.csv (300 customers, seed 42)
+fn run generate_synthetic_eda_folder /tmp/eda_folder 300 42
+```
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience import generate_synthetic_eda_folder
+
+res = generate_synthetic_eda_folder("/tmp/eda_folder", n_rows=300, seed=42)
+# res["files"] -> {"customers": ".../customers.csv", "orders": ..., "reviews": ...}
+# res["expected_relations"] -> orders.customer_id y reviews.customer_id -> customers.customer_id
+# Luego perfila la carpeta/base con el grupo eda:
+#   fn run profile_database /tmp/eda_folder
+```
+
+## Cuando usarla
+
+- Cuando necesites un fixture REPRODUCIBLE multi-tabla para evaluar el EDA de carpeta/base (`profile_database`, join graph, capitulo de relaciones inter-tabla) con relaciones FK reales y detectables.
+- Cuando escribas tests de la deteccion de claves foraneas por containment: orders y reviews referencian customer_id contenido al 100% en customers (inclusion 1.0 >= min_inclusion 0.9).
+- Como contraparte multi-tabla de `generate_synthetic_eda_table` (que cubre el EDA de UNA tabla).
+
+## Gotchas
+
+- **Impura**: escribe 3 CSV a disco (`mkdir -p` de la carpeta). Sobrescribe los CSV existentes con el mismo nombre.
+- **Requiere `faker`, `numpy` y `pandas`** en el venv. Sin `faker` devuelve `{status:'error'}` (no lanza).
+- **El containment depende del orden**: customers se genera PRIMERO y orders/reviews muestrean sus `customer_id`. Si se invierte el orden, la FK deja de estar contenida y el detector no la encuentra.
+- **`signup_date`/`ts` se escriben como texto ISO en el CSV** (`YYYY-MM-DD` / `YYYY-MM-DD HH:MM:SS`): es CSV, todo es texto; el profiler los promociona a datetime al leerlos.
+- **Determinismo dependiente del orden de llamadas**: se siembra `Faker.seed(seed)` + `np.random.default_rng(seed)` al inicio; mismo seed -> CSVs identicos byte a byte.
+- **Reutiliza helpers privados** de `generate_synthetic_eda_table` (`_make_fakers`, `_make_latlon`, `_make_reviews`, `_amount_with_outliers`): no romper esas firmas sin actualizar esta funcion.
+
+## Notas
+
+Estructura generada:
+
+| Archivo | PK | FK | Columnas clave |
+|---|---|---|---|
+| customers.csv | customer_id (uuid) | — | name, country, signup_date, latitude, longitude, email |
+| orders.csv | order_id (uuid) | customer_id -> customers | amount (lognormal + outliers), category, ts |
+| reviews.csv | review_id (uuid) | customer_id -> customers | review_text (multi-idioma, mediana palabras>=20), rating (1..5) |
+
+orders tiene ~2x filas que customers y reviews ~1x. Todos los `customer_id` de orders
+y reviews estan contenidos en customers (containment ⊆), por lo que la deteccion FK por
+inclusion descubre las dos relaciones declaradas en `expected_relations`.

python/functions/datascience/generate_synthetic_eda_folder.py

@@ -0,0 +1,177 @@
+"""generate_synthetic_eda_folder — fixture multi-tabla relacionado para el EDA de base/carpeta.
+
+Funcion impura (escribe CSVs a disco) y determinista por ``seed``: crea una
+carpeta con 3 CSV RELACIONADOS (customers, orders, reviews) cuyo contenido esta
+disenado para que el motor AutomaticEDA multi-tabla / `profile_database` detecte
+las relaciones FK por containment de valores (orders.customer_id y
+reviews.customer_id contenidos al 100% en customers.customer_id, por encima del
+``min_inclusion=0.9`` que usa la deteccion).
+
+Reutiliza los helpers de ``generate_synthetic_eda_table`` (texto multi-idioma,
+lat/lon validas, amount con outliers, listas fijas de paises/categorias) para no
+reimplementar logica.
+
+Estilo dict-no-throw del grupo `eda`: NUNCA lanza; devuelve
+``{"status": "error", "error": str}`` ante cualquier fallo.
+"""
+
+import os
+
+from .generate_synthetic_eda_table import (
+    _CATEGORIES,
+    _COUNTRIES,
+    _amount_with_outliers,
+    _make_fakers,
+    _make_latlon,
+    _make_reviews,
+)
+
+
+def generate_synthetic_eda_folder(out_dir, n_rows=2000, seed=42):
+    """Genera una carpeta con 3 CSV relacionados (customers/orders/reviews).
+
+    customers es la tabla padre (PK ``customer_id`` uuid unica). orders y reviews
+    referencian ``customer_id`` muestreandolo de customers, de modo que TODOS sus
+    valores estan contenidos en customers (inclusion 1.0 -> FK detectable).
+
+    Funcion impura (escribe a disco) y determinista por ``seed``. NUNCA lanza.
+
+    Args:
+        out_dir: carpeta de salida. Se crea con ``mkdir -p`` si no existe.
+        n_rows: numero de clientes (customers). orders ~= 2*n_rows, reviews ~= n_rows.
+            Default 2000.
+        seed: semilla para Faker y numpy. Default 42.
+
+    Returns:
+        dict dict-no-throw. En exito::
+
+            {"status": "ok", "out_dir": ..., "files": {customers, orders, reviews},
+             "n_customers": ..., "n_orders": ..., "n_reviews": ...,
+             "expected_relations": [{from_table, from_col, to_table, to_col}, ...],
+             "seed": seed}
+
+        En error (sin lanzar)::
+
+            {"status": "error", "error": str}
+    """
+    try:
+        import numpy as np
+        import pandas as pd
+
+        n = int(n_rows)
+        if n <= 0:
+            return {"status": "error", "error": f"n_rows debe ser > 0, dado {n_rows!r}"}
+
+        os.makedirs(out_dir, exist_ok=True)
+
+        fakers = _make_fakers(seed)
+        rng = np.random.default_rng(seed)
+
+        # ---------------- customers (tabla padre) ----------------
+        n_cust = n
+        customer_ids = [fakers["en_US"].uuid4() for _ in range(n_cust)]
+        names = [fakers["en_US"].name() for _ in range(n_cust)]
+        cust_country = rng.choice(_COUNTRIES, n_cust)
+        base = np.datetime64("2022-01-01")
+        signup_offsets = rng.integers(0, 730, n_cust)
+        signup_date = pd.to_datetime(base) + pd.to_timedelta(signup_offsets, unit="D")
+        signup_iso = [d.strftime("%Y-%m-%d") for d in signup_date]
+        lat, lon = _make_latlon(cust_country, rng)
+        cust_email = [fakers["en_US"].email() for _ in range(n_cust)]
+
+        customers = pd.DataFrame(
+            {
+                "customer_id": customer_ids,
+                "name": names,
+                "country": cust_country,
+                "signup_date": signup_iso,
+                "latitude": lat,
+                "longitude": lon,
+                "email": cust_email,
+            }
+        )
+
+        # ---------------- orders (FK -> customers) ----------------
+        n_orders = n_cust * 2
+        order_ids = [fakers["en_US"].uuid4() for _ in range(n_orders)]
+        order_cust = rng.choice(customer_ids, n_orders)  # subset/multiset de customers
+        amount = _amount_with_outliers(n_orders, rng, n_extreme=10)
+        order_cat = rng.choice(_CATEGORIES, n_orders)
+        ts_offsets = rng.integers(0, 730 * 24 * 3600, n_orders)
+        ts = pd.to_datetime(np.datetime64("2022-01-01T00:00:00")) + pd.to_timedelta(
+            ts_offsets, unit="s"
+        )
+        ts_iso = [t.strftime("%Y-%m-%d %H:%M:%S") for t in ts]
+
+        orders = pd.DataFrame(
+            {
+                "order_id": order_ids,
+                "customer_id": order_cust,
+                "amount": amount,
+                "category": order_cat,
+                "ts": ts_iso,
+            }
+        )
+
+        # ---------------- reviews (FK -> customers) ----------------
+        n_reviews = n_cust
+        review_ids = [fakers["en_US"].uuid4() for _ in range(n_reviews)]
+        # Subconjunto de customers (no todos) -> containment estricto ⊆ customers.
+        rev_cust = rng.choice(customer_ids, n_reviews)
+        review_text = _make_reviews(n_reviews, rng, fakers, null_frac=0.0)
+        rating = rng.integers(1, 6, n_reviews)
+
+        reviews = pd.DataFrame(
+            {
+                "review_id": review_ids,
+                "customer_id": rev_cust,
+                "review_text": review_text,
+                "rating": rating,
+            }
+        )
+
+        files = {
+            "customers": os.path.join(out_dir, "customers.csv"),
+            "orders": os.path.join(out_dir, "orders.csv"),
+            "reviews": os.path.join(out_dir, "reviews.csv"),
+        }
+        customers.to_csv(files["customers"], index=False)
+        orders.to_csv(files["orders"], index=False)
+        reviews.to_csv(files["reviews"], index=False)
+
+        return {
+            "status": "ok",
+            "out_dir": out_dir,
+            "files": files,
+            "n_customers": n_cust,
+            "n_orders": n_orders,
+            "n_reviews": n_reviews,
+            "expected_relations": [
+                {
+                    "from_table": "orders",
+                    "from_col": "customer_id",
+                    "to_table": "customers",
+                    "to_col": "customer_id",
+                },
+                {
+                    "from_table": "reviews",
+                    "from_col": "customer_id",
+                    "to_table": "customers",
+                    "to_col": "customer_id",
+                },
+            ],
+            "seed": seed,
+        }
+    except Exception as exc:  # noqa: BLE001 — dict-no-throw del grupo eda.
+        return {"status": "error", "error": str(exc)}
+
+
+if __name__ == "__main__":
+    import json
+    import sys
+
+    args = sys.argv[1:]
+    out = args[0] if len(args) > 0 else "/tmp/synthetic_eda_folder"
+    rows = int(args[1]) if len(args) > 1 else 2000
+    sd = int(args[2]) if len(args) > 2 else 42
+    print(json.dumps(generate_synthetic_eda_folder(out, rows, sd), indent=2))

