merge: 4b calidad — nueva formula (completeness 0.6+validity 0.4, dataset row_uniqueness, outliers fuera a Observaciones, sin doble conteo) report 2046 (verificado met)

python/functions/pipelines/profile_table.py

@@ -477,9 +477,18 @@ def profile_table(
                     if vals and (len(ok) / len(vals)) >= _PROMOTE_MIN_PARSE:
                         col["inferred_type"] = "numeric"
                         inferred = "numeric"
+                        # Tasa de parseo real de la muestra: alimenta la
+                        # dimension validity de column_quality_score (fraccion
+                        # de valores conformes al tipo numerico promovido).
+                        col["validity_rate"] = len(ok) / len(vals)
                 elif semantic in _DATETIME_SEMANTIC:
                     col["inferred_type"] = "datetime"
                     inferred = "datetime"
+                    # Tasa de parseo de la muestra a fecha (mismo papel que el
+                    # parse rate numerico) para la dimension validity.
+                    parsed_dt = [_to_ordinal_days(v) for v in vals]
+                    ok_dt = [d for d in parsed_dt if d is not None]
+                    col["validity_rate"] = (len(ok_dt) / len(vals)) if vals else None
 
             # 4) Enriquecer segun el inferred_type final.
             if inferred == "numeric":
@@ -506,11 +515,36 @@ def profile_table(
             # 5) Score de calidad por columna.
             col["quality_score"] = column_quality_score(col).get("score")
 
-        # 6) Score agregado de la tabla (media de columnas).
+        # 6) Score agregado de la tabla (report 2046): NO media simple.
+        #   cell_quality   = media de los scores de columna, en [0,1].
+        #   row_uniqueness = 1 - duplicate_pct (unicidad de registro).
+        #   score = 100 * (0.85*cell_quality + 0.15*row_uniqueness).
+        # Renormaliza a solo cell_quality si duplicate_pct no se pudo calcular.
         scores = [
             c["quality_score"] for c in cols if c.get("quality_score") is not None
         ]
-        prof["quality_score"] = round(sum(scores) / len(scores), 1) if scores else None
+        if scores:
+            cell_quality = (sum(scores) / len(scores)) / 100.0
+            dup_pct = prof.get("duplicate_pct")
+            if dup_pct is not None:
+                try:
+                    d = float(dup_pct)
+                except (TypeError, ValueError):
+                    d = None
+            else:
+                d = None
+            if d is not None:
+                # Tolerar escala 0-100 por si algun backend la entrega asi.
+                if d > 1.0:
+                    d = d / 100.0
+                row_uniqueness = max(0.0, min(1.0, 1.0 - d))
+                prof["quality_score"] = round(
+                    100.0 * (0.85 * cell_quality + 0.15 * row_uniqueness), 1
+                )
+            else:
+                prof["quality_score"] = round(100.0 * cell_quality, 1)
+        else:
+            prof["quality_score"] = None
 
         # 7) Candidatos a clave.
         key_candidates = []