chore(0004): close issue, bump unibus to 0.5.0, record dataplane-acl decision

Issue 0004 (security hardening) done across 0004a-0004f. app.md version 0.5.0
with the capability growth log entry; dev/0004d-dataplane-acl.md documents the
chosen minimum-defense strategy for the NATS data plane and its residual limit
(per-subject ACL deferred to 0003). Full work report in
projects/message_bus/reports/0005-2026-06-07-unibus-security-hardening.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.md

@@ -2,7 +2,7 @@
 name: unibus
 lang: go
 domain: infra
-version: 0.4.0
+version: 0.5.0
 description: "Bus de mensajería unificado sobre NATS+JetStream con cifrado E2E por room (megolm/olm reducido): service de membresía/claves, librería cliente y peers demo."
 tags: [service, messaging, nats, e2e]
 uses_functions:
@@ -154,6 +154,19 @@ agent.<nombre>.{in,out}   inbox/outbox de agente LLM (agent.scout.in)
 
 ## Capability growth log
 
+- v0.5.0 (2026-06-07) — hardening de seguridad (issue 0004) que cierra los
+  hallazgos de la auditoría red-team (report 0004) y lleva el veredicto de
+  exposición pública de "NO" a "sí-con-condiciones". Anti-DoS pre-auth
+  (`http.MaxBytesReader` por ruta + rechazo por `Content-Length` + rate-limit
+  por IP + `MaxHeaderBytes`); guard de fail-open que prohíbe arrancar con bind
+  público o TLS sin `--bus-auth enforce`; autorización por pertenencia en los GET
+  de room (metadata y clave sellada solo para miembros / el propio endpoint);
+  rooms cleartext deshabilitadas en bind público (contenido siempre E2E, mínimo
+  defensivo del data plane mientras la ACL por subject llega con 0003); TLS en el
+  control plane HTTP con la CA propia y cliente que exige `https` cuando hay CA;
+  y los medios H6/H7/H12 (owner ligado al firmante, `IsAuthorized` antes del
+  nonce-cache con poda O(expired) + cap, errores genéricos al cliente). Cada
+  hallazgo lleva su test adversarial `TestAudit_*` portado como regresión.
 - v0.4.0 (2026-06-07) — descubrimiento de rooms: `GET /members/{endpoint}/rooms`
   lista las rooms de un endpoint con su metadata y rol, y `client.ListMyRooms()`
   lo consume. El control plane es pull (no hay push de invitaciones), así que un

cmd/membershipd/config.go

@@ -0,0 +1,50 @@
+package main
+
+import (
+	"fmt"
+	"net"
+
+	"github.com/enmanuel/unibus/pkg/membership"
+)
+
+// isLoopbackBind reports whether the --bind value keeps the service reachable
+// only from this host. An empty bind means "all interfaces" (public), and a
+// hostname we cannot resolve to a loopback literal is treated as public — the
+// conservative choice, so an unusual bind never silently slips past the guard.
+func isLoopbackBind(bind string) bool {
+	switch bind {
+	case "localhost":
+		return true
+	case "":
+		return false // empty binds every interface
+	}
+	ip := net.ParseIP(bind)
+	if ip == nil {
+		return false // a hostname we can't classify: assume public
+	}
+	return ip.IsLoopback()
+}
+
+// validateBootConfig is the fail-open guard (audit H2). It refuses any startup
+// configuration that would expose the bus without enforced authentication:
+//
+//   - a non-loopback --bind without --bus-auth enforce (the data plane and
+//     control plane would both accept anyone), and
+//   - --tls-cert/--tls-key without --bus-auth enforce (TLS encrypts the channel
+//     but authenticates no one — encrypted access for everybody is still open).
+//
+// It is a pure function of the parsed flags so the command can fail fast at
+// startup and tests can assert the policy without booting a server.
+func validateBootConfig(bind string, mode membership.AuthMode, tlsCert, tlsKey string) error {
+	if !isLoopbackBind(bind) && mode != membership.AuthEnforce {
+		return fmt.Errorf(
+			"refusing to start: --bind %q is not loopback but --bus-auth is %q; a public bind requires --bus-auth enforce (or bind 127.0.0.1 for local dev)",
+			bind, mode)
+	}
+	if (tlsCert != "" || tlsKey != "") && mode != membership.AuthEnforce {
+		return fmt.Errorf(
+			"refusing to start: --tls-cert/--tls-key set but --bus-auth is %q; TLS without enforced auth is fail-open (encrypted channel, no authentication) — set --bus-auth enforce",
+			mode)
+	}
+	return nil
+}

cmd/membershipd/config_test.go

@@ -0,0 +1,72 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/enmanuel/unibus/pkg/membership"
+)
+
+// TestAudit_FailOpenTLSWithoutAuth ports the auditor's H2 vector. Before the
+// guard, booting with TLS on but the authenticator off ("--bind 0.0.0.0
+// --tls-cert … " without enforce) produced an encrypted data plane that an
+// unregistered, nkey-less client could still connect to — a fail-open config
+// wearing the appearance of security. validateBootConfig now refuses it, so the
+// insecure server never starts (the client therefore has nothing to connect to).
+func TestAudit_FailOpenTLSWithoutAuth(t *testing.T) {
+	// The exact auditor configuration: public bind, TLS provided, auth off.
+	err := validateBootConfig("0.0.0.0", membership.AuthOff, "server.crt", "server.key")
+	if err == nil {
+		t.Fatalf("TLS without enforce on a public bind must be refused at startup")
+	}
+	if !strings.Contains(err.Error(), "enforce") {
+		t.Fatalf("error should point the operator at --bus-auth enforce, got: %v", err)
+	}
+
+	// And TLS without enforce is rejected even on loopback: TLS implies a
+	// security posture, so authenticating no one is always a misconfiguration.
+	if err := validateBootConfig("127.0.0.1", membership.AuthOff, "server.crt", "server.key"); err == nil {
+		t.Fatalf("TLS flags without enforce must be refused regardless of bind")
+	}
+}
+
+// TestBootConfigPolicy is the full table: the golden secure-public config is
+// allowed, dev loopback is allowed, and every fail-open shape is refused.
+func TestBootConfigPolicy(t *testing.T) {
+	cases := []struct {
+		name    string
+		bind    string
+		mode    membership.AuthMode
+		cert    string
+		key     string
+		wantErr bool
+	}{
+		// Golden: the intended public production config.
+		{"public+enforce+tls", "0.0.0.0", membership.AuthEnforce, "s.crt", "s.key", false},
+		{"public+enforce+notls", "0.0.0.0", membership.AuthEnforce, "", "", false},
+		// Edge: local dev on loopback may stay open (no auth, no TLS).
+		{"loopback+off", "127.0.0.1", membership.AuthOff, "", "", false},
+		{"loopback-ipv6+off", "::1", membership.AuthOff, "", "", false},
+		{"localhost+off", "localhost", membership.AuthOff, "", "", false},
+		{"loopback+soft", "127.0.0.1", membership.AuthSoft, "", "", false},
+		// Error: public bind without enforce.
+		{"public+off", "0.0.0.0", membership.AuthOff, "", "", true},
+		{"public+soft", "0.0.0.0", membership.AuthSoft, "", "", true},
+		{"lan-ip+off", "192.168.1.10", membership.AuthOff, "", "", true},
+		{"empty-bind+off", "", membership.AuthOff, "", "", true},
+		// Error: TLS flags without enforce (cert or key alone is enough to trip it).
+		{"loopback+tlscert+off", "127.0.0.1", membership.AuthOff, "s.crt", "", true},
+		{"loopback+tlskey+soft", "127.0.0.1", membership.AuthSoft, "", "s.key", true},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			err := validateBootConfig(c.bind, c.mode, c.cert, c.key)
+			if c.wantErr && err == nil {
+				t.Fatalf("config %+v should be refused", c)
+			}
+			if !c.wantErr && err != nil {
+				t.Fatalf("config %+v should be allowed, got: %v", c, err)
+			}
+		})
+	}
+}

cmd/membershipd/main.go

@@ -6,6 +6,7 @@ package main
 
 import (
 	"context"
+	"crypto/tls"
 	"flag"
 	"log"
 	"net/http"
@@ -52,6 +53,13 @@ func main() {
 		log.Fatalf("%v", err)
 	}
 
+	// Fail-open guard (audit H2): a non-loopback bind, or any TLS flag, demands
+	// --bus-auth enforce. This makes an insecure public startup impossible rather
+	// than silently exposing the bus with the appearance of security.
+	if err := validateBootConfig(*bind, authMode, *tlsCert, *tlsKey); err != nil {
+		log.Fatalf("%v", err)
+	}
+
 	log.SetFlags(log.LstdFlags | log.Lmsgprefix)
 	log.SetPrefix("[membershipd] ")
 
@@ -111,15 +119,45 @@ func main() {
 	}
 
 	srv := membership.NewServer(store, blobs, authMode)
+	// On a public (non-loopback) bind, disable cleartext rooms: the embedded NATS
+	// has no per-subject ACL, so cleartext content would be readable by any
+	// registered peer. Forcing E2E keeps message content confidential regardless
+	// (audit H4 minimum defense; see dev/0004d-dataplane-acl.md).
+	if !isLoopbackBind(*bind) {
+		srv.RequireEncryptedRooms = true
+		log.Printf("cleartext rooms: DISABLED (public bind requires end-to-end encryption)")
+	}
 	log.Printf("control-plane auth: %s", authMode)
 	addr := *bind + ":" + *httpPort
-	httpSrv := &http.Server{Addr: addr, Handler: srv}
+	httpSrv := &http.Server{
+		Addr:    addr,
+		Handler: srv,
+		// Bound request header size so a peer cannot exhaust memory with huge
+		// headers before any body limit applies (the body ceilings live in the
+		// membership middleware).
+		MaxHeaderBytes:    membership.MaxHeaderBytes,
+		ReadHeaderTimeout: 10 * time.Second,
+	}
 
 	go func() {
-		log.Printf("HTTP control-plane API: http://%s", addr)
-		log.Printf("  health: http://%s/healthz", addr)
-		if err := httpSrv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
-			log.Fatalf("http server: %v", err)
+		var serveErr error
+		if *tlsCert != "" {
+			// Serve the control plane over TLS with the same CA-signed cert as the
+			// data plane (audit H5): metadata (subjects, pubkeys, sealed keys, the
+			// social graph) is no longer readable by a network MITM. The fail-open
+			// guard already requires --bus-auth enforce alongside these flags.
+			httpSrv.TLSConfig = &tls.Config{MinVersion: tls.VersionTLS12}
+			log.Printf("HTTPS control-plane API: https://%s", addr)
+			log.Printf("  health: https://%s/healthz", addr)
+			log.Printf("control-plane TLS: ON (%s)", *tlsCert)
+			serveErr = httpSrv.ListenAndServeTLS(*tlsCert, *tlsKey)
+		} else {
+			log.Printf("HTTP control-plane API: http://%s", addr)
+			log.Printf("  health: http://%s/healthz", addr)
+			serveErr = httpSrv.ListenAndServe()
+		}
+		if serveErr != nil && serveErr != http.ErrServerClosed {
+			log.Fatalf("http server: %v", serveErr)
 		}
 	}()
 