python/functions/datascience/generate_synthetic_eda_folder_test.py

@@ -0,0 +1,74 @@
+"""Tests para generate_synthetic_eda_folder."""
+
+import os
+import statistics
+
+import pandas as pd
+
+from datascience.generate_synthetic_eda_folder import generate_synthetic_eda_folder
+
+
+def test_genera_ok_y_archivos(tmp_path):
+    out = str(tmp_path / "folder")
+    res = generate_synthetic_eda_folder(out, n_rows=300, seed=42)
+    assert res["status"] == "ok"
+    assert res["n_customers"] == 300
+    assert res["n_orders"] == 600
+    assert res["n_reviews"] == 300
+    for key in ("customers", "orders", "reviews"):
+        assert os.path.exists(res["files"][key])
+    # Relaciones esperadas declaradas.
+    rels = {(r["from_table"], r["to_table"]) for r in res["expected_relations"]}
+    assert ("orders", "customers") in rels
+    assert ("reviews", "customers") in rels
+
+
+def test_determinismo_mismo_seed(tmp_path):
+    out1 = str(tmp_path / "f1")
+    out2 = str(tmp_path / "f2")
+    generate_synthetic_eda_folder(out1, n_rows=250, seed=11)
+    generate_synthetic_eda_folder(out2, n_rows=250, seed=11)
+    for name in ("customers.csv", "orders.csv", "reviews.csv"):
+        a = open(os.path.join(out1, name), "rb").read()
+        b = open(os.path.join(out2, name), "rb").read()
+        assert a == b, f"{name} difiere entre dos generaciones con el mismo seed"
+
+
+def test_seeds_distintos_difieren(tmp_path):
+    out1 = str(tmp_path / "f1")
+    out2 = str(tmp_path / "f2")
+    generate_synthetic_eda_folder(out1, n_rows=250, seed=11)
+    generate_synthetic_eda_folder(out2, n_rows=250, seed=12)
+    a = open(os.path.join(out1, "customers.csv"), "rb").read()
+    b = open(os.path.join(out2, "customers.csv"), "rb").read()
+    assert a != b
+
+
+def test_fk_containment(tmp_path):
+    out = str(tmp_path / "folder")
+    res = generate_synthetic_eda_folder(out, n_rows=300, seed=42)
+    customers = pd.read_csv(res["files"]["customers"])
+    orders = pd.read_csv(res["files"]["orders"])
+    reviews = pd.read_csv(res["files"]["reviews"])
+    cust_ids = set(customers["customer_id"])
+    # Todos los customer_id de orders y reviews ⊆ customers.
+    assert set(orders["customer_id"]) <= cust_ids
+    assert set(reviews["customer_id"]) <= cust_ids
+    # customer_id es PK unica en customers.
+    assert customers["customer_id"].is_unique
+    assert orders["order_id"].is_unique
+    assert reviews["review_id"].is_unique
+
+
+def test_review_text_mediana_palabras(tmp_path):
+    out = str(tmp_path / "folder")
+    res = generate_synthetic_eda_folder(out, n_rows=300, seed=42)
+    reviews = pd.read_csv(res["files"]["reviews"])
+    words = [len(str(t).split()) for t in reviews["review_text"].dropna()]
+    assert statistics.median(words) >= 20
+
+
+def test_n_rows_invalido(tmp_path):
+    out = str(tmp_path / "folder")
+    res = generate_synthetic_eda_folder(out, n_rows=0, seed=42)
+    assert res["status"] == "error"

python/functions/datascience/generate_synthetic_eda_table.md

@@ -0,0 +1,82 @@
+---
+name: generate_synthetic_eda_table
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def generate_synthetic_eda_table(out_db_path: str, table: str = 'synthetic', n_rows: int = 2000, seed: int = 42) -> dict"
+description: "Genera una tabla DuckDB sintetica (Faker + numpy, determinista por seed) cuyo contenido esta disenado para ACTIVAR el maximo de capitulos del motor AutomaticEDA del grupo eda: numericas continuas con correlacion lineal/no-lineal, numericas con outliers, categoricas desbalanceadas, texto libre multi-idioma con duplicados, fecha para serie temporal, lat/lon validas, semanticos/PII (uuid/email/iban/phone) y nulos con patron MCAR/MAR. Fixture para evaluar el EDA de punta a punta. Estilo dict-no-throw: nunca lanza."
+tags: [eda, synthetic, faker, testing, fixture, datascience]
+params:
+  - name: out_db_path
+    desc: "Ruta al archivo DuckDB de salida. Se crea (o reutiliza) y la tabla se reemplaza con CREATE OR REPLACE TABLE si ya existe."
+  - name: table
+    desc: "Nombre de la tabla a crear. Se valida contra ^[A-Za-z_][A-Za-z0-9_]*$ y se cita en el DDL. Default 'synthetic'."
+  - name: n_rows
+    desc: "Numero de filas (clientes unicos). Cada fila es un cliente con id/email/iban/phone propios. Default 2000."
+  - name: seed
+    desc: "Semilla para Faker (Faker.seed) y numpy (np.random.default_rng). Mismo seed -> tabla identica byte a byte. Default 42."
+output: "dict dict-no-throw. En exito {status:'ok', db_path, table, n_rows, columns:[19 nombres de columna], seed}. En error (sin lanzar, p.ej. nombre de tabla invalido o n_rows<=0) {status:'error', error:str}. Columnas: customer_id,email,iban,phone,income,spending,age,risk_score,tenure_months,engagement_quad,amount,n_purchases,country,category,plan,review,signup_date,latitude,longitude."
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: []
+tested: true
+tests: ["test_genera_ok_y_columnas", "test_determinismo_mismo_seed", "test_seeds_distintos_difieren", "test_latlon_en_rango", "test_plan_solo_niveles_validos", "test_income_spending_co_nulos", "test_review_mediana_palabras_y_signup_datetime", "test_phone_matchea_regex_internacional", "test_outliers_y_correlaciones", "test_tabla_invalida_devuelve_error"]
+test_file_path: "python/functions/datascience/generate_synthetic_eda_table_test.py"
+file_path: "python/functions/datascience/generate_synthetic_eda_table.py"
+---
+
+## Ejemplo
+
+```bash
+# Genera /tmp/x.duckdb con la tabla `synthetic` (2000 filas, seed 42)
+fn run generate_synthetic_eda_table /tmp/x.duckdb synthetic 2000 42
+```
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience import generate_synthetic_eda_table
+
+res = generate_synthetic_eda_table("/tmp/x.duckdb", "synthetic", n_rows=2000, seed=42)
+# res == {"status":"ok", "db_path":"/tmp/x.duckdb", "table":"synthetic",
+#         "n_rows":2000, "columns":[...19...], "seed":42}
+# Luego perfilala con el grupo eda:
+#   fn run profile_table /tmp/x.duckdb synthetic
+```
+
+## Cuando usarla
+
+- Cuando necesites un dataset de prueba REPRODUCIBLE para evaluar el motor AutomaticEDA de punta a punta: su contenido dispara, a proposito, num_distr, cat_distr, text_distr, correlacion, missingness (MCAR/MAR), modelos (PCA/KMeans/outliers), timeseries, geospatial, calidad, agregacion y los detectores semanticos / PII (`infer_semantic_type`).
+- Cuando escribas tests de capitulos del EDA y quieras una tabla con una columna que active CADA detector sin montar datos a mano.
+- Cuando quieras un fixture determinista (mismo seed -> misma tabla) para comparar el render del EDA entre versiones.
+
+## Gotchas
+
+- **Impura**: escribe a disco (crea/reutiliza el archivo DuckDB). Reemplaza la tabla destino con `CREATE OR REPLACE`.
+- **Requiere `faker`, `duckdb`, `numpy` y `pandas`** instalados en el venv. Sin `faker` la generacion devuelve `{status:'error'}` (no lanza).
+- **`signup_date` queda como TIMESTAMP/DATE en DuckDB** (se construye con `datetime64[ns]`), NO VARCHAR — condicion para que `detect_time_column` la elija y se active el capitulo timeseries. Si fuese VARCHAR, el detector de fecha fallaria.
+- **El texto de `review` debe superar el gate de text_distr**: media de caracteres >= 50 y mediana de palabras >= 20. Por eso cada review concatena dos parrafos Faker (~50 palabras de mediana); no reducir el numero de frases o el capitulo text_distr no activa.
+- **Determinismo dependiente del orden de llamadas**: se siembra `Faker.seed(seed)` + `np.random.default_rng(seed)` al inicio; cambiar el orden de las extracciones cambia la salida aunque el seed sea el mismo.
+- **PII real-istica**: `email`/`iban`/`phone`/`customer_id` matchean los regex de `infer_semantic_type` (email/iban/phone_intl/uuid) al 100%; son datos sinteticos de Faker, no personas reales.
+
+## Notas
+
+Mapa columna -> detector que activa:
+
+| Columna(s) | Tipo | Detector / capitulo |
+|---|---|---|
+| income, spending | num continua | correlacion POSITIVA fuerte (Pearson > 0.8) |
+| age, risk_score | num continua | correlacion NEGATIVA |
+| tenure_months, engagement_quad | num continua | relacion NO LINEAL (cuadratica) |
+| amount, n_purchases | num + outliers | num_distr / outliers (cola pesada + extremos inyectados) |
+| country (12), category (6), plan (3 desbalanceado) | categorica | cat_distr / agregacion (entropia baja en plan) |
+| review | texto libre multi-idioma | text_distr (len_mean>=50, mediana palabras>=20) + duplicados exactos |
+| signup_date | DATE/TIMESTAMP | timeseries |
+| latitude, longitude | num [-90,90]/[-180,180] | geospatial (detect_latlon_columns) |
+| customer_id, email, iban, phone | texto | semantic_type uuid/email/iban/phone_intl (PII) |
+| income+spending (co-nulos 12%), risk_score (nulo si plan=alta), review (8%) | nulos con patron | missingness MCAR/MAR |

python/functions/datascience/generate_synthetic_eda_table.py

@@ -0,0 +1,314 @@
+"""generate_synthetic_eda_table — fixture sintetico para ejercitar el motor AutomaticEDA.
+
+Funcion impura (escribe un archivo DuckDB a disco) y determinista por ``seed``:
+construye una unica tabla cuyo CONTENIDO esta disenado para ACTIVAR el maximo
+numero de capitulos del motor AutomaticEDA del grupo `eda` (num_distr, cat_distr,
+text_distr, correlacion, missingness, modelos, timeseries, geospatial, relaciones,
+calidad, agregacion) y los detectores semanticos / PII (`infer_semantic_type`).
+
+Estilo dict-no-throw del grupo `eda`: NUNCA lanza; captura cualquier error y
+devuelve ``{"status": "error", "error": str}``.
+
+Determinismo: con el mismo ``seed`` el DataFrame y, por tanto, la tabla DuckDB
+resultante son identicos byte a byte. Se siembra Faker (``Faker.seed``) y numpy
+(``np.random.default_rng(seed)``) al inicio de cada generacion.
+"""
+
+import re
+
+# Lista fija de paises (12 -> cardinalidad media para cat_distr / agregacion).
+_COUNTRIES = [
+    "ES", "FR", "DE", "IT", "PT", "NL",
+    "BE", "US", "GB", "IE", "SE", "PL",
+]
+
+# Lista fija de categorias de producto (6 -> cardinalidad media).
+_CATEGORIES = [
+    "electronics", "clothing", "home", "sports", "books", "toys",
+]
+
+# Niveles de plan con probabilidades DESBALANCEADAS (entropia baja para cat_distr).
+_PLANS = ["baja", "media", "alta"]
+_PLAN_PROBS = [0.70, 0.25, 0.05]
+
+# Centroides (lat, lon) aproximados por pais: muestrean coordenadas validas
+# dentro de [-90, 90] x [-180, 180] para que detect_latlon_columns las acepte.
+_CENTROIDS = {
+    "ES": (40.4, -3.7), "FR": (46.6, 2.2), "DE": (51.1, 10.4), "IT": (41.9, 12.5),
+    "PT": (39.4, -8.2), "NL": (52.1, 5.3), "BE": (50.5, 4.5), "US": (39.0, -98.0),
+    "GB": (54.0, -2.0), "IE": (53.4, -8.0), "SE": (60.1, 18.6), "PL": (52.0, 19.1),
+}
+
+# Locales rotados para generar texto multi-idioma (es/en/fr).
+_TEXT_LOCALES = ["es_ES", "en_US", "fr_FR"]
+
+# Identificador SQL valido (DuckDB no parametriza el nombre de tabla en DDL).
+_IDENT_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
+
+
+def _make_fakers(seed):
+    """Crea los Faker por locale tras sembrar el generador compartido.
+
+    ``Faker.seed(seed)`` siembra el ``random.Random`` compartido por todas las
+    instancias Faker que usan el generador por defecto, asi que el orden de
+    llamadas determina por completo la salida (determinismo).
+    """
+    from faker import Faker
+
+    Faker.seed(seed)
+    es_es, en_us, fr_fr = (Faker(loc) for loc in _TEXT_LOCALES)
+    return {"es_ES": es_es, "en_US": en_us, "fr_FR": fr_fr}
+
+
+# Texto duplicado canonico (multi-idioma, > 20 palabras) que se inyecta en una
+# fraccion de las filas para que el analisis de duplicados exactos lo detecte.
+_DUP_REVIEW = (
+    "Servicio excelente y entrega muy rapida, el producto llego en perfecto "
+    "estado y coincide con la descripcion publicada en la tienda. The customer "
+    "support team answered every question quickly and the packaging was solid "
+    "and well protected during shipping. Je recommande vivement ce vendeur a "
+    "tous mes amis, la qualite est vraiment au rendez-vous cette fois."
+)
+
+
+def _make_reviews(n, rng, fakers, dup_frac=0.04, null_frac=0.08):
+    """Genera ``n`` reviews de texto libre largo multi-idioma (es/en/fr).
+
+    Cada review concatena dos parrafos de Faker en el idioma rotado por fila, de
+    modo que la MEDIANA de palabras por documento queda muy por encima de 20 y la
+    media de caracteres por encima de 50 (gates del capitulo text_distr). Se
+    inyectan duplicados exactos (``dup_frac``) y nulos (``null_frac``).
+
+    Devuelve una ``list`` de ``str`` o ``None`` (nulos) de longitud ``n``.
+    """
+    # Numero de frases por parrafo precomputado con numpy (determinista) para no
+    # interleavar draws de rng dentro del bucle de faker.
+    nb1 = rng.integers(4, 8, n)
+    nb2 = rng.integers(3, 7, n)
+
+    reviews = []
+    for i in range(n):
+        fk = fakers[_TEXT_LOCALES[i % 3]]
+        p1 = fk.paragraph(nb_sentences=int(nb1[i]))
+        p2 = fk.paragraph(nb_sentences=int(nb2[i]))
+        reviews.append(f"{p1} {p2}")
+
+    # Duplicados exactos: una fraccion de filas comparte un review identico.
+    if n > 0 and dup_frac > 0:
+        k_dup = max(1, int(n * dup_frac))
+        dup_idx = rng.choice(n, size=min(k_dup, n), replace=False)
+        for j in dup_idx:
+            reviews[int(j)] = _DUP_REVIEW
+
+    # Nulos MCAR-ish: una fraccion de filas al azar queda en None.
+    if n > 0 and null_frac > 0:
+        k_null = max(1, int(n * null_frac))
+        null_idx = rng.choice(n, size=min(k_null, n), replace=False)
+        for j in null_idx:
+            reviews[int(j)] = None
+
+    return reviews
+
+
+def _make_phone_intl(rng):
+    """Construye un telefono en formato internacional que casa phone_intl.
+
+    Regex objetivo (fullmatch): ``\\+\\d[\\d\\s()-]{6,}\\d``. Empieza por '+',
+    digito, bloques de digitos separados por espacios y termina en digito.
+    """
+    cc = int(rng.integers(1, 99))
+    a = int(rng.integers(100, 999))
+    b = int(rng.integers(100, 999))
+    c = int(rng.integers(100, 999))
+    return f"+{cc} {a} {b} {c}"
+
+
+def _make_latlon(countries, rng):
+    """Devuelve (latitudes, longitudes) muestreando centroides de pais + jitter.
+
+    Mantiene los valores dentro de [-90, 90] y [-180, 180] (validez exigida por
+    detect_latlon_columns). El jitter es pequeno para no salirse del rango.
+    """
+    import numpy as np
+
+    lats = np.empty(len(countries), dtype=float)
+    lons = np.empty(len(countries), dtype=float)
+    jitter_lat = rng.normal(0.0, 0.5, len(countries))
+    jitter_lon = rng.normal(0.0, 0.5, len(countries))
+    for i, code in enumerate(countries):
+        base_lat, base_lon = _CENTROIDS[code]
+        lats[i] = float(np.clip(base_lat + jitter_lat[i], -90.0, 90.0))
+        lons[i] = float(np.clip(base_lon + jitter_lon[i], -180.0, 180.0))
+    return lats, lons
+
+
+def _amount_with_outliers(n, rng, n_extreme=6, factor=50.0):
+    """Serie lognormal de cola pesada con ~``n_extreme`` outliers altos (x``factor``)."""
+    import numpy as np
+
+    amount = rng.lognormal(mean=4.0, sigma=1.0, size=n)
+    if n > 0 and n_extreme > 0:
+        idx = rng.choice(n, size=min(n_extreme, n), replace=False)
+        amount[idx] = amount[idx] * factor
+    return amount
+
+
+def generate_synthetic_eda_table(
+    out_db_path, table="synthetic", n_rows=2000, seed=42
+):
+    """Genera una tabla DuckDB sintetica que activa el maximo de capitulos del EDA.
+
+    Construye un DataFrame de ``n_rows`` clientes unicos con columnas elegidas para
+    disparar detectores concretos del motor AutomaticEDA (numericas continuas con
+    correlaciones lineal/no-lineal, numericas con outliers, categoricas
+    desbalanceadas, texto libre multi-idioma con duplicados, fecha para serie
+    temporal, lat/lon validas, semanticos/PII y nulos con patron MCAR/MAR), y la
+    materializa en ``out_db_path`` con ``CREATE OR REPLACE TABLE``.
+
+    Funcion impura (escribe a disco) y determinista por ``seed``: con el mismo
+    seed la tabla resultante es identica byte a byte. NUNCA lanza.
+
+    Args:
+        out_db_path: ruta al archivo DuckDB de salida. Se crea (o reutiliza) y la
+            tabla se reemplaza si ya existe.
+        table: nombre de la tabla a crear. Se valida contra
+            ``^[A-Za-z_][A-Za-z0-9_]*$`` y se cita en el DDL.
+        n_rows: numero de filas (clientes unicos). Default 2000.
+        seed: semilla para Faker y numpy. Default 42.
+
+    Returns:
+        dict dict-no-throw. En exito::
+
+            {"status": "ok", "db_path": out_db_path, "table": table,
+             "n_rows": n_rows, "columns": [<nombres de columna>], "seed": seed}
+
+        En error (sin lanzar)::
+
+            {"status": "error", "error": str}
+    """
+    try:
+        import duckdb
+        import numpy as np
+        import pandas as pd
+
+        if not _IDENT_RE.match(table or ""):
+            return {
+                "status": "error",
+                "error": (
+                    f"nombre de tabla invalido: {table!r} "
+                    "(debe casar con ^[A-Za-z_][A-Za-z0-9_]*$)"
+                ),
+            }
+        n = int(n_rows)
+        if n <= 0:
+            return {"status": "error", "error": f"n_rows debe ser > 0, dado {n_rows!r}"}
+
+        fakers = _make_fakers(seed)
+        rng = np.random.default_rng(seed)
+
+        # --- Numericas continuas (distinct alto, correlaciones) ---
+        income = np.clip(rng.normal(40000.0, 12000.0, n), 1000.0, None)
+        spending = income * 0.35 + rng.normal(0.0, 2000.0, n)  # corr POSITIVA fuerte
+        age = rng.integers(18, 91, n)
+        risk_score = 90.0 - age * 0.7 + rng.normal(0.0, 5.0, n)  # corr NEGATIVA con age
+        tenure_months = rng.uniform(0.0, 60.0, n)
+        engagement_quad = ((tenure_months - 30.0) ** 2) / 30.0 + rng.normal(0.0, 1.0, n)
+
+        # --- Numericas con outliers claros ---
+        amount = _amount_with_outliers(n, rng)
+        n_purchases = rng.poisson(3.0, n).astype(float)
+        if n > 0:
+            k_hi = min(max(1, int(n * 0.002)) + 2, n)  # ~3-5 valores altisimos
+            hi_idx = rng.choice(n, size=k_hi, replace=False)
+            n_purchases[hi_idx] = rng.integers(200, 400, len(hi_idx)).astype(float)
+
+        # --- Categoricas ---
+        country = rng.choice(_COUNTRIES, n)
+        category = rng.choice(_CATEGORIES, n)
+        plan = rng.choice(_PLANS, n, p=_PLAN_PROBS)
+
+        # --- Texto libre multi-idioma con duplicados ---
+        review = _make_reviews(n, rng, fakers)
+
+        # --- Fecha / serie temporal (rango ~2 anios, cadencia ~diaria) ---
+        base = np.datetime64("2022-01-01")
+        offsets = rng.integers(0, 730, n)
+        signup_date = pd.to_datetime(base) + pd.to_timedelta(offsets, unit="D")
+
+        # --- Geo lat/lon validas ---
+        latitude, longitude = _make_latlon(country, rng)
+
+        # --- Semanticos / PII (>=80% match para infer_semantic_type) ---
+        customer_id = [fakers["en_US"].uuid4() for _ in range(n)]
+        email = [fakers["en_US"].email() for _ in range(n)]
+        iban = [fakers["en_US"].iban() for _ in range(n)]
+        phone = [_make_phone_intl(rng) for _ in range(n)]
+
+        df = pd.DataFrame(
+            {
+                "customer_id": customer_id,
+                "email": email,
+                "iban": iban,
+                "phone": phone,
+                "income": income,
+                "spending": spending,
+                "age": age,
+                "risk_score": risk_score,
+                "tenure_months": tenure_months,
+                "engagement_quad": engagement_quad,
+                "amount": amount,
+                "n_purchases": n_purchases,
+                "country": country,
+                "category": category,
+                "plan": plan,
+                "review": review,
+                "signup_date": signup_date,
+                "latitude": latitude,
+                "longitude": longitude,
+            }
+        )
+
+        # --- Nulos con patron ---
+        # income + spending faltan JUNTAS en las MISMAS filas (co-ocurrencia -> MAR).
+        k_co = max(1, int(n * 0.12))
+        co_idx = rng.choice(n, size=min(k_co, n), replace=False)
+        df.loc[co_idx, "income"] = np.nan
+        df.loc[co_idx, "spending"] = np.nan
+        # risk_score falta cuando plan == "alta" (mas una pizca de azar) -> MAR.
+        risk_mask = (df["plan"] == "alta").to_numpy() | (rng.random(n) < 0.02)
+        df.loc[risk_mask, "risk_score"] = np.nan
+
+        columns = list(df.columns)
+
+        con = duckdb.connect(out_db_path)
+        try:
+            con.register("df_synth_eda", df)
+            con.execute(
+                f'CREATE OR REPLACE TABLE "{table}" AS SELECT * FROM df_synth_eda'
+            )
+            con.unregister("df_synth_eda")
+        finally:
+            con.close()
+
+        return {
+            "status": "ok",
+            "db_path": out_db_path,
+            "table": table,
+            "n_rows": n,
+            "columns": columns,
+            "seed": seed,
+        }
+    except Exception as exc:  # noqa: BLE001 — dict-no-throw del grupo eda.
+        return {"status": "error", "error": str(exc)}
+
+
+if __name__ == "__main__":
+    import json
+    import sys
+
+    args = sys.argv[1:]
+    db_path = args[0] if len(args) > 0 else "/tmp/synthetic_eda.duckdb"
+    tbl = args[1] if len(args) > 1 else "synthetic"
+    rows = int(args[2]) if len(args) > 2 else 2000
+    sd = int(args[3]) if len(args) > 3 else 42
+    print(json.dumps(generate_synthetic_eda_table(db_path, tbl, rows, sd), indent=2))