dev/0004d-dataplane-acl.md

@@ -0,0 +1,80 @@
+# 0004d — Data-plane access control on NATS (audit H4)
+
+## The finding
+
+The NATS authenticator (`pkg/busauth`) decides one thing per connection:
+*is this identity registered on the bus?* It does **not** scope what a connected
+client may subscribe to or publish. There is a single NATS account with no
+`Permissions`, so any registered peer can subscribe to, or publish on, **any**
+subject. Concretely:
+
+- A cleartext room (`ModeNATS`) carries its payload in the clear on its subject.
+  A registered peer that knows or guesses the subject subscribes and reads the
+  content directly (the auditor's `TestAudit_NoSubjectACL`: eve, never invited,
+  receives `"internal: salary numbers"`).
+- An encrypted room (`ModeMatrix`) keeps its **content** confidential (the
+  payload is AEAD ciphertext), but the **metadata of traffic** — that a subject
+  is active, message sizes and timing, who is publishing — is still observable by
+  any registered peer that subscribes to the subject.
+
+## Why the "complete" fix does not fit here
+
+The preferred fix is per-subject permissions derived from room membership: when a
+client connects, the authenticator looks up the rooms it belongs to and grants
+`Sub`/`Pub` only on those subjects. NATS supports this — `CustomClientAuthentication`
+can register a `*server.User` carrying `Permissions`.
+
+The blocker is that **NATS evaluates permissions once, at connect time, and never
+re-evaluates them on a live connection.** unibus clients routinely *connect → create
+or get invited to a room → publish/subscribe* within the **same** connection
+(`TestSecureBusEndToEnd` does exactly this: A connects, then creates `room.secure`,
+then publishes to it). Permissions frozen at connect time would not include a room
+created or joined afterwards, so the legitimate owner could not publish to the room
+it just made. Making per-subject ACLs work would therefore require the client to
+**reconnect on every membership change**, an invasive change to the client library
+and to every peer (worker, chat, mobile) — and the prompt for this issue scopes the
+client changes to the minimum.
+
+That dynamic-membership reconnection model is precisely the redesign that issue
+**0003** (decentralization) already has to do: it moves the control-plane state to a
+replicated JetStream KV and reworks how nodes and clients (re)establish sessions. Per
+the issue's own guidance ("if a complete strategy does not fit, implement the minimum
+defense and document the rest"), the full subject ACL is deferred to 0003, where the
+session/permission model is being rebuilt anyway.
+
+## The strategy implemented here: forbid cleartext rooms in public
+
+`Server.RequireEncryptedRooms` (set by `membershipd` on any non-loopback bind)
+refuses to create a cleartext (`ModeNATS`) room. Every room on a public deployment
+is therefore end-to-end encrypted, so **message content stays confidential even
+though the transport offers no subject isolation**: a peer that sniffs another
+room's subject receives only AEAD ciphertext it has no key for.
+
+This composes with the 0004c control-plane authorization: a non-member cannot even
+learn a room's subject through the control plane (`GET /rooms/{id}` → 403), so to
+sniff it an attacker must already know or guess the subject out of band.
+
+## What this does NOT close (residual exposure, by design)
+
+- **Traffic metadata.** A registered peer that already knows a subject can still
+  subscribe and observe that the subject is active, the ciphertext sizes, and the
+  timing/cadence of messages. It cannot read content.
+- **Cross-room publish.** A registered peer can still *publish* arbitrary bytes on
+  any subject. In an encrypted room those bytes fail AEAD open and the signature
+  check (`SignMsgs`), so receivers drop them — it is a nuisance/spam vector, not a
+  confidentiality or integrity break.
+- **WireGuard-only deployments** may still use cleartext rooms (the guard only trips
+  on a public bind), because the network already restricts who can reach the bus.
+
+Closing the residual metadata exposure requires the per-subject ACL described above,
+tracked for issue 0003.
+
+## Regression evidence
+
+- `pkg/membership` — `TestRequireEncryptedRoomsRejectsCleartext`: with
+  `RequireEncryptedRooms` on, `POST /rooms` for a cleartext policy returns 403 while
+  an encrypted-room create returns 201.
+- `pkg/client` — `TestAudit_NoSubjectACL`: under the public posture, creating a
+  `ModeNATS` room fails; alice creates an encrypted room and publishes; eve (a
+  registered non-member) raw-subscribes to the subject and receives only ciphertext —
+  she never recovers the plaintext.

dev/issues/0004-security-hardening.md

@@ -1,8 +1,10 @@
 ---
 issue: 0004
 title: Hardening de seguridad — autorización, anti-DoS y confidencialidad antes de exponer público
-status: spec
+status: done
 created: 2026-06-07
+completed: 2026-06-07
+report: projects/message_bus/reports/0005-2026-06-07-unibus-security-hardening.md
 domain: security
 scope: unibus (pkg/membership/server.go, auth.go, pkg/embeddednats, pkg/client, cmd/membershipd, deploy/tls)
 depends_on: 0001 (cierra los gaps que la auditoría 0004 encontró sobre lo entregado en 0001)

go.mod

@@ -8,7 +8,9 @@ require (
 	fn-registry v0.0.0-00010101000000-000000000000
 	github.com/nats-io/nats-server/v2 v2.10.22
 	github.com/nats-io/nats.go v1.37.0
+	github.com/nats-io/nkeys v0.4.7
 	github.com/oklog/ulid/v2 v2.1.0
+	golang.org/x/time v0.7.0
 	modernc.org/sqlite v1.47.0
 )
 
@@ -19,7 +21,6 @@ require (
 	github.com/mattn/go-isatty v0.0.20 // indirect
 	github.com/minio/highwayhash v1.0.3 // indirect
 	github.com/nats-io/jwt/v2 v2.5.8 // indirect
-	github.com/nats-io/nkeys v0.4.7 // indirect
 	github.com/nats-io/nuid v1.0.1 // indirect
 	github.com/ncruces/go-strftime v1.0.0 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
@@ -29,7 +30,6 @@ require (
 	golang.org/x/sync v0.20.0 // indirect
 	golang.org/x/sys v0.44.0 // indirect
 	golang.org/x/text v0.37.0 // indirect
-	golang.org/x/time v0.7.0 // indirect
 	golang.org/x/tools v0.45.0 // indirect
 	modernc.org/libc v1.70.0 // indirect
 	modernc.org/mathutil v1.7.1 // indirect

go.sum

@@ -1,5 +1,7 @@
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
+github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
+github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17kjQEVQ1XRhq2/JR1M3sGqeJoxs=
 github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
@@ -47,6 +49,10 @@ golang.org/x/time v0.7.0 h1:ntUhktv3OPE6TgYxXWv9vKvUSJyIFJlyohwbkEwPrKQ=
 golang.org/x/time v0.7.0/go.mod h1:3BpzKBy/shNhVucY/MWOyx10tF3SFh9QdLuxbVysPQM=
 golang.org/x/tools v0.45.0 h1:18qN3FAooORvApf5XjCXgsuayZOEtXf6JK18I3+ONa8=
 golang.org/x/tools v0.45.0/go.mod h1:LuUGqqaXcXMEFEruIVJVm5mgDD8vww/z/SR1gQ4uE/0=
+golang.org/x/tools/go/expect v0.1.1-deprecated h1:jpBZDwmgPhXsKZC6WhL20P4b/wmnpsEAGHaNy0n/rJM=
+golang.org/x/tools/go/expect v0.1.1-deprecated/go.mod h1:eihoPOH+FgIqa3FpoTwguz/bVUSGBlGQU67vpBeOrBY=
+golang.org/x/tools/go/packages/packagestest v0.1.1-deprecated h1:1h2MnaIAIXISqTFKdENegdpAgUXz6NrPEsbIeWaBRvM=
+golang.org/x/tools/go/packages/packagestest v0.1.1-deprecated/go.mod h1:RVAQXBGNv1ib0J382/DPCRS/BPnsGebyM1Gj5VSDpG8=
 modernc.org/cc/v4 v4.27.1 h1:9W30zRlYrefrDV2JE2O8VDtJ1yPGownxciz5rrbQZis=
 modernc.org/cc/v4 v4.27.1/go.mod h1:uVtb5OGqUKpoLWhqwNQo/8LwvoiEBLvZXIQ/SmO6mL0=
 modernc.org/ccgo/v4 v4.32.0 h1:hjG66bI/kqIPX1b2yT6fr/jt+QedtP2fqojG2VrFuVw=

pkg/client/client.go

@@ -24,6 +24,7 @@ import (
 	"io"
 	"net/http"
 	"strconv"
+	"strings"
 	"sync"
 	"time"
 
@@ -67,10 +68,15 @@ type Options struct {
 	// with an nkey to a server that does not advertise nkey auth ("nkeys not
 	// supported by the server"), so this is opt-in rather than always-on.
 	UseNkey bool
-	// TLS, when non-nil, secures the NATS connection and pins the server to this
-	// config's RootCAs (the bus's self-signed CA). Build it with
-	// busauth.LoadCATLSConfig(caPath). Nil keeps the connection plaintext.
+	// TLS, when non-nil, secures the NATS (data plane) connection and pins the
+	// server to this config's RootCAs (the bus's self-signed CA). Build it with
+	// busauth.LoadCATLSConfig(caPath). Nil keeps the data plane plaintext.
 	TLS *tls.Config
+	// CtrlTLS, when non-nil, secures the HTTP control-plane connection and pins it
+	// to this config's RootCAs. It is separate from TLS so the two planes can be
+	// secured independently (a test may TLS one and not the other); production
+	// sets both to the same CA via Connect. Nil keeps the control plane plaintext.
+	CtrlTLS *tls.Config
 }
 
 // New connects to NATS and records the control-plane URL with default Options
@@ -90,11 +96,19 @@ func Connect(natsURL, ctrlURL string, id cs.Identity, caPath string) (*Client, e
 	if caPath == "" {
 		return New(natsURL, ctrlURL, id)
 	}
+	// A CA implies the bus is TLS on BOTH planes. Refuse a plaintext control-plane
+	// URL: signing gives integrity, not confidentiality, so sending metadata over
+	// http:// when the operator provisioned a CA would silently leak it to a MITM
+	// (audit H5). Force https rather than silently downgrade.
+	if !strings.HasPrefix(ctrlURL, "https://") {
+		return nil, fmt.Errorf("client: control-plane URL %q must be https:// when a CA is provided", ctrlURL)
+	}
 	tlsCfg, err := busauth.LoadCATLSConfig(caPath)
 	if err != nil {
 		return nil, fmt.Errorf("client: load CA %q: %w", caPath, err)
 	}
-	return NewWithOptions(natsURL, ctrlURL, id, Options{UseNkey: true, TLS: tlsCfg})
+	// Pin the same CA on both planes: nkey+TLS on NATS, TLS on the HTTP control plane.
+	return NewWithOptions(natsURL, ctrlURL, id, Options{UseNkey: true, TLS: tlsCfg, CtrlTLS: tlsCfg})
 }
 