python/functions/datascience/generate_synthetic_eda_table_test.py

@@ -0,0 +1,129 @@
+"""Tests para generate_synthetic_eda_table."""
+
+import os
+import re
+import statistics
+
+import duckdb
+
+from datascience.generate_synthetic_eda_table import generate_synthetic_eda_table
+
+_EXPECTED_COLS = [
+    "customer_id", "email", "iban", "phone", "income", "spending", "age",
+    "risk_score", "tenure_months", "engagement_quad", "amount", "n_purchases",
+    "country", "category", "plan", "review", "signup_date", "latitude", "longitude",
+]
+_PHONE_RE = re.compile(r"\+\d[\d\s()-]{6,}\d")
+
+
+def _load(db_path, table="synthetic"):
+    con = duckdb.connect(db_path, read_only=True)
+    try:
+        return con.execute(f'SELECT * FROM "{table}"').fetch_df()
+    finally:
+        con.close()
+
+
+def test_genera_ok_y_columnas(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    res = generate_synthetic_eda_table(db, "synthetic", n_rows=500, seed=42)
+    assert res["status"] == "ok"
+    assert res["table"] == "synthetic"
+    assert res["n_rows"] == 500
+    assert res["columns"] == _EXPECTED_COLS
+    assert os.path.exists(db)
+    df = _load(db)
+    assert list(df.columns) == _EXPECTED_COLS
+    assert len(df) == 500
+
+
+def test_determinismo_mismo_seed(tmp_path):
+    db1 = str(tmp_path / "a.duckdb")
+    db2 = str(tmp_path / "b.duckdb")
+    generate_synthetic_eda_table(db1, "synthetic", n_rows=400, seed=7)
+    generate_synthetic_eda_table(db2, "synthetic", n_rows=400, seed=7)
+    df1 = _load(db1).astype(str)
+    df2 = _load(db2).astype(str)
+    # Misma semilla -> tabla identica fila a fila.
+    assert df1.equals(df2)
+
+
+def test_seeds_distintos_difieren(tmp_path):
+    db1 = str(tmp_path / "a.duckdb")
+    db2 = str(tmp_path / "b.duckdb")
+    generate_synthetic_eda_table(db1, "synthetic", n_rows=400, seed=7)
+    generate_synthetic_eda_table(db2, "synthetic", n_rows=400, seed=8)
+    df1 = _load(db1).astype(str)
+    df2 = _load(db2).astype(str)
+    assert not df1.equals(df2)
+
+
+def test_latlon_en_rango(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=500, seed=42)
+    df = _load(db)
+    assert df["latitude"].between(-90, 90).all()
+    assert df["longitude"].between(-180, 180).all()
+
+
+def test_plan_solo_niveles_validos(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=500, seed=42)
+    df = _load(db)
+    assert set(df["plan"].unique()) <= {"baja", "media", "alta"}
+
+
+def test_income_spending_co_nulos(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=600, seed=42)
+    df = _load(db)
+    inc_null = df["income"].isna()
+    sp_null = df["spending"].isna()
+    # income y spending faltan exactamente en las MISMAS filas.
+    assert (inc_null == sp_null).all()
+    assert inc_null.sum() > 0
+
+
+def test_review_mediana_palabras_y_signup_datetime(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=500, seed=42)
+    df = _load(db)
+    words = [len(str(r).split()) for r in df["review"].dropna()]
+    assert statistics.median(words) >= 20
+    # signup_date debe ser datetime/date en DuckDB (no VARCHAR).
+    con = duckdb.connect(db, read_only=True)
+    try:
+        dtype = con.execute(
+            "SELECT column_type FROM (DESCRIBE synthetic) WHERE column_name='signup_date'"
+        ).fetchone()[0]
+    finally:
+        con.close()
+    assert dtype.upper().startswith(("DATE", "TIMESTAMP"))
+
+
+def test_phone_matchea_regex_internacional(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=500, seed=42)
+    df = _load(db)
+    phones = [p for p in df["phone"].tolist() if p is not None]
+    assert all(_PHONE_RE.fullmatch(str(p)) for p in phones)
+
+
+def test_outliers_y_correlaciones(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    generate_synthetic_eda_table(db, "synthetic", n_rows=800, seed=42)
+    df = _load(db)
+    # amount tiene cola con outliers altos evidentes.
+    assert df["amount"].max() > df["amount"].median() * 20
+    # correlacion positiva fuerte income~spending y negativa age~risk_score.
+    sub = df[["income", "spending"]].dropna()
+    assert sub["income"].corr(sub["spending"]) > 0.8
+    sub2 = df[["age", "risk_score"]].dropna()
+    assert sub2["age"].corr(sub2["risk_score"]) < -0.6
+
+
+def test_tabla_invalida_devuelve_error(tmp_path):
+    db = str(tmp_path / "t.duckdb")
+    res = generate_synthetic_eda_table(db, "bad name;", n_rows=10, seed=42)
+    assert res["status"] == "error"
+    assert "invalido" in res["error"]

python/functions/datascience/missingness_corr_heatmap_figure.md

@@ -0,0 +1,103 @@
+---
+id: missingness_corr_heatmap_figure_py_datascience
+name: missingness_corr_heatmap_figure
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: impure
+signature: "def missingness_corr_heatmap_figure(matrix, labels, title=\"Co-ocurrencia de ausencias\") -> \"matplotlib.figure.Figure\""
+description: "Construye una figura matplotlib (heatmap) de la matriz NxN de correlación de ausencias entre columnas: +1 = dos columnas suelen ser nulas a la vez, -1 = cuando una falta la otra está presente, 0 = ausencias independientes. Usa ax.imshow con coolwarm fijado a [-1,1], ticks con los labels truncados (X rotados 45º), colorbar y anota el valor de cada celda si N<=12. Devuelve un matplotlib.figure.Figure listo para rasterizar por el renderer del informe EDA (capítulo de datos faltantes). Backend Agg sin pyplot global; defensivo ante matrix/labels vacíos o celdas no numéricas (nunca lanza)."
+tags: [eda, missing, missingness, correlation, heatmap, matplotlib, figure, visualization, datascience, impure]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: "error_go_core"
+imports: [matplotlib]
+example: |
+  from datascience.missingness_corr_heatmap_figure import missingness_corr_heatmap_figure
+  matrix = [
+      [1.0, 0.82, -0.10],
+      [0.82, 1.0, 0.05],
+      [-0.10, 0.05, 1.0],
+  ]
+  labels = ["telefono", "movil", "email"]
+  fig = missingness_corr_heatmap_figure(matrix, labels, title="Co-ocurrencia de ausencias")
+tested: true
+tests:
+  - "test_returns_figure_with_axes"
+  - "test_empty_matrix_does_not_raise_and_returns_figure"
+  - "test_empty_labels_returns_message_figure"
+  - "test_large_matrix_omits_annotations"
+  - "test_ragged_and_non_numeric_cells_are_handled"
+test_file_path: "python/functions/datascience/missingness_corr_heatmap_figure_test.py"
+file_path: "python/functions/datascience/missingness_corr_heatmap_figure.py"
+params:
+  - name: matrix
+    desc: "Lista de listas (NxN) de floats en [-1,1]: la correlación de ausencias por pares de columnas. Puede venir vacía. Filas de longitud desigual se toleran (se rellenan/recortan a N); celdas None, NaN o no numéricas se coercen a 0.0. No se muta el original."
+  - name: labels
+    desc: "Lista de N nombres de columna, paralela a matrix. Puede venir vacía (devuelve figura \"sin columnas con ausencia variable\"). Se truncan a ~14 chars con elipsis para los ticks; los originales no se mutan."
+  - name: title
+    desc: "Título de la figura. Se trunca a ~60 chars con elipsis si es muy largo. Default \"Co-ocurrencia de ausencias\"."
+output: "Un matplotlib.figure.Figure (figsize 6.4x5.2, dpi 150) con un Axes heatmap (imshow vmin=-1, vmax=1, cmap coolwarm) más una colorbar etiquetada \"correlación de ausencias\". Ticks en ambos ejes con los labels truncados (X rotados 45º). Si N<=12 cada celda lleva su valor numérico anotado (texto blanco sobre celdas saturadas, oscuro sobre pálidas); con N grande se omiten las anotaciones para no saturar. Si matrix o labels vienen vacíos devuelve una Figure con texto centrado \"sin columnas con ausencia variable\"; cualquier error inesperado se captura y devuelve una Figure con el mensaje de error (nunca lanza). El caller rasteriza/cierra la figura; la función no la muestra ni la guarda."
+---
+
+## Ejemplo
+
+```python
+from datascience.missingness_corr_heatmap_figure import missingness_corr_heatmap_figure
+
+# Correlación de ausencias entre 3 columnas de contacto:
+# telefono y movil tienden a faltar juntos (0.82); email es casi independiente.
+matrix = [
+    [1.00, 0.82, -0.10],
+    [0.82, 1.00,  0.05],
+    [-0.10, 0.05, 1.00],
+]
+labels = ["telefono", "movil", "email"]
+
+fig = missingness_corr_heatmap_figure(
+    matrix,
+    labels,
+    title="Co-ocurrencia de ausencias",
+)
+
+# El renderer del informe lo rasteriza; aquí solo persistimos para inspección.
+fig.savefig("/tmp/missingness_heatmap.png")
+```
+
+## Cuando usarla
+
+Úsala en el capítulo de datos faltantes de un informe EDA cuando quieras ver de
+un vistazo qué columnas faltan juntas (mismo formulario sin rellenar, mismo
+proceso roto) frente a columnas cuyas ausencias son independientes. Pásale la
+matriz de correlación de ausencias (calculada sobre la máscara de nulos, p. ej.
+`df.isnull().corr()`) restringida a las columnas que de verdad tienen ausencia
+variable, junto con sus nombres. Es la pareja "estructura" del ranking de % de
+nulos: las barras dicen *cuánto* falta cada columna, este heatmap dice *si las
+ausencias están relacionadas* entre columnas.
+
+## Gotchas
+
+- **Impura por matplotlib.** Toca la maquinaria de render. Usa el backend `Agg`
+  y la API orientada a objetos `Figure`/`add_subplot` — NUNCA `pyplot.*` aquí,
+  para no tocar el estado global ni filtrar figuras entre llamadas. `pyplot` NO
+  es thread-safe; esta función evita ese riesgo construyendo el `Figure`
+  directamente, así que es segura de llamar en bucle desde el renderer.
+- **El caller cierra la figura.** Devuelve el `Figure` pero no lo muestra ni lo
+  guarda. Quien la consume debe rasterizarla y luego liberarla
+  (`matplotlib.pyplot.close(fig)`) para no acumular memoria en lotes grandes.
+- **Escala de color fija en [-1, 1].** `vmin=-1`, `vmax=1` están fijados a
+  propósito para que el color sea comparable entre informes y entre columnas. No
+  se autoescala al rango real de la matriz; valores fuera de `[-1, 1]` se
+  saturan al extremo del colormap.
+- **Anotaciones solo con N<=12.** Por encima de 12 columnas el grid de números
+  se vuelve ilegible y se omite; queda solo el color + la colorbar. Filtra a las
+  columnas con ausencia variable antes de llamar para no llegar a matrices
+  enormes.
+- **Defensiva, nunca lanza.** `matrix=[]`, `labels=[]`, filas cortas, celdas
+  `None`/`NaN`/no numéricas o cualquier error inesperado se manejan sin propagar:
+  en el peor caso devuelve una `Figure` con "sin columnas con ausencia variable"
+  o con el texto del error. No envuelvas la llamada en try/except por miedo a un
+  raise — no lo hay.

python/functions/datascience/missingness_corr_heatmap_figure.py

@@ -0,0 +1,158 @@
+"""Impure EDA helper: heatmap of missingness co-occurrence (`eda` group).
+
+Builds a matplotlib heatmap of the pairwise missingness correlation matrix of a
+dataset: a value near ``+1`` means two columns tend to be null together, near
+``-1`` means when one is null the other tends to be present, and ``0`` means
+their absences are independent. Returns a ready-to-rasterize
+``matplotlib.figure.Figure``; it never shows nor saves it.
+
+Impure because it touches matplotlib's rendering machinery. It uses the headless
+Agg backend and the object-oriented ``Figure`` API (no ``pyplot``) so it leaks no
+global state and is safe to call repeatedly from a report renderer.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+from matplotlib.figure import Figure  # noqa: E402
+
+# Muted gray for secondary text (no-data / fallback messages).
+_MUTED_TEXT = "#5f6b7a"
+# Soft red for the error fallback message (kept readable, not alarming).
+_ERROR_TEXT = "#b00020"
+
+
+def _truncate(text, width: int = 14) -> str:
+    """Truncate ``text`` to ``width`` chars, appending an ellipsis if cut."""
+    s = "" if text is None else str(text)
+    if len(s) <= width:
+        return s
+    if width <= 1:
+        return s[:width]
+    return s[: width - 1] + "…"
+
+
+def _message_figure(message: str, color: str = _MUTED_TEXT) -> "Figure":
+    """Return a fallback ``Figure`` carrying a single centered message."""
+    fig = Figure(figsize=(6.4, 4.0), dpi=150)
+    ax = fig.add_subplot(111)
+    ax.axis("off")
+    ax.text(
+        0.5,
+        0.5,
+        message,
+        ha="center",
+        va="center",
+        fontsize=12,
+        color=color,
+        wrap=True,
+        transform=ax.transAxes,
+    )
+    fig.tight_layout()
+    return fig
+
+
+def missingness_corr_heatmap_figure(
+    matrix,
+    labels,
+    title: str = "Co-ocurrencia de ausencias",
+) -> "matplotlib.figure.Figure":
+    """Build a heatmap figure of a missingness correlation matrix.
+
+    Renders an ``NxN`` matrix of missingness correlations in ``[-1, 1]`` with a
+    diverging ``coolwarm`` colormap (fixed ``vmin=-1``, ``vmax=1`` so the color
+    scale is comparable across reports). Both axes are tick-labelled with the
+    column names (truncated to ~14 chars; the X labels rotated 45°). A colorbar
+    is attached. When the matrix is small (``N <= 12``) each cell is annotated
+    with its numeric value; for larger matrices the annotations are omitted to
+    avoid an unreadable grid.
+
+    The function is fully defensive: empty/ragged/non-numeric input never raises.
+    When there is nothing valid to draw it returns a ``Figure`` carrying a
+    centered "sin columnas con ausencia variable" message, and any unexpected
+    error is caught and turned into a fallback ``Figure`` carrying the error text.
+
+    Args:
+        matrix: List of lists (``NxN``) of floats in ``[-1, 1]`` — the pairwise
+            missingness correlation. May be empty; rows of unequal length are
+            tolerated by treating the matrix as invalid only when it is empty or
+            its label count does not match. Non-numeric/``None`` cells are
+            coerced to ``0.0``.
+        labels: List of ``N`` column names, parallel to ``matrix``. May be empty.
+            Truncated for display; the originals are not mutated.
+        title: Figure title. Default "Co-ocurrencia de ausencias".
+
+    Returns:
+        A ``matplotlib.figure.Figure`` with a single heatmap Axes plus a
+        colorbar. The caller is responsible for rasterizing/closing it.
+    """
+    try:
+        # --- Validate shape: need a non-empty square-ish matrix with labels.
+        if (
+            not isinstance(matrix, (list, tuple))
+            or not isinstance(labels, (list, tuple))
+            or len(matrix) == 0
+            or len(labels) == 0
+        ):
+            return _message_figure("sin columnas con ausencia variable")
+
+        n = len(labels)
+        # Build a clean NxN grid: coerce each cell to float, default 0.0, pad/clip
+        # rows so a ragged input never crashes imshow.
+        grid = []
+        for i in range(n):
+            row_src = matrix[i] if i < len(matrix) else []
+            if not isinstance(row_src, (list, tuple)):
+                row_src = []
+            row = []
+            for j in range(n):
+                cell = row_src[j] if j < len(row_src) else 0.0
+                try:
+                    val = float(cell)
+                except (TypeError, ValueError):
+                    val = 0.0
+                if val != val:  # NaN guard.
+                    val = 0.0
+                row.append(val)
+            grid.append(row)
+
+        fig = Figure(figsize=(6.4, 5.2), dpi=150)
+        ax = fig.add_subplot(111)
+
+        im = ax.imshow(grid, vmin=-1, vmax=1, cmap="coolwarm", aspect="equal")
+
+        short = [_truncate(lab, 14) for lab in labels]
+        ax.set_xticks(range(n))
+        ax.set_yticks(range(n))
+        ax.set_xticklabels(short, rotation=45, ha="right", fontsize=8)
+        ax.set_yticklabels(short, fontsize=8)
+
+        # Annotate each cell only when the grid is small enough to stay legible.
+        if n <= 12:
+            for i in range(n):
+                for j in range(n):
+                    val = grid[i][j]
+                    # White text over saturated (dark) cells, dark over pale.
+                    txt_color = "white" if abs(val) >= 0.55 else "#202020"
+                    ax.text(
+                        j,
+                        i,
+                        f"{val:.2f}",
+                        ha="center",
+                        va="center",
+                        fontsize=7,
+                        color=txt_color,
+                    )
+
+        cbar = fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
+        cbar.ax.tick_params(labelsize=8)
+        cbar.set_label("correlación de ausencias", fontsize=8)
+
+        if title:
+            ax.set_title(_truncate(title, 60), fontsize=12, loc="center", pad=10)
+
+        fig.tight_layout()
+        return fig
+    except Exception as exc:  # noqa: BLE001 — never raise from a figure builder.
+        return _message_figure(f"error al dibujar heatmap: {exc}", color=_ERROR_TEXT)