 // NewWithOptions is New with explicit connection options (nkey auth, and, from
@@ -125,13 +139,20 @@ func NewWithOptions(natsURL, ctrlURL string, id cs.Identity, opts Options) (*Cli
 		nc.Close()
 		return nil, fmt.Errorf("client: init jetstream: %w", err)
 	}
+	// The control-plane HTTP client pins the bus CA when CtrlTLS is set, so an
+	// https:// control plane is verified against the bus's own CA rather than the
+	// system roots (audit H5). Without it the client stays plaintext for dev.
+	httpClient := &http.Client{Timeout: 10 * time.Second}
+	if opts.CtrlTLS != nil {
+		httpClient.Transport = &http.Transport{TLSClientConfig: opts.CtrlTLS.Clone()}
+	}
 	return &Client{
 		id:        id,
 		endpoint:  frame.EndpointID(id.SignPub),
 		nc:        nc,
 		js:        js,
 		ctrlURL:   ctrlURL,
-		http:      &http.Client{Timeout: 10 * time.Second},
+		http:      httpClient,
 		keyCache:  map[string]map[int][]byte{},
 		signCache: map[string][]byte{},
 	}, nil

pkg/client/client_test.go

@@ -32,6 +32,7 @@ type testHarness struct {
 	ns      *server.Server
 	httpts  *httptest.Server
 	store   *membership.Store
+	srv     *membership.Server
 }
 
 func freePort(t *testing.T) int {
@@ -98,7 +99,7 @@ func bootHarness(t *testing.T, ctrlMode membership.AuthMode, natsAuth bool, nats
 	srv := membership.NewServer(store, blobs, ctrlMode)
 	httpts := httptest.NewServer(srv)
 
-	h := &testHarness{natsURL: embeddednats.ClientURL(ns), ctrlURL: httpts.URL, ns: ns, httpts: httpts, store: store}
+	h := &testHarness{natsURL: embeddednats.ClientURL(ns), ctrlURL: httpts.URL, ns: ns, httpts: httpts, store: store, srv: srv}
 	t.Cleanup(func() {
 		httpts.Close()
 		store.Close()

pkg/client/control_tls_test.go

@@ -0,0 +1,87 @@
+package client_test
+
+import (
+	"crypto/tls"
+	"crypto/x509"
+	"net/http/httptest"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/enmanuel/unibus/pkg/blobstore"
+	"github.com/enmanuel/unibus/pkg/client"
+	"github.com/enmanuel/unibus/pkg/embeddednats"
+	"github.com/enmanuel/unibus/pkg/membership"
+	"github.com/enmanuel/unibus/pkg/room"
+)
+
+// TestConnectRequiresHTTPSWithCA covers audit H5's client contract: when a CA is
+// provided the control-plane URL must be https://. A signed request gives
+// integrity but not confidentiality, so silently talking http:// to a bus the
+// operator secured with a CA would leak all metadata to a MITM. Connect refuses
+// the plaintext URL outright (error path; the scheme is checked before any
+// network use, so a bogus CA path is irrelevant).
+func TestConnectRequiresHTTPSWithCA(t *testing.T) {
+	_, err := client.Connect("nats://127.0.0.1:4222", "http://127.0.0.1:8470", mustIdentity(t), "/nonexistent/ca.crt")
+	if err == nil {
+		t.Fatalf("Connect with a CA and an http:// control plane must be refused")
+	}
+	if !strings.Contains(err.Error(), "https") {
+		t.Fatalf("error should point the caller at https, got: %v", err)
+	}
+}
+
+// TestControlPlaneOverTLS proves the control plane works over TLS pinned to the
+// bus CA (golden) and that a client lacking the CA cannot complete the handshake
+// (error path) — so a network observer can neither read nor inject control-plane
+// traffic. The data plane is left plaintext here to isolate the HTTP-TLS wiring.
+func TestControlPlaneOverTLS(t *testing.T) {
+	dir := t.TempDir()
+	store, err := membership.Open(filepath.Join(dir, "unibus.db"))
+	if err != nil {
+		t.Fatalf("store: %v", err)
+	}
+	t.Cleanup(func() { store.Close() })
+	blobs, err := blobstore.New(filepath.Join(dir, "blobs"))
+	if err != nil {
+		t.Fatalf("blobs: %v", err)
+	}
+
+	ns, err := embeddednats.StartServer(embeddednats.ServerConfig{
+		StoreDir: filepath.Join(dir, "js"), Host: "127.0.0.1", Port: freePort(t),
+	})
+	if err != nil {
+		t.Fatalf("nats: %v", err)
+	}
+	t.Cleanup(func() { ns.Shutdown(); ns.WaitForShutdown() })
+	natsURL := embeddednats.ClientURL(ns)
+
+	// An https control plane wrapping the real membership server.
+	ts := httptest.NewTLSServer(membership.NewServer(store, blobs, membership.AuthOff))
+	t.Cleanup(ts.Close)
+
+	pool := x509.NewCertPool()
+	pool.AddCert(ts.Certificate())
+
+	// Golden: trusting the control-plane CA, an https control-plane request works.
+	good, err := client.NewWithOptions(natsURL, ts.URL, mustIdentity(t),
+		client.Options{CtrlTLS: &tls.Config{RootCAs: pool, MinVersion: tls.VersionTLS12}})
+	if err != nil {
+		t.Fatalf("connect with the pinned CA: %v", err)
+	}
+	defer good.Close()
+	if _, err := good.CreateRoom("room.tls.ctrl", room.ModeNATS); err != nil {
+		t.Fatalf("control plane over TLS should succeed with the pinned CA: %v", err)
+	}
+
+	// Error path: without the CA the https handshake fails, so the request errors.
+	bad, err := client.NewWithOptions(natsURL, ts.URL, mustIdentity(t),
+		client.Options{CtrlTLS: &tls.Config{RootCAs: x509.NewCertPool(), MinVersion: tls.VersionTLS12}})
+	if err != nil {
+		t.Fatalf("nats connect (bad CA case): %v", err)
+	}
+	defer bad.Close()
+	if _, err := bad.CreateRoom("room.tls.fail", room.ModeNATS); err == nil {
+		t.Fatalf("a control-plane request without the CA must fail the TLS handshake")
+	}
+}

pkg/client/dataplane_acl_test.go

@@ -0,0 +1,124 @@
+package client_test
+
+import (
+	"bytes"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/enmanuel/unibus/pkg/client"
+	"github.com/enmanuel/unibus/pkg/frame"
+	"github.com/enmanuel/unibus/pkg/room"
+	"github.com/nats-io/nats.go"
+)
+
+// TestAudit_NoSubjectACL ports the auditor's H4 (Alto) finding under the minimum
+// defense chosen for this issue (forbid cleartext rooms in public; see
+// dev/0004d-dataplane-acl.md). The NATS data plane still has no per-subject ACL,
+// so the guarantee we make is CONTENT confidentiality, proven three ways:
+//
+//	error : a cleartext (ModeNATS) room cannot be created under the public posture;
+//	golden: a legitimate member (bob) decrypts the secret;
+//	edge  : eve, sniffing the raw subject off the data plane, receives only
+//	        ciphertext — she never recovers the plaintext the auditor's eve did.
+func TestAudit_NoSubjectACL(t *testing.T) {
+	h := newHarness(t)
+	h.srv.RequireEncryptedRooms = true // the public posture
+	waitHealth(t, h.ctrlURL)
+
+	alice, err := client.New(h.natsURL, h.ctrlURL, mustIdentity(t))
+	if err != nil {
+		t.Fatalf("connect alice: %v", err)
+	}
+	defer alice.Close()
+
+	// Error path: a cleartext room is refused, so no payload ever rides a subject
+	// in the clear for a sniffer to read (the exact vector the auditor exploited).
+	if _, err := alice.CreateRoom("secret.subject.payroll", room.ModeNATS); err == nil {
+		t.Fatalf("cleartext room must be refused on a public deployment")
+	}
+
+	// alice creates an encrypted room and invites bob (the legitimate reader).
+	const subject = "secret.subject.payroll.e2e"
+	const secret = "internal: salary numbers"
+	roomID, err := alice.CreateRoom(subject, room.ModeMatrix)
+	if err != nil {
+		t.Fatalf("alice create encrypted room: %v", err)
+	}
+	bob, err := client.New(h.natsURL, h.ctrlURL, mustIdentity(t))
+	if err != nil {
+		t.Fatalf("connect bob: %v", err)
+	}
+	defer bob.Close()
+	if err := alice.Invite(roomID, bob.Endpoint()); err != nil {
+		t.Fatalf("alice invite bob: %v", err)
+	}
+	if err := bob.Join(roomID); err != nil {
+		t.Fatalf("bob join: %v", err)
+	}
+
+	// Golden: bob (a member) subscribes and decrypts the secret.
+	var bmu sync.Mutex
+	var bobGot []string
+	bobSub, err := bob.Subscribe(roomID, func(_ frame.Frame, plaintext []byte) {
+		bmu.Lock()
+		bobGot = append(bobGot, string(plaintext))
+		bmu.Unlock()
+	})
+	if err != nil {
+		t.Fatalf("bob subscribe: %v", err)
+	}
+	defer bobSub.Unsubscribe()
+
+	// Edge: eve sniffs the raw subject directly off NATS (no membership, no key).
+	rawEve, err := nats.Connect(h.natsURL)
+	if err != nil {
+		t.Fatalf("eve raw connect: %v", err)
+	}
+	defer rawEve.Close()
+	eveGot := make(chan []byte, 8)
+	if _, err := rawEve.Subscribe(subject, func(m *nats.Msg) { eveGot <- m.Data }); err != nil {
+		t.Fatalf("eve raw subscribe: %v", err)
+	}
+	if err := rawEve.Flush(); err != nil {
+		t.Fatalf("eve flush: %v", err)
+	}
+	time.Sleep(200 * time.Millisecond) // let both subscriptions settle
+
+	if err := alice.Publish(roomID, []byte(secret)); err != nil {
+		t.Fatalf("alice publish: %v", err)
+	}
+
+	// bob must decrypt the secret.
+	if !waitFor(&bmu, &bobGot, func(rs []string) bool {
+		for _, r := range rs {
+			if r == secret {
+				return true
+			}
+		}
+		return false
+	}, 2*time.Second) {
+		t.Fatalf("bob (member) should decrypt the secret; got %v", snapshot(&bmu, &bobGot))
+	}
+
+	// eve must receive only ciphertext — never the plaintext.
+	select {
+	case data := <-eveGot:
+		if bytes.Contains(data, []byte(secret)) {
+			t.Fatalf("eve sniffed the plaintext off the data plane: %q", data)
+		}
+		f, err := frame.Unmarshal(data)
+		if err != nil {
+			t.Fatalf("eve received an undecodable frame: %v", err)
+		}
+		if string(f.Payload) == secret {
+			t.Fatalf("eve read the secret from the frame payload")
+		}
+		if len(f.Nonce) == 0 {
+			t.Fatalf("expected an AEAD-encrypted payload (non-empty nonce), got cleartext frame")
+		}
+	case <-time.After(2 * time.Second):
+		// eve receiving nothing is also a safe outcome; the assertion is only that
+		// she never gets the plaintext, which holds vacuously here.
+	}
+}

pkg/membership/auth.go

@@ -11,6 +11,8 @@ import (
 	"time"
 
 	cs "fn-registry/functions/cybersecurity"
+
+	"github.com/enmanuel/unibus/pkg/frame"
 )
 