python/functions/datascience/missingness_corr_heatmap_figure_test.py

@@ -0,0 +1,62 @@
+"""Tests para missingness_corr_heatmap_figure (heatmap de ausencias, grupo eda).
+
+Usa el backend Agg sin pyplot; no muestra ni guarda figuras. Cada test cierra
+explícitamente la Figure construida (matplotlib.pyplot.close) para no acumular
+estado entre tests.
+"""
+
+import matplotlib
+
+matplotlib.use("Agg")
+
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.figure import Figure  # noqa: E402
+
+from missingness_corr_heatmap_figure import missingness_corr_heatmap_figure
+
+
+def _identity_matrix(n):
+    """Matriz NxN con diagonal 1.0 y resto 0.0 (correlación de ausencias)."""
+    return [[1.0 if i == j else 0.0 for j in range(n)] for i in range(n)]
+
+
+def test_returns_figure_with_axes():
+    matrix = [[1.0, 0.3, -0.2], [0.3, 1.0, 0.5], [-0.2, 0.5, 1.0]]
+    labels = ["edad", "ingresos", "ciudad"]
+    fig = missingness_corr_heatmap_figure(matrix, labels, title="ausencias")
+    assert isinstance(fig, Figure)
+    # Heatmap (>=1 axes) + colorbar añade su propio Axes -> al menos 1.
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_empty_matrix_does_not_raise_and_returns_figure():
+    fig = missingness_corr_heatmap_figure([], [], title="vacía")
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_empty_labels_returns_message_figure():
+    fig = missingness_corr_heatmap_figure([[1.0]], [], title="sin labels")
+    assert isinstance(fig, Figure)
+    plt.close(fig)
+
+
+def test_large_matrix_omits_annotations():
+    n = 16
+    fig = missingness_corr_heatmap_figure(
+        _identity_matrix(n), [f"col_{i}" for i in range(n)]
+    )
+    assert isinstance(fig, Figure)
+    assert len(fig.axes) >= 1
+    plt.close(fig)
+
+
+def test_ragged_and_non_numeric_cells_are_handled():
+    # Fila corta + celda None + celda string -> se rellenan/coercen sin lanzar.
+    matrix = [[1.0, None], ["x", 1.0, 0.5]]
+    labels = ["a", "b"]
+    fig = missingness_corr_heatmap_figure(matrix, labels)
+    assert isinstance(fig, Figure)
+    plt.close(fig)

python/functions/datascience/missingness_correlation.md

@@ -0,0 +1,68 @@
+---
+name: missingness_correlation
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def missingness_correlation(null_mask: dict, top_k: int = 20) -> dict"
+description: "Co-ocurrencia de ausencias: nucleo del capitulo de missingness del grupo eda. Recibe la mascara binaria de nulos de una tabla (1 = falta, 0 = presente, alineada por fila) y mide hasta que punto las columnas faltan juntas. Calcula la matriz de correlacion de Pearson entre los vectores binarios de ausencia de las columnas con varianza (al menos un 1 y un 0), mas las cifras de solapamiento de conjuntos por par (co-missing, either-missing, Jaccard). Excluye las columnas constantes en su ausencia (correlacion indefinida) y reporta cuantas. Compone la funcion atomica pearson del registry; no la reimplementa. Lectura defensiva; NUNCA lanza."
+tags: [eda, missingness, correlation, pearson, co-occurrence, jaccard, datascience]
+params:
+  - name: null_mask
+    desc: "dict {col: [int 0/1, ...]} con la mascara de ausencias de la tabla, alineada por fila: 1 = el valor falta en esa fila, 0 = presente. Todas las listas se asumen de la misma longitud (numero de filas). Valores truthy distintos de 0 se tratan como ausencia; entradas no-lista se ignoran sin romper."
+  - name: top_k
+    desc: "Numero maximo de pares a devolver en `pairs`, ordenados por valor absoluto de correlacion descendente. Default 20. Solo limita la lista de pares; la matriz cubre siempre todas las columnas con varianza."
+output: "dict con: columns (columnas con varianza en la ausencia, en orden de entrada); matrix (len(columns) x len(columns) de correlacion de Pearson entre las mascaras binarias, diagonal 1.0); pairs (hasta top_k pares i<j ordenados por |corr| desc, cada uno {a, b, corr, co_missing, either_missing, jaccard} donde co_missing = filas en que ambas faltan, either_missing = filas en que al menos una falta, jaccard = co_missing/either_missing o 0.0 si either_missing=0); n_excluded (nº de columnas con algun nulo pero sin varianza, constantes en la ausencia); excluded_cols (esas columnas en orden de entrada). Si hay <2 columnas con varianza, columns/matrix/pairs van vacios pero n_excluded/excluded_cols se rellenan. NUNCA lanza."
+uses_functions: [pearson_py_datascience]
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: []
+tested: true
+tests: ["test_co_ocurrencia_fuerte_corr_uno_jaccard_uno", "test_ausencias_disjuntas_corr_negativa_jaccard_cero", "test_columna_sin_varianza_se_excluye", "test_menos_de_dos_columnas_con_varianza_vacio_pero_cuenta_excluidas", "test_mask_vacio_todo_vacio", "test_top_k_limita_pares", "test_no_lanza_con_entradas_raras"]
+test_file_path: "python/functions/datascience/missingness_correlation_test.py"
+file_path: "python/functions/datascience/missingness_correlation.py"
+---
+
+## Ejemplo
+
+```python
+import sys, os
+sys.path.insert(0, os.path.join("python", "functions"))
+from datascience.missingness_correlation import missingness_correlation
+
+# Mascara de ausencias de 6 filas. 1 = falta, 0 = presente.
+mask = {
+    "ingresos":  [1, 0, 1, 0, 1, 0],   # falta junto a "deducciones"
+    "deducciones": [1, 0, 1, 0, 1, 0], # mismas filas que "ingresos"
+    "telefono":  [0, 0, 0, 1, 0, 0],   # casi siempre presente
+    "verificado": [1, 1, 1, 1, 1, 1],  # siempre ausente -> constante, excluida
+}
+out = missingness_correlation(mask, top_k=10)
+
+print(out["columns"])        # ['ingresos', 'deducciones', 'telefono']
+print(out["n_excluded"])     # 1
+print(out["excluded_cols"])  # ['verificado']
+
+# El par mas fuerte: ingresos y deducciones faltan siempre juntas.
+top = out["pairs"][0]
+print(top["a"], top["b"], round(top["corr"], 3))  # ingresos deducciones 1.0
+print(top["co_missing"], top["either_missing"], top["jaccard"])  # 3 3 1.0
+```
+
+## Cuando usarla
+
+- Usala en el capitulo de **missingness** de `AutomaticEDA` cuando ya tengas la mascara binaria de nulos por columna y quieras detectar **patrones de ausencia conjunta**: que columnas faltan siempre juntas (posible misma fuente/proceso roto) y cuales faltan de forma independiente.
+- Cuando necesites ordenar los pares de columnas por fuerza de co-ocurrencia (|corr|) para priorizar que bloques de ausencia investigar o imputar juntos.
+- Cuando quieras la cifra de solapamiento de conjuntos (Jaccard, co-missing) ademas de la correlacion lineal, para distinguir "faltan juntas" de "estan presentes juntas".
+- Antes de elegir una estrategia de imputacion: dos columnas con corr de ausencia ~1.0 no aportan informacion independiente sobre por que falta la otra.
+
+## Gotchas
+
+- Funcion pura, sin I/O y determinista. Lectura defensiva: entradas no-dict, columnas no-lista o vacias se ignoran sin lanzar.
+- Solo entran al calculo las columnas con **varianza en la ausencia** (al menos un 1 y al menos un 0). Una columna siempre-presente (todo 0) no aporta ausencia y **no** se cuenta como excluida; una columna siempre-ausente o constante con nulos (todo 1) tiene correlacion indefinida y se excluye, sumando a `n_excluded` / `excluded_cols`.
+- Con menos de 2 columnas con varianza, `columns`/`matrix`/`pairs` quedan vacios pero `n_excluded`/`excluded_cols` se rellenan igual — el caller debe contemplar el caso "sin pares".
+- La correlacion es la de Pearson sobre vectores binarios (equivale al coeficiente phi). El signo importa: corr negativa = las ausencias tienden a ser **complementarias** (cuando una falta, la otra suele estar presente).
+- Asume todas las listas alineadas por fila y de la misma longitud. Si vienen de longitudes distintas, `pearson` opera sobre el solapamiento que permita `zip` y degrada a 0.0 cuando no hay varianza efectiva; alinea la mascara antes de llamar.

python/functions/datascience/missingness_correlation.py

@@ -0,0 +1,120 @@
+"""Co-ocurrencia de ausencias: matriz de correlacion de Pearson entre mascaras de nulos.
+
+Funcion pura del grupo eda, nucleo del capitulo de missingness. Recibe la mascara
+binaria de ausencias de una tabla (1 = falta, 0 = presente, alineada por fila) y
+mide hasta que punto las columnas faltan juntas. Para cada par de columnas con
+varianza en su ausencia calcula la correlacion de Pearson entre los vectores
+binarios, mas las cifras de solapamiento de conjuntos (co-missing, either-missing,
+Jaccard). Compone la funcion atomica `pearson` del registry; no reimplementa la
+correlacion. Lectura defensiva; NUNCA lanza.
+"""
+
+from datascience import pearson
+
+
+def missingness_correlation(null_mask, top_k=20) -> dict:
+    """Correlacion de co-ocurrencia de ausencias entre columnas.
+
+    Args:
+        null_mask: dict {col: [int 0/1, ...]} alineado por fila (1 = el valor
+            falta en esa fila). Todas las listas se asumen de la misma longitud.
+        top_k: numero maximo de pares a devolver, ordenados por |corr| desc.
+
+    Returns:
+        dict con:
+          - columns: columnas con varianza en la ausencia (al menos un 1 y al
+            menos un 0), en orden de entrada.
+          - matrix: matriz len(columns) x len(columns) de correlacion de Pearson
+            entre las mascaras binarias, diagonal 1.0.
+          - pairs: lista de hasta top_k pares (i<j) ordenados por |corr| desc.
+            Cada par: {a, b, corr, co_missing, either_missing, jaccard}.
+          - n_excluded: numero de columnas con algun nulo pero sin varianza
+            (constantes en la ausencia: siempre presentes o siempre ausentes).
+          - excluded_cols: lista de esas columnas (en orden de entrada).
+
+        Si hay menos de 2 columnas con varianza, columns/matrix/pairs van vacios
+        pero n_excluded/excluded_cols se rellenan igualmente. NUNCA lanza.
+    """
+    # Salida base, defensiva ante entradas no-dict.
+    result = {
+        "columns": [],
+        "matrix": [],
+        "pairs": [],
+        "n_excluded": 0,
+        "excluded_cols": [],
+    }
+
+    if not isinstance(null_mask, dict) or not null_mask:
+        return result
+
+    varying = []          # columnas con varianza en la ausencia
+    varying_vecs = []     # sus vectores binarios saneados (floats 0.0/1.0)
+    excluded_cols = []    # columnas con nulos pero sin varianza (constantes)
+
+    for col, raw in null_mask.items():
+        if not isinstance(raw, (list, tuple)):
+            continue
+        # Sanea a 0/1: cualquier valor truthy distinto de 0 cuenta como ausencia.
+        vec = [1 if bool(v) else 0 for v in raw]
+        if not vec:
+            continue
+        ones = sum(vec)
+        zeros = len(vec) - ones
+        if ones > 0 and zeros > 0:
+            varying.append(col)
+            varying_vecs.append([float(v) for v in vec])
+        elif ones > 0:
+            # Tiene nulos pero todos (constante en la ausencia): sin varianza.
+            excluded_cols.append(col)
+        # ones == 0 -> columna siempre presente, sin nulos: no se cuenta como
+        # excluida (no aporta ausencia al analisis de co-ocurrencia).
+
+    result["n_excluded"] = len(excluded_cols)
+    result["excluded_cols"] = excluded_cols
+
+    n = len(varying)
+    if n < 2:
+        return result
+
+    result["columns"] = list(varying)
+
+    # Matriz de correlacion de Pearson, diagonal 1.0.
+    matrix = [[0.0] * n for _ in range(n)]
+    for i in range(n):
+        matrix[i][i] = 1.0
+    for i in range(n):
+        for j in range(i + 1, n):
+            r = pearson(varying_vecs[i], varying_vecs[j])
+            matrix[i][j] = r
+            matrix[j][i] = r
+    result["matrix"] = matrix
+
+    # Pares con cifras de solapamiento de conjuntos.
+    pairs = []
+    for i in range(n):
+        vi = varying_vecs[i]
+        for j in range(i + 1, n):
+            vj = varying_vecs[j]
+            co_missing = 0
+            either_missing = 0
+            for a, b in zip(vi, vj):
+                a_miss = a != 0.0
+                b_miss = b != 0.0
+                if a_miss and b_miss:
+                    co_missing += 1
+                if a_miss or b_miss:
+                    either_missing += 1
+            jaccard = co_missing / either_missing if either_missing > 0 else 0.0
+            pairs.append({
+                "a": varying[i],
+                "b": varying[j],
+                "corr": matrix[i][j],
+                "co_missing": co_missing,
+                "either_missing": either_missing,
+                "jaccard": jaccard,
+            })
+
+    pairs.sort(key=lambda p: abs(p["corr"]), reverse=True)
+    result["pairs"] = pairs[:top_k] if top_k is not None and top_k >= 0 else pairs
+
+    return result

python/functions/datascience/missingness_correlation_test.py

@@ -0,0 +1,115 @@
+"""Tests para missingness_correlation."""
+
+from datascience.missingness_correlation import missingness_correlation
+
+
+def test_co_ocurrencia_fuerte_corr_uno_jaccard_uno():
+    # a y b faltan EXACTAMENTE en las mismas filas -> corr 1.0, jaccard 1.0.
+    mask = {
+        "a": [1, 0, 1, 0, 1, 0],
+        "b": [1, 0, 1, 0, 1, 0],
+    }
+    out = missingness_correlation(mask)
+    assert out["columns"] == ["a", "b"]
+    assert out["n_excluded"] == 0
+    # Diagonal 1.0, off-diagonal ~1.0.
+    assert out["matrix"][0][0] == 1.0
+    assert out["matrix"][1][1] == 1.0
+    assert abs(out["matrix"][0][1] - 1.0) < 1e-9
+    assert len(out["pairs"]) == 1
+    pair = out["pairs"][0]
+    assert {pair["a"], pair["b"]} == {"a", "b"}
+    assert abs(pair["corr"] - 1.0) < 1e-9
+    assert pair["co_missing"] == 3       # filas 0,2,4
+    assert pair["either_missing"] == 3   # mismas filas
+    assert abs(pair["jaccard"] - 1.0) < 1e-9
+
+
+def test_ausencias_disjuntas_corr_negativa_jaccard_cero():
+    # a y b nunca faltan en la misma fila -> co_missing 0, jaccard 0, corr <= 0.
+    mask = {
+        "a": [1, 1, 0, 0],
+        "b": [0, 0, 1, 1],
+    }
+    out = missingness_correlation(mask)
+    assert out["columns"] == ["a", "b"]
+    pair = out["pairs"][0]
+    assert pair["co_missing"] == 0
+    assert pair["either_missing"] == 4
+    assert pair["jaccard"] == 0.0
+    # Solapamiento nulo + ausencias complementarias -> correlacion negativa.
+    assert pair["corr"] < 0.0
+    assert abs(pair["corr"] - out["matrix"][0][1]) < 1e-12
+
+
+def test_columna_sin_varianza_se_excluye():
+    # c esta siempre presente (todo 0): no aporta ausencia -> no entra ni como
+    # excluida. d esta siempre ausente (todo 1): tiene nulos pero sin varianza
+    # -> excluida y n_excluded incrementa. a y b tienen varianza.
+    mask = {
+        "a": [1, 0, 1, 0],
+        "b": [1, 0, 0, 0],
+        "c": [0, 0, 0, 0],   # siempre presente
+        "d": [1, 1, 1, 1],   # siempre ausente, constante
+    }
+    out = missingness_correlation(mask)
+    assert out["columns"] == ["a", "b"]
+    assert "d" in out["excluded_cols"]
+    assert "c" not in out["excluded_cols"]
+    assert out["n_excluded"] == 1
+    # Matriz solo de las columnas con varianza.
+    assert len(out["matrix"]) == 2
+    assert len(out["matrix"][0]) == 2
+
+
+def test_menos_de_dos_columnas_con_varianza_vacio_pero_cuenta_excluidas():
+    # Solo una columna con varianza (a) + una constante-ausente (d).
+    mask = {
+        "a": [1, 0, 1, 0],
+        "d": [1, 1, 1, 1],
+    }
+    out = missingness_correlation(mask)
+    assert out["columns"] == []
+    assert out["matrix"] == []
+    assert out["pairs"] == []
+    assert out["n_excluded"] == 1
+    assert out["excluded_cols"] == ["d"]
+
+
+def test_mask_vacio_todo_vacio():
+    out = missingness_correlation({})
+    assert out == {
+        "columns": [],
+        "matrix": [],
+        "pairs": [],
+        "n_excluded": 0,
+        "excluded_cols": [],
+    }
+
+
+def test_top_k_limita_pares():
+    # 4 columnas con varianza -> 6 pares; top_k=2 deja 2.
+    mask = {
+        "a": [1, 0, 1, 0, 0],
+        "b": [1, 0, 0, 1, 0],
+        "c": [0, 1, 1, 0, 1],
+        "d": [1, 1, 0, 0, 1],
+    }
+    out = missingness_correlation(mask, top_k=2)
+    assert len(out["columns"]) == 4
+    assert len(out["pairs"]) == 2
+    # Ordenados por |corr| desc.
+    assert abs(out["pairs"][0]["corr"]) >= abs(out["pairs"][1]["corr"])
+
+
+def test_no_lanza_con_entradas_raras():
+    # Valores no-lista y no-dict no deben romper.
+    assert missingness_correlation(None)["columns"] == []
+    mask = {
+        "a": [1, 0, 1, 0],
+        "b": [1, 0, 1, 0],
+        "bad": "not a list",
+        "empty": [],
+    }
+    out = missingness_correlation(mask)
+    assert out["columns"] == ["a", "b"]