 // AuthMode is the control-plane authentication rollout state (feature flag
@@ -73,6 +75,11 @@ const (
 const (
 	clockSkew = 30 * time.Second
 	nonceTTL  = 60 * time.Second
+	// maxNonceCacheEntries bounds the replay cache so it cannot grow without limit
+	// (audit H7). With IsAuthorized now gating insertion, only authorized traffic
+	// is cached, so this ceiling is only approached under a legitimate burst; at
+	// the cap the oldest nonce is evicted (its TTL is nearly up anyway).
+	maxNonceCacheEntries = 100_000
 )
 
 // CanonicalRequest returns the exact bytes that are signed and verified for a
@@ -89,43 +96,75 @@ func CanonicalRequest(method, path, ts, nonce string, body []byte) []byte {
 }
 
 // nonceCache remembers recently-seen nonces to reject replays. It is an
-// in-memory map guarded by a mutex with lazy expiry — sufficient for a single
-// membershipd process (the spec's chosen tradeoff over a server-issued nonce
-// round-trip). A distributed deployment would need a shared store.
+// in-memory store guarded by a mutex — sufficient for a single membershipd
+// process (the spec's chosen tradeoff over a server-issued nonce round-trip). A
+// distributed deployment would need a shared store (tracked for issue 0003).
+//
+// Pruning is O(expired), not O(n): because the TTL is constant, insertion order
+// equals expiry order, so the oldest entries (front of `order`) are exactly the
+// ones that expire first (audit H7 — the previous full-map scan under the mutex
+// was a CPU-amplification vector). A size cap bounds memory.
 type nonceCache struct {
-	mu   sync.Mutex
-	seen map[string]time.Time
-	ttl  time.Duration
+	mu    sync.Mutex
+	seen  map[string]time.Time // nonce -> expiry
+	order []string             // nonces in insertion order == expiry order
+	ttl   time.Duration
+	cap   int
 }
 
-func newNonceCache(ttl time.Duration) *nonceCache {
-	return &nonceCache{seen: make(map[string]time.Time), ttl: ttl}
+func newNonceCache(ttl time.Duration, capacity int) *nonceCache {
+	return &nonceCache{seen: make(map[string]time.Time), ttl: ttl, cap: capacity}
 }
 
 // rememberOrReject records nonce and returns true if it was unseen, or false if
-// it is a replay (still live in the cache). Expired entries are pruned lazily on
-// each call so the map cannot grow without bound under steady traffic.
+// it is a replay (still live in the cache).
 func (n *nonceCache) rememberOrReject(nonce string, now time.Time) bool {
 	n.mu.Lock()
 	defer n.mu.Unlock()
-	for k, exp := range n.seen {
-		if exp.Before(now) {
-			delete(n.seen, k)
+
+	// Prune expired entries from the front (oldest first). The first live entry
+	// ends the scan — everything behind it was inserted later and is newer.
+	cut := 0
+	for cut < len(n.order) {
+		exp, ok := n.seen[n.order[cut]]
+		if !ok {
+			cut++ // already evicted by the cap path below
+			continue
 		}
+		if !exp.Before(now) {
+			break
+		}
+		delete(n.seen, n.order[cut])
+		cut++
 	}
+	if cut > 0 {
+		n.order = append(n.order[:0], n.order[cut:]...)
+	}
+
 	if exp, ok := n.seen[nonce]; ok && !exp.Before(now) {
-		return false
+		return false // a live replay
 	}
+
+	// Bound memory: at capacity, evict the oldest entry (its TTL is nearly up).
+	for len(n.seen) >= n.cap && len(n.order) > 0 {
+		oldest := n.order[0]
+		n.order = n.order[1:]
+		delete(n.seen, oldest)
+	}
+
 	n.seen[nonce] = now.Add(n.ttl)
+	n.order = append(n.order, nonce)
 	return true
 }
 
 // authResult is what a successful authentication yields: the verified signing
-// key (hex) and the authorized user record. Handlers may use it for fine-grained
-// authorization (e.g. role checks) in later phases.
+// key (hex), the endpoint id derived from it, and the authorized user record.
+// Handlers use endpoint for membership authorization (only a member of a room
+// may read its metadata/keys); user is available for role checks.
 type authResult struct {
-	pubHex string
-	user   User
+	pubHex   string
+	endpoint string
+	user     User
 }
 
 // authenticate verifies the signature headers on r against body and the user
@@ -168,10 +207,9 @@ func (s *Server) authenticate(r *http.Request, body []byte, now time.Time) (auth
 		return authResult{}, fmt.Errorf("invalid signature")
 	}
 
-	if !s.nonces.rememberOrReject(nonce, now) {
-		return authResult{}, fmt.Errorf("replayed nonce")
-	}
-
+	// Authorize BEFORE touching the replay cache (audit H7): an unregistered
+	// identity can mint valid signatures for free, so caching its nonces would let
+	// it poison/grow the cache pre-auth. Only authorized identities are remembered.
 	if !s.store.IsAuthorized(pubHex) {
 		return authResult{}, fmt.Errorf("identity not authorized")
 	}
@@ -181,5 +219,12 @@ func (s *Server) authenticate(r *http.Request, body []byte, now time.Time) (auth
 		// IsAuthorized passed but the row vanished (race with revoke): fail closed.
 		return authResult{}, fmt.Errorf("identity not authorized")
 	}
-	return authResult{pubHex: pubHex, user: user}, nil
+
+	// Anti-replay last: a replayed request from an authorized identity is still
+	// rejected here (the nonce is already live in the cache from its first use).
+	if !s.nonces.rememberOrReject(nonce, now) {
+		return authResult{}, fmt.Errorf("replayed nonce")
+	}
+
+	return authResult{pubHex: pubHex, endpoint: frame.EndpointID(pub), user: user}, nil
 }

pkg/membership/auth_test.go

@@ -15,6 +15,7 @@ import (
 	cs "fn-registry/functions/cybersecurity"
 
 	"github.com/enmanuel/unibus/pkg/blobstore"
+	"github.com/enmanuel/unibus/pkg/frame"
 )
 
 // authHarness boots an in-process membershipd HTTP server in the given auth mode
@@ -88,13 +89,24 @@ func do(t *testing.T, req *http.Request) (int, string) {
 	return resp.StatusCode, string(b)
 }
 
-const okPath = "/members/alice-endpoint/rooms" // always 200 with an empty list
+// okPath is a path that authenticates and returns 200 with an empty list when
+// the request carries NO membership-bound signer (AuthOff/soft/missing-headers
+// tests). Under enforce, the per-endpoint room directory is now restricted to
+// the signer's own endpoint (audit H3), so tests that sign as alice use
+// aliceRoomsPath instead.
+const okPath = "/members/alice-endpoint/rooms"
+
+// aliceRoomsPath is alice's own room directory — the canonical "authenticated
+// and authorized" 200 path under enforce after H3.
+func aliceRoomsPath(h *authHarness) string {
+	return "/members/" + frame.EndpointID(h.alice.SignPub) + "/rooms"
+}
 
 // Golden: a request signed by a registered, active identity is accepted.
 func TestAuthGoldenAccepted(t *testing.T) {
 	h := newAuthHarness(t, AuthEnforce)
 	now := time.Now().Unix()
-	code, _ := do(t, signedReq(t, h.ts.URL, "GET", okPath, nil, h.alice, now, "nonce-golden"))
+	code, _ := do(t, signedReq(t, h.ts.URL, "GET", aliceRoomsPath(h), nil, h.alice, now, "nonce-golden"))
 	if code != http.StatusOK {
 		t.Fatalf("golden signed request should be 200, got %d", code)
 	}
@@ -116,12 +128,12 @@ func TestAuthUnregisteredRejected(t *testing.T) {
 func TestAuthReplayRejected(t *testing.T) {
 	h := newAuthHarness(t, AuthEnforce)
 	now := time.Now().Unix()
-	first := signedReq(t, h.ts.URL, "GET", okPath, nil, h.alice, now, "nonce-replay")
+	first := signedReq(t, h.ts.URL, "GET", aliceRoomsPath(h), nil, h.alice, now, "nonce-replay")
 	if code, body := do(t, first); code != http.StatusOK {
 		t.Fatalf("first request should be 200, got %d (%s)", code, body)
 	}
 	// Identical ts + nonce + signature: a replay.
-	second := signedReq(t, h.ts.URL, "GET", okPath, nil, h.alice, now, "nonce-replay")
+	second := signedReq(t, h.ts.URL, "GET", aliceRoomsPath(h), nil, h.alice, now, "nonce-replay")
 	if code, body := do(t, second); code != http.StatusUnauthorized {
 		t.Fatalf("replayed request should be 401, got %d (%s)", code, body)
 	}