python/functions/datascience/missingness_overview.md

@@ -0,0 +1,99 @@
+---
+id: missingness_overview_py_datascience
+name: missingness_overview
+kind: function
+lang: py
+domain: datascience
+version: "1.0.0"
+purity: pure
+signature: "def missingness_overview(null_mask) -> dict"
+description: "Resumen de ausencias a nivel de dataset a partir de una máscara de nulos 0/1 por columna ({col: [1=falta, 0=presente]} alineada por fila). Calcula celdas y porcentaje de datos faltantes, cuántas columnas tienen algún nulo y cuántas filas son completas vs. incompletas. Estilo dict-no-throw del grupo eda: nunca lanza. Lectura defensiva — no-dict o dict vacío devuelve todo a 0; columnas no-lista se tratan como vacías; listas de longitud distinta se alinean a la longitud máxima rellenando la cola corta como presente (0); valores None/no-int cuentan como presente; sin ZeroDivisionError."
+tags: [eda, missing, missingness, nulls, profiling, datascience, pure]
+uses_functions: []
+uses_types: []
+returns: []
+returns_optional: false
+error_type: ""
+imports: []
+example: |
+  from datascience.missingness_overview import missingness_overview
+  mask = {
+      "a": [1, 0, 0, 0, 1],
+      "b": [1, 0, 1, 0, 0],
+      "c": [0, 0, 0, 0, 1],
+  }
+  missingness_overview(mask)
+  # n_missing_cells=5, missing_cell_pct≈33.33, complete_rows=2, incomplete_rows=3
+tested: true
+tests:
+  - "test_cooccurrence_three_cols_exact"
+  - "test_empty_dict_all_zero"
+  - "test_output_keys_contract"
+  - "test_not_a_dict_returns_zero"
+  - "test_no_nulls_all_complete"
+  - "test_none_values_treated_as_present"
+  - "test_unequal_lengths_pad_with_max"
+  - "test_columns_present_but_no_rows"
+  - "test_never_raises_on_garbage"
+test_file_path: "python/functions/datascience/missingness_overview_test.py"
+file_path: "python/functions/datascience/missingness_overview.py"
+params:
+  - name: null_mask
+    desc: "Dict {col_name: [int 0/1, ...]} con la máscara de nulos por columna, alineada por fila (1 = el valor falta, 0 = el valor está presente). Normalmente todas las listas tienen la misma longitud = nº de filas. Lectura defensiva: si no es dict o está vacío se devuelve todo a 0; columnas cuyo valor no es lista/tupla se tratan como vacías; listas de longitud distinta se alinean a la longitud máxima (las posiciones inexistentes de las columnas más cortas cuentan como presentes, 0); valores None o no enteros cuentan como presentes."
+output: "Dict con exactamente 9 claves, todas siempre presentes (la función nunca lanza): n_rows (longitud de fila = longitud máxima entre columnas, 0 si vacío), n_cols (nº de columnas), n_cols_with_null (columnas con >=1 falta), n_missing_cells (suma total de 1s), missing_cell_pct (0-100 = n_missing_cells / (n_rows*n_cols) * 100), complete_rows (filas sin ninguna falta), incomplete_rows (filas con >=1 falta), complete_pct (0-100), incomplete_pct (0-100). Los porcentajes son 0.0 cuando el denominador es 0 (sin ZeroDivisionError)."
+---
+
+## Ejemplo
+
+```python
+from datascience.missingness_overview import missingness_overview
+
+# Máscara de nulos por columna: 1 = falta, 0 = presente, alineada por fila.
+mask = {
+    "a": [1, 0, 0, 0, 1],
+    "b": [1, 0, 1, 0, 0],
+    "c": [0, 0, 0, 0, 1],
+}
+
+missingness_overview(mask)
+# {
+#   "n_rows": 5,
+#   "n_cols": 3,
+#   "n_cols_with_null": 3,      # a, b y c tienen al menos una falta
+#   "n_missing_cells": 5,       # 2 (a) + 2 (b) + 1 (c)
+#   "missing_cell_pct": 33.33,  # 5 / (5*3) * 100
+#   "complete_rows": 2,         # filas 1 y 3 sin ninguna falta
+#   "incomplete_rows": 3,       # filas 0 (a&b), 2 (b), 4 (a&c)
+#   "complete_pct": 40.0,       # 2 / 5 * 100
+#   "incomplete_pct": 60.0,     # 3 / 5 * 100
+# }
+
+missingness_overview({})
+# Todo a 0: {"n_rows": 0, "n_cols": 0, "n_cols_with_null": 0,
+#            "n_missing_cells": 0, "missing_cell_pct": 0.0,
+#            "complete_rows": 0, "incomplete_rows": 0,
+#            "complete_pct": 0.0, "incomplete_pct": 0.0}
+```
+
+## Cuando usarla
+
+Úsala al perfilar un dataset cuando ya tienes una máscara de nulos 0/1 por
+columna (p. ej. derivada del paso de carga/perfilado del EDA) y quieres la foto
+global de ausencias en una llamada: cuánta proporción de celdas falta, cuántas
+columnas están afectadas y, sobre todo, cuántas filas quedan completas vs.
+incompletas. Es el bloque resumen del capítulo de calidad/missingness de un EDA,
+y la base para decidir estrategias de imputación o de borrado de filas. Como es
+pura y dict-no-throw, puedes alimentarla con la máscara tal cual sin validarla
+antes: entradas malformadas degradan a ceros en vez de romper el pipeline.
+
+## Gotchas
+
+- **`n_rows` es la longitud máxima entre columnas.** Con listas de longitud
+  desigual, las posiciones que faltan en las columnas más cortas se cuentan como
+  presentes (`0`); no se descartan filas. En el caso normal (todas las listas de
+  igual longitud) `n_rows` es simplemente esa longitud.
+- **Solo el valor exacto `1` cuenta como falta.** `None`, `0`, cadenas y
+  cualquier otro valor se tratan como presentes. `True` (== 1) también cuenta
+  como falta por la igualdad.
+- **Porcentajes en escala 0-100**, no fracciones. División por cero protegida:
+  con `n_rows*n_cols == 0` los porcentajes salen `0.0`.

python/functions/datascience/missingness_overview.py

@@ -0,0 +1,116 @@
+"""Pure EDA helper: dataset-level missingness overview from a 0/1 null mask.
+
+Part of the `eda` capability group. Consumes a per-column null mask
+(``{col_name: [int 0/1, ...]}`` aligned by row, ``1`` = value is missing,
+``0`` = value is present) and derives dataset-wide missingness metrics: cell
+count and percentage of missing data, how many columns carry any null, and how
+many rows are complete vs. incomplete.
+
+Dict-no-throw style of the `eda` group: it NEVER raises. A non-dict, an empty
+dict, malformed columns, ragged lists or non-int cell values all degrade
+gracefully to the zero/contract output. Stdlib only.
+
+Ragged-length policy: columns are allowed to have different lengths. ``n_rows``
+is the **maximum** column length; positions that don't exist in a shorter
+column are treated as present (``0``). This keeps the ``n_rows * n_cols`` cell
+grid well defined without dropping rows.
+"""
+
+
+def _is_missing(value) -> int:
+    """Return ``1`` iff ``value`` denotes a missing cell, else ``0``.
+
+    Only an exact equality to ``1`` (covers ``int`` ``1`` and ``float`` ``1.0``)
+    counts as missing. ``None``, ``0``, strings and any other value are treated
+    as present. The comparison cannot raise for standard inputs.
+    """
+    try:
+        return 1 if value == 1 else 0
+    except Exception:
+        return 0
+
+
+def missingness_overview(null_mask) -> dict:
+    """Summarize dataset-level missingness from a 0/1 null mask.
+
+    Args:
+        null_mask: Dict ``{col_name: [int 0/1, ...]}`` where each list is aligned
+            by row (``1`` = missing, ``0`` = present). Lists are normally all the
+            same length (= number of rows). Defensive: a non-dict or empty dict
+            returns the all-zero contract; non-list columns are treated as empty;
+            ragged lists are aligned to the maximum length, padding the missing
+            tail of shorter columns as present (``0``); ``None`` / non-int cells
+            count as present.
+
+    Returns:
+        Dict with exactly these keys, all always present (the function never
+        raises): ``n_rows``, ``n_cols``, ``n_cols_with_null``,
+        ``n_missing_cells``, ``missing_cell_pct`` (0-100), ``complete_rows``,
+        ``incomplete_rows``, ``complete_pct`` (0-100), ``incomplete_pct``
+        (0-100). Percentages are ``0.0`` when the denominator is zero (no
+        ``ZeroDivisionError``).
+    """
+    zero = {
+        "n_rows": 0,
+        "n_cols": 0,
+        "n_cols_with_null": 0,
+        "n_missing_cells": 0,
+        "missing_cell_pct": 0.0,
+        "complete_rows": 0,
+        "incomplete_rows": 0,
+        "complete_pct": 0.0,
+        "incomplete_pct": 0.0,
+    }
+
+    if not isinstance(null_mask, dict) or not null_mask:
+        return dict(zero)
+
+    # Normalize every column to a list; non-list columns become empty.
+    cols = {}
+    for name, seq in null_mask.items():
+        cols[name] = seq if isinstance(seq, (list, tuple)) else []
+
+    n_cols = len(cols)
+    lengths = [len(seq) for seq in cols.values()]
+    n_rows = max(lengths) if lengths else 0
+
+    if n_rows == 0:
+        # Columns exist but carry no rows: everything zero except n_cols.
+        out = dict(zero)
+        out["n_cols"] = n_cols
+        return out
+
+    n_missing_cells = 0
+    n_cols_with_null = 0
+    row_has_missing = [False] * n_rows
+
+    for seq in cols.values():
+        col_len = len(seq)
+        col_has_null = False
+        for r in range(n_rows):
+            if r < col_len and _is_missing(seq[r]):
+                n_missing_cells += 1
+                row_has_missing[r] = True
+                col_has_null = True
+        if col_has_null:
+            n_cols_with_null += 1
+
+    incomplete_rows = sum(1 for flag in row_has_missing if flag)
+    complete_rows = n_rows - incomplete_rows
+
+    total_cells = n_rows * n_cols
+    missing_cell_pct = (n_missing_cells / total_cells * 100.0) if total_cells else 0.0
+    complete_pct = complete_rows / n_rows * 100.0
+    incomplete_pct = incomplete_rows / n_rows * 100.0
+
+    return {
+        "n_rows": n_rows,
+        "n_cols": n_cols,
+        "n_cols_with_null": n_cols_with_null,
+        "n_missing_cells": n_missing_cells,
+        "missing_cell_pct": missing_cell_pct,
+        "complete_rows": complete_rows,
+        "incomplete_rows": incomplete_rows,
+        "complete_pct": complete_pct,
+        "incomplete_pct": incomplete_pct,
+    }