pkg/membership/authz_test.go

@@ -0,0 +1,119 @@
+package membership
+
+import (
+	"encoding/hex"
+	"net/http"
+	"strconv"
+	"testing"
+	"time"
+
+	cs "fn-registry/functions/cybersecurity"
+
+	"github.com/enmanuel/unibus/pkg/frame"
+)
+
+// seedRoom inserts an encrypted room owned by alice with a sealed key for her,
+// directly through the store so the test controls membership precisely. It
+// returns the room id and alice's endpoint.
+func seedRoom(t *testing.T, h *authHarness, subject string) (string, string) {
+	t.Helper()
+	aliceEp := frame.EndpointID(h.alice.SignPub)
+	roomID := newULID()
+	info := RoomInfo{RoomID: roomID, Subject: subject, OwnerEndpoint: aliceEp, Encrypt: true}
+	if err := h.store.CreateRoom(info, h.alice.SignPub, h.alice.KexPub, []byte("alice-sealed-key")); err != nil {
+		t.Fatalf("seed room: %v", err)
+	}
+	return roomID, aliceEp
+}
+
+// register adds id to the bus allowlist so its signed requests clear auth and
+// reach the handler, where membership authorization (not mere registration) is
+// what the test exercises.
+func register(t *testing.T, h *authHarness, id cs.Identity, handle string) {
+	t.Helper()
+	if err := h.store.AddUser(hex.EncodeToString(id.SignPub), handle, RoleMember); err != nil {
+		t.Fatalf("register %s: %v", handle, err)
+	}
+}
+
+// TestAudit_HorizontalMetadataLeak ports the auditor's H3 (Alto) finding: bob is
+// REGISTERED on the bus but is NOT a member of alice's room. Before the fix the
+// GET endpoints checked registration, not membership, so bob could read the
+// room's subject, the full member list (with everyone's public keys), alice's
+// room directory, and even alice's sealed key. Now every one of those returns
+// 403 to bob, while alice (owner/member) and carol (plain member) get 200.
+func TestAudit_HorizontalMetadataLeak(t *testing.T) {
+	h := newAuthHarness(t, AuthEnforce)
+	roomID, aliceEp := seedRoom(t, h, "secret.subject.payroll")
+
+	// bob: registered, never invited.
+	bob, _ := cs.GenerateIdentity()
+	register(t, h, bob, "bob")
+
+	// carol: registered AND a plain (non-owner) member — the legitimate-member edge.
+	carol, _ := cs.GenerateIdentity()
+	register(t, h, carol, "carol")
+	carolEp := frame.EndpointID(carol.SignPub)
+	if err := h.store.AddMember(roomID, Member{Endpoint: carolEp, Role: RoleMember, SignPub: carol.SignPub, KexPub: carol.KexPub}, 1, []byte("carol-sealed")); err != nil {
+		t.Fatalf("add carol: %v", err)
+	}
+
+	n := 0
+	get := func(id cs.Identity, path string) int {
+		n++
+		code, _ := do(t, signedReq(t, h.ts.URL, "GET", path, nil, id, time.Now().Unix(), nonceN(n)))
+		return code
+	}
+
+	// Error path: bob (non-member) is forbidden on every room endpoint.
+	bobChecks := []struct {
+		name string
+		path string
+	}{
+		{"get room", "/rooms/" + roomID},
+		{"list members", "/rooms/" + roomID + "/members"},
+		{"alice room directory", "/members/" + aliceEp + "/rooms"},
+		{"alice sealed key", "/rooms/" + roomID + "/key?endpoint=" + aliceEp},
+		{"bob sealed key in alices room", "/rooms/" + roomID + "/key?endpoint=" + frame.EndpointID(bob.SignPub)},
+	}
+	for _, c := range bobChecks {
+		if code := get(bob, c.path); code != http.StatusForbidden {
+			t.Fatalf("bob (non-member) %s should be 403, got %d", c.name, code)
+		}
+	}
+
+	// Golden: alice (owner/member) reads her room's metadata, members, directory, key.
+	aliceChecks := []string{
+		"/rooms/" + roomID,
+		"/rooms/" + roomID + "/members",
+		"/members/" + aliceEp + "/rooms",
+		"/rooms/" + roomID + "/key?endpoint=" + aliceEp,
+	}
+	for _, p := range aliceChecks {
+		if code := get(h.alice, p); code != http.StatusOK {
+			t.Fatalf("alice (owner) %s should be 200, got %d", p, code)
+		}
+	}
+
+	// Edge: carol is a plain member, not the owner — she may still read the room.
+	if code := get(carol, "/rooms/"+roomID); code != http.StatusOK {
+		t.Fatalf("carol (member) get room should be 200, got %d", code)
+	}
+	if code := get(carol, "/rooms/"+roomID+"/members"); code != http.StatusOK {
+		t.Fatalf("carol (member) list members should be 200, got %d", code)
+	}
+
+	// Edge: carol may fetch her OWN sealed key but not alice's.
+	if code := get(carol, "/rooms/"+roomID+"/key?endpoint="+carolEp); code != http.StatusOK {
+		t.Fatalf("carol fetching her own key should be 200, got %d", code)
+	}
+	if code := get(carol, "/rooms/"+roomID+"/key?endpoint="+aliceEp); code != http.StatusForbidden {
+		t.Fatalf("carol fetching alice's key should be 403, got %d", code)
+	}
+}
+
+// nonceN yields a distinct nonce per request so the anti-replay cache never
+// rejects a fresh, legitimately-different request inside one test.
+func nonceN(i int) string {
+	return "authz-nonce-" + strconv.Itoa(i)
+}

pkg/membership/dos_test.go

@@ -0,0 +1,206 @@
+package membership
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/enmanuel/unibus/pkg/blobstore"
+)
+
+// dosServer builds a Server backed by a fresh store + blob store so a test can
+// drive ServeHTTP in-process (white-box) and observe its memory behavior without
+// a network round trip — the same in-process technique the auditor used.
+func dosServer(t *testing.T, mode AuthMode) *Server {
+	t.Helper()
+	dir := t.TempDir()
+	store, err := Open(filepath.Join(dir, "unibus.db"))
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	blobs, err := blobstore.New(filepath.Join(dir, "blobs"))
+	if err != nil {
+		t.Fatalf("open blobs: %v", err)
+	}
+	t.Cleanup(func() { store.Close() })
+	return NewServer(store, blobs, mode)
+}
+
+// zeroReader yields up to remaining zero bytes without ever allocating them, so
+// the test process itself never materializes a huge buffer (which would taint the
+// RSS measurement we are trying to make about the SERVER).
+type zeroReader struct{ remaining int64 }
+
+func (z *zeroReader) Read(p []byte) (int, error) {
+	if z.remaining <= 0 {
+		return 0, io.EOF
+	}
+	n := int64(len(p))
+	if n > z.remaining {
+		n = z.remaining
+	}
+	for i := int64(0); i < n; i++ {
+		p[i] = 0
+	}
+	z.remaining -= n
+	return int(n), nil
+}
+
+// vmRSSkB reads the resident set size (kB) of this process from /proc. Linux-only;
+// the caller skips on other platforms.
+func vmRSSkB(t *testing.T) int64 {
+	t.Helper()
+	b, err := os.ReadFile("/proc/self/status")
+	if err != nil {
+		t.Skipf("cannot read /proc/self/status: %v", err)
+	}
+	for _, line := range strings.Split(string(b), "\n") {
+		if strings.HasPrefix(line, "VmRSS:") {
+			f := strings.Fields(line)
+			if len(f) >= 2 {
+				v, _ := strconv.ParseInt(f[1], 10, 64)
+				return v
+			}
+		}
+	}
+	t.Skip("VmRSS not present in /proc/self/status")
+	return 0
+}
+
+// TestAudit_DoSBodyLimitNoAuth ports the auditor's H1 (Critical) vector: a peer
+// with NO valid signature posts an oversized body. Before the fix the middleware
+// io.ReadAll'd it unbounded (the auditor sent 400 MB and watched RSS jump from
+// 18 MB to 898 MB). Now the request is rejected 413 and the resident set does NOT
+// spike. Two shapes are covered:
+//
+//	(1) a truthful, over-ceiling Content-Length  -> rejected before any byte is read;
+//	(2) a lying / unknown length (chunked)       -> MaxBytesReader trips mid-read,
+//	    capping the buffered bytes at the ceiling instead of the attacker's 400 MB.
+func TestAudit_DoSBodyLimitNoAuth(t *testing.T) {
+	if runtime.GOOS != "linux" {
+		t.Skip("RSS probe is Linux-only")
+	}
+	srv := dosServer(t, AuthEnforce) // enforce: the request carries no signature
+
+	const huge = int64(400) << 20 // 400 MiB — the auditor's figure
+	// A spike threshold an order of magnitude below the attack. The old code would
+	// add ~400 MB+; the fix keeps the delta to at most one bounded buffer.
+	const maxSpikeKB = int64(96) << 10 // 96 MiB
+
+	// Shape 1: declared Content-Length over the blob ceiling -> early 413, no read.
+	runtime.GC()
+	before := vmRSSkB(t)
+	req := httptest.NewRequest(http.MethodPost, "/blobs", &zeroReader{remaining: huge})
+	req.ContentLength = huge
+	rec := httptest.NewRecorder()
+	srv.ServeHTTP(rec, req)
+	if rec.Code != http.StatusRequestEntityTooLarge {
+		t.Fatalf("over-declared body should be 413, got %d", rec.Code)
+	}
+	runtime.GC()
+	if d := vmRSSkB(t) - before; d > maxSpikeKB {
+		t.Fatalf("RSS spiked %d kB on a pre-declared oversized body (limit %d kB)", d, maxSpikeKB)
+	}
+
+	// Shape 2: unknown length (chunked-style). The middleware cannot reject by
+	// Content-Length, so MaxBytesReader must cap the read at maxBlobBytes.
+	runtime.GC()
+	before = vmRSSkB(t)
+	req = httptest.NewRequest(http.MethodPost, "/blobs", &zeroReader{remaining: huge})
+	req.ContentLength = -1
+	rec = httptest.NewRecorder()
+	srv.ServeHTTP(rec, req)
+	if rec.Code != http.StatusRequestEntityTooLarge {
+		t.Fatalf("unknown-length oversized body should be 413, got %d", rec.Code)
+	}
+	runtime.GC()
+	if d := vmRSSkB(t) - before; d > maxSpikeKB {
+		t.Fatalf("RSS spiked %d kB on a chunked oversized body (limit %d kB)", d, maxSpikeKB)
+	}
+}
+
+// TestBlobLimitGoldenAndBoundary covers the golden path (a normal blob is stored)
+// and the boundary (a body exactly at the ceiling is accepted; one byte over by
+// truthful Content-Length is rejected before buffering).
+func TestBlobLimitGoldenAndBoundary(t *testing.T) {
+	srv := dosServer(t, AuthOff) // AuthOff: the limits apply regardless of auth mode
+
+	// Golden: a small blob is accepted and hashed.
+	rec := httptest.NewRecorder()
+	srv.ServeHTTP(rec, httptest.NewRequest(http.MethodPost, "/blobs", strings.NewReader("hello blob")))
+	if rec.Code != http.StatusOK {
+		t.Fatalf("normal blob should be 200, got %d (%s)", rec.Code, rec.Body.String())
+	}
+
+	// Boundary: exactly at the ceiling is allowed (MaxBytesReader permits N bytes).
+	atLimit := strings.Repeat("a", maxBlobBytes)
+	rec = httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/blobs", strings.NewReader(atLimit))
+	req.ContentLength = int64(len(atLimit))
+	srv.ServeHTTP(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("blob exactly at the ceiling should be 200, got %d", rec.Code)
+	}
+
+	// Error: one byte over the ceiling (truthful Content-Length) -> 413 pre-read.
+	rec = httptest.NewRecorder()
+	req = httptest.NewRequest(http.MethodPost, "/blobs", &zeroReader{remaining: maxBlobBytes + 1})
+	req.ContentLength = maxBlobBytes + 1
+	srv.ServeHTTP(rec, req)
+	if rec.Code != http.StatusRequestEntityTooLarge {
+		t.Fatalf("blob one byte over the ceiling should be 413, got %d", rec.Code)
+	}
+}
+
+// TestControlBodyLimit checks the smaller JSON ceiling on a non-blob route: a body
+// over maxControlBodyBytes is rejected 413 before the handler runs.
+func TestControlBodyLimit(t *testing.T) {
+	srv := dosServer(t, AuthOff)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/rooms", &zeroReader{remaining: maxControlBodyBytes + 1})
+	req.ContentLength = maxControlBodyBytes + 1
+	srv.ServeHTTP(rec, req)
+	if rec.Code != http.StatusRequestEntityTooLarge {
+		t.Fatalf("control body over 1 MiB should be 413, got %d", rec.Code)
+	}
+}
+
+// TestRateLimitPerIP exercises the per-IP throttle: a burst from one IP eventually
+// gets 429 (error path), while a spread across distinct IPs is never throttled
+// (edge — the bucket is keyed per source, not global).
+func TestRateLimitPerIP(t *testing.T) {
+	srv := dosServer(t, AuthOff)
+
+	// Same IP: well past the burst -> at least one 429.
+	got429 := false
+	for i := 0; i < defaultRateBurst+50; i++ {
+		rec := httptest.NewRecorder()
+		req := httptest.NewRequest(http.MethodGet, "/rooms/none", nil)
+		req.RemoteAddr = "203.0.113.7:5555"
+		srv.ServeHTTP(rec, req)
+		if rec.Code == http.StatusTooManyRequests {
+			got429 = true
+			break
+		}
+	}
+	if !got429 {
+		t.Fatalf("a flood from one IP should eventually be rate-limited (429)")
+	}
+
+	// Distinct IPs: each gets a fresh bucket, so none is throttled.
+	for i := 0; i < 100; i++ {
+		rec := httptest.NewRecorder()
+		req := httptest.NewRequest(http.MethodGet, "/rooms/none", nil)
+		req.RemoteAddr = "198.51.100." + strconv.Itoa(i%254+1) + ":4444"
+		srv.ServeHTTP(rec, req)
+		if rec.Code == http.StatusTooManyRequests {
+			t.Fatalf("distinct IPs must not share a rate bucket; IP #%d got 429", i)
+		}
+	}
+}

pkg/membership/nonce_cache_test.go

@@ -0,0 +1,51 @@
+package membership
+
+import (
+	"strconv"
+	"testing"
+	"time"
+)
+
+// TestNonceCacheRememberPrune covers the replay/expiry behavior directly on the
+// cache: a fresh nonce is accepted (golden), an immediate repeat is rejected
+// (error), and after the TTL the same nonce is accepted again because its entry
+// was pruned (edge).
+func TestNonceCacheRememberPrune(t *testing.T) {
+	nc := newNonceCache(50*time.Millisecond, 1000)
+	base := time.Now()
+
+	if !nc.rememberOrReject("a", base) {
+		t.Fatalf("first sighting should be accepted")
+	}
+	if nc.rememberOrReject("a", base) {
+		t.Fatalf("an immediate replay should be rejected")
+	}
+	if !nc.rememberOrReject("a", base.Add(60*time.Millisecond)) {
+		t.Fatalf("after the TTL the nonce should be accepted again (pruned)")
+	}
+}
+
+// TestNonceCacheCapBounded covers the memory bound (audit H7): with a long TTL so
+// nothing expires, inserting far more nonces than the cap must still keep the
+// cache at or under the cap (oldest evicted), and the order queue must not drift
+// from the map.
+func TestNonceCacheCapBounded(t *testing.T) {
+	const capacity = 100
+	nc := newNonceCache(time.Hour, capacity)
+	base := time.Now()
+	for i := 0; i < 500; i++ {
+		nc.rememberOrReject("n"+strconv.Itoa(i), base)
+	}
+
+	nc.mu.Lock()
+	size := len(nc.seen)
+	orderLen := len(nc.order)
+	nc.mu.Unlock()
+
+	if size > capacity {
+		t.Fatalf("cache exceeded its cap: %d > %d", size, capacity)
+	}
+	if orderLen != size {
+		t.Fatalf("order queue drifted from the map: order=%d seen=%d", orderLen, size)
+	}
+}

pkg/membership/owner_nonce_test.go

@@ -0,0 +1,88 @@
+package membership
+
+import (
+	"encoding/json"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	cs "fn-registry/functions/cybersecurity"
+
+	"github.com/enmanuel/unibus/pkg/frame"
+)
+
+// TestAudit_OwnerSpoof ports the auditor's H6 finding: handleCreateRoom did not
+// bind the body's declared owner to the request signer, so a registered peer
+// could create rooms in another identity's name. Now the owner endpoint AND the
+// owner signing key must both be the authenticated signer's.
+func TestAudit_OwnerSpoof(t *testing.T) {
+	h := newAuthHarness(t, AuthEnforce)
+
+	bob, _ := cs.GenerateIdentity()
+	register(t, h, bob, "bob")
+	bobEp := frame.EndpointID(bob.SignPub)
+	victim, _ := cs.GenerateIdentity()
+
+	post := func(id cs.Identity, owner endpointJSON, nonce string) int {
+		body, _ := json.Marshal(createRoomReq{Subject: "some.room", Owner: owner})
+		code, _ := do(t, signedReq(t, h.ts.URL, "POST", "/rooms", body, id, time.Now().Unix(), nonce))
+		return code
+	}
+
+	// Error path: bob signs, body claims victim as owner -> 403.
+	if code := post(bob, endpointJSON{Endpoint: frame.EndpointID(victim.SignPub), SignPub: victim.SignPub, KexPub: victim.KexPub}, "spoof-1"); code != http.StatusForbidden {
+		t.Fatalf("owner-spoofed create should be 403, got %d", code)
+	}
+
+	// Edge: bob declares his own endpoint but a foreign signing key -> 403 (the
+	// key, not just the endpoint string, is bound to the signer).
+	if code := post(bob, endpointJSON{Endpoint: bobEp, SignPub: victim.SignPub, KexPub: victim.KexPub}, "spoof-2"); code != http.StatusForbidden {
+		t.Fatalf("create with a foreign owner key should be 403, got %d", code)
+	}
+
+	// Golden: alice creates a room owned by herself -> 201.
+	aliceEp := frame.EndpointID(h.alice.SignPub)
+	if code := post(h.alice, endpointJSON{Endpoint: aliceEp, SignPub: h.alice.SignPub, KexPub: h.alice.KexPub}, "owner-ok"); code != http.StatusCreated {
+		t.Fatalf("self-owned create should be 201, got %d", code)
+	}
+}
+
+// TestAudit_NonceCachePoisonPreAuth ports the auditor's H7 finding: the replay
+// cache was populated BEFORE the allowlist check, so any unregistered identity
+// (Ed25519 keys are free) could seed nonces into it. Now IsAuthorized runs first,
+// so an unauthorized identity's nonce is never cached: a repeat of the same nonce
+// still fails as "not authorized", not "replayed nonce".
+func TestAudit_NonceCachePoisonPreAuth(t *testing.T) {
+	h := newAuthHarness(t, AuthEnforce)
+
+	eve, _ := cs.GenerateIdentity() // valid signatures, NOT on the allowlist
+	now := time.Now().Unix()
+
+	code1, body1 := do(t, signedReq(t, h.ts.URL, "GET", "/rooms/x", nil, eve, now, "poison-nonce"))
+	if code1 != http.StatusUnauthorized || !strings.Contains(body1, "not authorized") {
+		t.Fatalf("unregistered first request should be 401 not-authorized, got %d (%s)", code1, body1)
+	}
+
+	// Same nonce again: if the nonce had been cached, this would report "replayed
+	// nonce". It must still be "not authorized" — proving the nonce was NOT cached.
+	code2, body2 := do(t, signedReq(t, h.ts.URL, "GET", "/rooms/x", nil, eve, now, "poison-nonce"))
+	if code2 != http.StatusUnauthorized {
+		t.Fatalf("unregistered replay should still be 401, got %d", code2)
+	}
+	if strings.Contains(body2, "replayed") {
+		t.Fatalf("an unauthorized identity's nonce was cached pre-auth: %s", body2)
+	}
+	if !strings.Contains(body2, "not authorized") {
+		t.Fatalf("second unregistered request should still be not-authorized, got: %s", body2)
+	}
+
+	// Positive control: an AUTHORIZED identity's replay IS still rejected, so the
+	// reorder did not weaken anti-replay for legitimate traffic.
+	if code, _ := do(t, signedReq(t, h.ts.URL, "GET", aliceRoomsPath(h), nil, h.alice, now, "alice-live")); code != http.StatusOK {
+		t.Fatalf("alice's first request should be 200, got %d", code)
+	}
+	if code, body := do(t, signedReq(t, h.ts.URL, "GET", aliceRoomsPath(h), nil, h.alice, now, "alice-live")); code != http.StatusUnauthorized || !strings.Contains(body, "replayed") {
+		t.Fatalf("alice's replay should be 401 replayed nonce, got %d (%s)", code, body)
+	}
+}

pkg/membership/ratelimit.go

@@ -0,0 +1,93 @@
+package membership
+
+import (
+	"net"
+	"net/http"
+	"sync"
+	"time"
+
+	"golang.org/x/time/rate"
+)
+
+// ipRateLimiter is a per-source-IP token-bucket rate limiter for the control
+// plane. It exists to blunt pre-auth flooding: an unauthenticated peer that
+// hammers the HTTP API (signature verification is not free, and io is bounded
+// but still real) is throttled before it can amplify load. Like the nonceCache,
+// this is transport glue specific to unibus, not a registry primitive — the
+// report 0003 made the same call for the nonce cache (it would only drag a NATS
+// dependency into the multi-domain registry go.mod for one helper).
+//
+// Each distinct IP gets its own golang.org/x/time/rate.Limiter (a standard
+// token bucket already in the module graph, so no new dependency). Idle buckets
+// are reaped so the map cannot grow without bound under a churn of source IPs.
+type ipRateLimiter struct {
+	mu      sync.Mutex
+	buckets map[string]*ipBucket
+	r       rate.Limit
+	burst   int
+	ttl     time.Duration
+}
+
+type ipBucket struct {
+	lim  *rate.Limiter
+	seen time.Time
+}
+
+// newIPRateLimiter builds a limiter granting r tokens/second with the given
+// burst per IP. ttl bounds how long an idle bucket is retained before being
+// reaped. r<=0 disables limiting (Allow always true) so dev/loopback stacks are
+// unaffected.
+func newIPRateLimiter(r rate.Limit, burst int, ttl time.Duration) *ipRateLimiter {
+	return &ipRateLimiter{
+		buckets: make(map[string]*ipBucket),
+		r:       r,
+		burst:   burst,
+		ttl:     ttl,
+	}
+}
+
+// allow reports whether a request from ip may proceed now, consuming one token
+// on success. A disabled limiter (r<=0) always allows. Reaping of stale buckets
+// is amortized: it runs only when the map has grown past a small threshold, so
+// the common path is a single map lookup under the mutex.
+func (l *ipRateLimiter) allow(ip string, now time.Time) bool {
+	if l == nil || l.r <= 0 {
+		return true
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+
+	if len(l.buckets) > 1024 {
+		l.reapLocked(now)
+	}
+	b, ok := l.buckets[ip]
+	if !ok {
+		b = &ipBucket{lim: rate.NewLimiter(l.r, l.burst)}
+		l.buckets[ip] = b
+	}
+	b.seen = now
+	return b.lim.AllowN(now, 1)
+}
+
+// reapLocked drops buckets idle for longer than ttl. The caller holds l.mu.
+func (l *ipRateLimiter) reapLocked(now time.Time) {
+	for ip, b := range l.buckets {
+		if now.Sub(b.seen) > l.ttl {
+			delete(l.buckets, ip)
+		}
+	}
+}
+
+// clientIP extracts the source IP of an HTTP request, stripping the port. It
+// trusts the transport's RemoteAddr only (no X-Forwarded-For parsing): a public
+// deployment terminates TLS at this process or behind a proxy that the operator
+// controls, and honoring an attacker-supplied header would let a single IP fan
+// its quota across forged identities. If parsing fails the whole RemoteAddr is
+// used as the key (still a stable per-connection bucket).
+func clientIP(r *http.Request) string {
+	host, _, err := net.SplitHostPort(r.RemoteAddr)
+	if err != nil {
+		return r.RemoteAddr
+	}
+	return host
+}

pkg/membership/require_e2e_test.go

@@ -0,0 +1,46 @@
+package membership
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+// TestRequireEncryptedRoomsRejectsCleartext is the control-plane half of the
+// audit H4 minimum defense: with RequireEncryptedRooms on (the public posture),
+// creating a cleartext (ModeNATS) room is refused 403, while an encrypted room is
+// created normally. This is what guarantees no message ever rides the un-ACL'd
+// NATS subject in the clear on a public deployment.
+func TestRequireEncryptedRoomsRejectsCleartext(t *testing.T) {
+	srv := dosServer(t, AuthOff)
+	srv.RequireEncryptedRooms = true
+
+	create := func(encrypt bool) int {
+		body, _ := json.Marshal(createRoomReq{
+			Subject:       "payroll.subject",
+			Policy:        policyJSON{Encrypt: encrypt, Persist: encrypt, SignMsgs: encrypt},
+			Owner:         endpointJSON{Endpoint: "owner-ep", SignPub: []byte("sp"), KexPub: []byte("kp")},
+			SealedKeySelf: []byte("sealed"),
+		})
+		rec := httptest.NewRecorder()
+		srv.ServeHTTP(rec, httptest.NewRequest(http.MethodPost, "/rooms", bytes.NewReader(body)))
+		return rec.Code
+	}
+
+	// Error path: a cleartext room is refused.
+	if code := create(false); code != http.StatusForbidden {
+		t.Fatalf("cleartext room under RequireEncryptedRooms should be 403, got %d", code)
+	}
+	// Golden: an encrypted room is created.
+	if code := create(true); code != http.StatusCreated {
+		t.Fatalf("encrypted room should be 201, got %d", code)
+	}
+
+	// Edge: with the flag OFF (loopback/dev), cleartext rooms are allowed again.
+	srv.RequireEncryptedRooms = false
+	if code := create(false); code != http.StatusCreated {
+		t.Fatalf("cleartext room with the flag off should be 201, got %d", code)
+	}
+}

pkg/membership/server.go

@@ -2,6 +2,7 @@ package membership
 
 import (
 	"bytes"
+	"context"
 	"database/sql"
 	"encoding/json"
 	"errors"
@@ -15,7 +16,35 @@ import (
 
 	cs "fn-registry/functions/cybersecurity"
 
+	"golang.org/x/time/rate"
+
 	"github.com/enmanuel/unibus/pkg/blobstore"
+	"github.com/enmanuel/unibus/pkg/frame"
+)
+
+// Body-size ceilings for the control plane. They bound how much an unauthenticated
+// peer can make the server buffer in RAM before the request is even authenticated
+// (the signature is verified over the full body, so the body must be read — but
+// not unboundedly). maxControlBodyBytes covers JSON metadata requests; /blobs gets
+// a separate, larger ceiling because media ciphertext is legitimately bigger. A
+// request whose declared Content-Length already exceeds its ceiling is rejected
+// before a single byte is buffered.
+const (
+	maxControlBodyBytes = 1 << 20  // 1 MiB for JSON control-plane requests
+	maxBlobBytes        = 16 << 20 // 16 MiB for a single media blob upload
+	// MaxHeaderBytes caps request header size; wired into the http.Server by the
+	// command. Exported so the bound lives next to its body-size siblings.
+	MaxHeaderBytes = 1 << 20 // 1 MiB
+)
+
+// Per-IP rate-limit defaults for the control plane. Tuned for an interactive
+// human/agent bus rather than a high-QPS API: a steady ~20 req/s with a burst of
+// 40 absorbs a chat client's bursty polling while throttling a flood. Loopback
+// dev stacks pass r<=0 to disable limiting entirely.
+const (
+	defaultRatePerSec = rate.Limit(20)
+	defaultRateBurst  = 40
+	rateBucketTTL     = 10 * time.Minute
 )
 
 // Server is the HTTP control plane: the authoritative source of room metadata,
@@ -32,18 +61,32 @@ type Server struct {
 	mux      *http.ServeMux
 	authMode AuthMode
 	nonces   *nonceCache
+	limiter  *ipRateLimiter
+
+	// RequireEncryptedRooms, when true, refuses to create cleartext (ModeNATS)
+	// rooms. It is the minimum-defensive control for the data plane (audit H4):
+	// the embedded NATS has no per-subject ACL, so a cleartext room is readable by
+	// any registered peer that knows (or guesses) its subject. Forcing every room
+	// to be end-to-end encrypted keeps message CONTENT confidential even when the
+	// transport offers no subject isolation. The command sets this on a public
+	// (non-loopback) bind. See dev/0004d-dataplane-acl.md for the full rationale
+	// and the residual metadata exposure this does NOT close.
+	RequireEncryptedRooms bool
 }
 
 // NewServer wires the membership store and blob store into an http.Handler. The
 // authMode selects the control-plane auth rollout state (AuthOff for callers and
-// tests that have not migrated to signed requests yet).
+// tests that have not migrated to signed requests yet). It installs a per-IP
+// rate limiter with the package defaults; loopback dev behavior is unchanged
+// because the burst comfortably exceeds any single client's request rate.
 func NewServer(store *Store, blobs *blobstore.Store, authMode AuthMode) *Server {
 	s := &Server{
 		store:    store,
 		blobs:    blobs,
 		mux:      http.NewServeMux(),
 		authMode: authMode,
-		nonces:   newNonceCache(nonceTTL),
+		nonces:   newNonceCache(nonceTTL, maxNonceCacheEntries),
+		limiter:  newIPRateLimiter(defaultRatePerSec, defaultRateBurst, rateBucketTTL),
 	}
 	s.routes()
 	return s
@@ -53,23 +96,54 @@ func NewServer(store *Store, blobs *blobstore.Store, authMode AuthMode) *Server
 // (signature verification + anti-replay + allowlist) ahead of the router
 // according to authMode, then dispatches to the matched handler.
 func (s *Server) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	now := time.Now()
+
+	// Per-IP rate limit runs first, ahead of auth and body reads, so a flood is
+	// shed at the cheapest possible point. The health probe is exempt so liveness
+	// checks are never throttled.
+	if !isAuthExempt(r) && !s.limiter.allow(clientIP(r), now) {
+		writeErr(w, http.StatusTooManyRequests, "rate limit exceeded")
+		return
+	}
+
+	// Cap how much body we will buffer, BEFORE reading a single byte. The ceiling
+	// is per-route: /blobs may legitimately carry a media ciphertext, everything
+	// else is small JSON. A declared Content-Length over the ceiling is rejected
+	// outright (no buffering); MaxBytesReader then guards against a lying or
+	// chunked sender by failing the read once the limit is crossed. This is the
+	// fix for the pre-auth DoS: without it an unauthenticated peer could make the
+	// server buffer an unbounded body in RAM before authenticate() ever ran.
+	limit := int64(maxControlBodyBytes)
+	if r.Method == http.MethodPost && r.URL.Path == "/blobs" {
+		limit = int64(maxBlobBytes)
+	}
+	if r.ContentLength > limit {
+		writeErr(w, http.StatusRequestEntityTooLarge, "request body too large")
+		return
+	}
+	r.Body = http.MaxBytesReader(w, r.Body, limit)
+
 	if s.authMode == AuthOff || isAuthExempt(r) {
 		s.mux.ServeHTTP(w, r)
 		return
 	}
 
-	// Buffer the body so the signature can be verified over it and the handler
-	// still reads it. Bodies on the control plane are small (JSON metadata or a
-	// media blob already capped upstream), so full buffering is acceptable.
+	// Buffer the (now bounded) body so the signature can be verified over it and
+	// the handler still reads it.
 	body, err := io.ReadAll(r.Body)
 	if err != nil {
-		writeErr(w, http.StatusBadRequest, "read body: "+err.Error())
+		if isBodyTooLarge(err) {
+			writeErr(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
+		writeErr(w, http.StatusBadRequest, "read body")
 		return
 	}
 	_ = r.Body.Close()
 	r.Body = io.NopCloser(bytes.NewReader(body))
 
-	if _, err := s.authenticate(r, body, time.Now()); err != nil {
+	res, err := s.authenticate(r, body, now)
+	if err != nil {
 		if s.authMode == AuthSoft {
 			log.Printf("[auth] soft: would reject %s %s: %v", r.Method, r.URL.Path, err)
 			s.mux.ServeHTTP(w, r)
@@ -78,7 +152,53 @@ func (s *Server) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, http.StatusUnauthorized, "unauthorized: "+err.Error())
 		return
 	}
-	s.mux.ServeHTTP(w, r)
+	// Carry the authenticated signer's endpoint into the handler so room handlers
+	// can authorize by membership (audit H3). Only set on a verified identity.
+	s.mux.ServeHTTP(w, r.WithContext(withSigner(r.Context(), res.endpoint)))
+}
+
+// isBodyTooLarge reports whether err is the sentinel returned by MaxBytesReader
+// when the body exceeds its limit, so the middleware can map it to 413.
+func isBodyTooLarge(err error) bool {
+	var maxErr *http.MaxBytesError
+	return errors.As(err, &maxErr)
+}
+
+// ctxKey is the unexported type for this package's request-context keys, so the
+// values cannot collide with keys set by other packages.
+type ctxKey int
+
+const ctxSignerEndpoint ctxKey = iota
+
+// withSigner returns a context carrying the authenticated signer's endpoint id.
+func withSigner(ctx context.Context, endpoint string) context.Context {
+	return context.WithValue(ctx, ctxSignerEndpoint, endpoint)
+}
+
+// signerEndpoint returns the authenticated signer's endpoint id and whether one
+// is present. It is absent under AuthOff (no verification) and when a soft-mode
+// request was let through unauthenticated — in both cases membership
+// authorization is skipped, preserving dev/legacy behavior.
+func signerEndpoint(r *http.Request) (string, bool) {
+	v, ok := r.Context().Value(ctxSignerEndpoint).(string)
+	return v, ok && v != ""
+}
+
+// requireMember authorizes a room request by membership (audit H3): it returns
+// the signer endpoint and true when the request may proceed, or writes 403 and
+// returns false when an authenticated signer is not a member of roomID. When no
+// authenticated signer is present (AuthOff/dev, or soft pass-through) it allows
+// the request — membership is only enforced once the caller's identity is known.
+func (s *Server) requireMember(w http.ResponseWriter, r *http.Request, roomID string) (string, bool) {
+	signer, ok := signerEndpoint(r)
+	if !ok {
+		return "", true
+	}
+	if _, err := s.store.GetMember(roomID, signer); err != nil {
+		writeErr(w, http.StatusForbidden, "forbidden: not a member of this room")
+		return signer, false
+	}
+	return signer, true
 }
 
 // isAuthExempt lists requests that bypass control-plane auth even under enforce.
@@ -188,6 +308,15 @@ func writeErr(w http.ResponseWriter, code int, msg string) {
 	writeJSON(w, code, map[string]string{"error": msg})
 }
 
+// writeServerErr logs the internal error detail and returns ONLY a generic
+// message to the client (audit H12): raw store/blob errors embed SQL fragments
+// and filesystem paths, which must not leak to a caller. Use it for any error
+// that originates inside the server (5xx, or a not-found wrapping a store error).
+func writeServerErr(w http.ResponseWriter, r *http.Request, code int, publicMsg string, err error) {
+	log.Printf("[handler] %s %s -> %d: %v", r.Method, r.URL.Path, code, err)
+	writeErr(w, code, publicMsg)
+}
+
 // canonicalSig returns the bytes to verify for a request: the request struct
 // re-marshaled with its Sig field cleared. The caller passes a copy with Sig
 // already zeroed. This is symmetric with how the client signs.
@@ -232,6 +361,24 @@ func (s *Server) handleCreateRoom(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, http.StatusBadRequest, "subject and owner.endpoint required")
 		return
 	}
+	// Data-plane minimum defense (audit H4): on a public deployment cleartext
+	// rooms are disabled, so no message ever rides the un-ACL'd NATS subject in
+	// the clear for another registered peer to sniff.
+	if s.RequireEncryptedRooms && !req.Policy.Encrypt {
+		writeErr(w, http.StatusForbidden,
+			"cleartext rooms are disabled on this deployment; create an encrypted (Matrix-policy) room")
+		return
+	}
+	// Owner binding (audit H6): the declared owner must BE the authenticated
+	// signer — both the endpoint id and the signing key. Otherwise a registered
+	// peer could create rooms in another identity's name. Enforced only when an
+	// authenticated signer is present (AuthOff/dev trusts the caller).
+	if signer, ok := signerEndpoint(r); ok {
+		if req.Owner.Endpoint != signer || frame.EndpointID(req.Owner.SignPub) != signer {
+			writeErr(w, http.StatusForbidden, "forbidden: room owner must be the authenticated signer")
+			return
+		}
+	}
 	roomID := newULID()
 	info := RoomInfo{
 		RoomID:        roomID,
@@ -242,7 +389,7 @@ func (s *Server) handleCreateRoom(w http.ResponseWriter, r *http.Request) {
 		OwnerEndpoint: req.Owner.Endpoint,
 	}
 	if err := s.store.CreateRoom(info, req.Owner.SignPub, req.Owner.KexPub, req.SealedKeySelf); err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 		return
 	}
 	writeJSON(w, http.StatusCreated, createRoomResp{RoomID: roomID})
@@ -264,7 +411,7 @@ func (s *Server) handleInvite(w http.ResponseWriter, r *http.Request) {
 	}
 	info, err := s.store.GetRoom(roomID)
 	if err != nil {
-		writeErr(w, http.StatusNotFound, err.Error())
+		writeServerErr(w, r, http.StatusNotFound, "room not found", err)
 		return
 	}
 	m := Member{
@@ -274,7 +421,7 @@ func (s *Server) handleInvite(w http.ResponseWriter, r *http.Request) {
 		KexPub:   req.Member.KexPub,
 	}
 	if err := s.store.AddMember(roomID, m, info.Epoch, req.SealedKey); err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 		return
 	}
 	writeJSON(w, http.StatusOK, map[string]string{"status": "invited"})
@@ -287,6 +434,20 @@ func (s *Server) handleGetKey(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, http.StatusBadRequest, "endpoint query param required")
 		return
 	}
+	// A sealed room key is sealed to one identity's X25519 key. Serving it only to
+	// that identity (the signer) stops a registered peer from harvesting another
+	// member's sealed key (audit H3). Membership is implied by owning a sealed key,
+	// but we also require the signer to be a member for defense in depth.
+	if signer, ok := signerEndpoint(r); ok {
+		if endpoint != signer {
+			writeErr(w, http.StatusForbidden, "forbidden: may only fetch your own sealed key")
+			return
+		}
+		if _, err := s.store.GetMember(roomID, signer); err != nil {
+			writeErr(w, http.StatusForbidden, "forbidden: not a member of this room")
+			return
+		}
+	}
 	epoch := 0
 	if e := r.URL.Query().Get("epoch"); e != "" {
 		if n, err := strconv.Atoi(e); err == nil {
@@ -300,7 +461,7 @@ func (s *Server) handleGetKey(w http.ResponseWriter, r *http.Request) {
 				"not invited to this encrypted room: no key has been sealed for your identity. Ask the room owner to invite you before joining.")
 			return
 		}
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 		return
 	}
 	writeJSON(w, http.StatusOK, keyResp{Epoch: ep, SealedKey: sealed})
@@ -308,9 +469,14 @@ func (s *Server) handleGetKey(w http.ResponseWriter, r *http.Request) {
 
 func (s *Server) handleListMembers(w http.ResponseWriter, r *http.Request) {
 	roomID := r.PathValue("id")
+	// Membership authorization (audit H3): the member list exposes every member's
+	// sign_pub + kex_pub, so it must not be served to a non-member.
+	if _, ok := s.requireMember(w, r, roomID); !ok {
+		return
+	}
 	members, err := s.store.ListMembers(roomID)
 	if err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeErr(w, http.StatusInternalServerError, "internal error")
 		return
 	}
 	out := make([]memberJSON, 0, len(members))
@@ -326,9 +492,15 @@ func (s *Server) handleListMemberRooms(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, http.StatusBadRequest, "endpoint required")
 		return
 	}
+	// A peer may only enumerate its OWN room directory (audit H3): otherwise any
+	// registered identity could map another's entire social graph of rooms.
+	if signer, ok := signerEndpoint(r); ok && endpoint != signer {
+		writeErr(w, http.StatusForbidden, "forbidden: may only list your own rooms")
+		return
+	}
 	rooms, err := s.store.ListRoomsForEndpoint(endpoint)
 	if err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeErr(w, http.StatusInternalServerError, "internal error")
 		return
 	}
 	out := make([]memberRoomJSON, 0, len(rooms))
@@ -346,9 +518,12 @@ func (s *Server) handleListMemberRooms(w http.ResponseWriter, r *http.Request) {
 
 func (s *Server) handleGetRoom(w http.ResponseWriter, r *http.Request) {
 	roomID := r.PathValue("id")
+	if _, ok := s.requireMember(w, r, roomID); !ok {
+		return
+	}
 	info, err := s.store.GetRoom(roomID)
 	if err != nil {
-		writeErr(w, http.StatusNotFound, err.Error())
+		writeErr(w, http.StatusNotFound, "room not found")
 		return
 	}
 	writeJSON(w, http.StatusOK, roomResp{
@@ -378,7 +553,7 @@ func (s *Server) handleRekey(w http.ResponseWriter, r *http.Request) {
 	// Bump epoch, then store the fresh sealed keys for the remaining members,
 	// then remove the kicked/left members.
 	if err := s.store.BumpEpoch(roomID, req.NewEpoch); err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 		return
 	}
 	keys := make(map[string][]byte, len(req.Keys))
@@ -387,13 +562,13 @@ func (s *Server) handleRekey(w http.ResponseWriter, r *http.Request) {
 	}
 	if len(keys) > 0 {
 		if err := s.store.PutSealedKeys(roomID, req.NewEpoch, keys); err != nil {
-			writeErr(w, http.StatusInternalServerError, err.Error())
+			writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 			return
 		}
 	}
 	for _, ep := range req.Remove {
 		if err := s.store.RemoveMember(roomID, ep); err != nil {
-			writeErr(w, http.StatusInternalServerError, err.Error())
+			writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 			return
 		}
 	}
@@ -401,14 +576,23 @@ func (s *Server) handleRekey(w http.ResponseWriter, r *http.Request) {
 }
 
 func (s *Server) handlePutBlob(w http.ResponseWriter, r *http.Request) {
+	// The body arrives already bounded: ServeHTTP wraps it in a MaxBytesReader
+	// (maxBlobBytes) and rejects an over-declared Content-Length before this
+	// handler runs, in every auth mode. Reading here therefore cannot buffer
+	// more than the ceiling; a sender that lies about its length (e.g. chunked)
+	// trips MaxBytesReader and we map that to 413 rather than a generic 400.
 	data, err := io.ReadAll(r.Body)
 	if err != nil {
-		writeErr(w, http.StatusBadRequest, "read body: "+err.Error())
+		if isBodyTooLarge(err) {
+			writeErr(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
+		writeErr(w, http.StatusBadRequest, "read body")
 		return
 	}
 	hash, err := s.blobs.Put(data)
 	if err != nil {
-		writeErr(w, http.StatusInternalServerError, err.Error())
+		writeServerErr(w, r, http.StatusInternalServerError, "internal error", err)
 		return
 	}
 	writeJSON(w, http.StatusOK, blobResp{Hash: hash})
@@ -422,7 +606,7 @@ func (s *Server) handleGetBlob(w http.ResponseWriter, r *http.Request) {
 	}
 	data, err := s.blobs.Get(hash)
 	if err != nil {
-		writeErr(w, http.StatusNotFound, err.Error())
+		writeServerErr(w, r, http.StatusNotFound, "not found", err)
 		return
 	}
 	w.Header().Set("Content-Type", "application/octet-stream")