diff --git a/.gitignore b/.gitignore
index d712160..2bcfc01 100644
--- a/.gitignore
+++ b/.gitignore
@@ -47,3 +47,4 @@ scripts/_*
 sources_cache/legal/
 scratch_rtc_e2e/
 scratch_sandbox/
+scratch_enrich/
diff --git a/agent/pyproject.toml b/agent/pyproject.toml
index 50c295a..89a595c 100644
--- a/agent/pyproject.toml
+++ b/agent/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "gti-fiscal-agent"
-version = "1.2.5"
+version = "1.3.0"
 requires-python = ">=3.10"
 dependencies = ["httpx>=0.27,<1"]
 
diff --git a/agent/scripts/run_agent_service.py b/agent/scripts/run_agent_service.py
index 24eeabd..8d0ec21 100644
--- a/agent/scripts/run_agent_service.py
+++ b/agent/scripts/run_agent_service.py
@@ -107,7 +107,7 @@ class Config:
         self.poll_interval_seconds = float(env.get("POLL_INTERVAL_SECONDS", "5"))
         # RTC auto-apply is a SEPARATE switch (default OFF): FIREBIRD_WRITES alone never lets the Agent fill cClassTrib.
         self.rtc_auto_apply_max_per_job = int(env.get("RTC_AUTO_APPLY_MAX_PER_JOB", str(_rtc_apply.DEFAULT_MAX_PER_JOB)))
-        self.agent_version = env.get("AGENT_VERSION", "1.2.5")
+        self.agent_version = env.get("AGENT_VERSION", "1.3.0")
         self.schema_fingerprint = env.get("SCHEMA_FINGERPRINT", REFERENCE_SCHEMA_V1)
         self.state_db = env.get("AGENT_STATE_DB", str(AGENT_ROOT / "agent_state.sqlite"))
         # Pilot write authorization: both must be explicitly set by the installer for
@@ -349,7 +349,7 @@ def execute_job(job, config, executor, state, env):
             "fiscal_context": fiscal_context,
             "rtc_auto_apply": rtc_apply,
             "fiscal_mode": getattr(config, "fiscal_mode", None),
-            "release_version": env.get("AGENT_VERSION") or "1.2.5",
+            "release_version": env.get("AGENT_VERSION") or "1.3.0",
             "products_total": len(products),
             "products_changed": len(to_process),
             "products_processed": summary["processed"],
@@ -520,7 +520,7 @@ def run(config, env, *, max_iterations=None, stopper=None):
 def enforce_startup_preflight(env):
     """Full read-only preflight at startup (Hub, auth, Firebird, capabilities, company context, guards). Fail closed."""
     from gti_fiscal_agent.preflight import run_preflight
-    report = run_preflight(env, agent_version=env.get("AGENT_VERSION") or "1.2.5")
+    report = run_preflight(env, agent_version=env.get("AGENT_VERSION") or "1.3.0")
     logger.info("startup preflight: %s", json.dumps({k: report[k] for k in ("PREFLIGHT", "blockers", "fiscal_mode", "effective_firebird_writes")}, ensure_ascii=False))
     print(json.dumps({"PREFLIGHT": report["PREFLIGHT"], "blockers": report["blockers"]}, ensure_ascii=False))
     if report["PREFLIGHT"] != "PASS":
diff --git a/agent/src/gti_fiscal_agent/enrichment.py b/agent/src/gti_fiscal_agent/enrichment.py
index c135c30..e635ff1 100644
--- a/agent/src/gti_fiscal_agent/enrichment.py
+++ b/agent/src/gti_fiscal_agent/enrichment.py
@@ -22,8 +22,8 @@ logger = logging.getLogger("gti_fiscal_agent.enrichment")
 
 CAPABILITY = "PRODUCT_ENRICHMENT_WRITE_V1"
 ALLOWED_FIELDS = ("CLASSIFICACAO_FISCAL", "CEST", "IPI")            # explicit allowlist; anything else raises before touching SQL
-READ_FIELDS = ("CODIGO", "PRODUTO", "GTIN") + ALLOWED_FIELDS
-POLL_SECONDS = 3.0
+READ_FIELDS = ("CODIGO", "PRODUTO", "GTIN", "CODBARRA") + ALLOWED_FIELDS      # the SGV operator types the EAN in CODBARRA; GTIN (fiscal) is mostly empty
+POLL_SECONDS = 2.0
 SAFETY_DIFF_SECONDS = 60.0
 OVERRIDE_CHECK_SECONDS = 20.0
 WATCH_WINDOW_SECONDS = 120.0            # a fresh product is re-read for a while: the operator may add the EAN seconds after the first save
@@ -112,6 +112,21 @@ def _digits(value) -> str:
     return "".join(ch for ch in _text(value) if ch.isdigit())
 
 
+def gtin_check_ok(value) -> bool:
+    """Same acceptance as central.app.enrichment.gtin (8/12/13/14 digits + check digit); parity is asserted by a test. Only used to pick WHICH column to send: the Hub decides."""
+    d = _digits(value)
+    if len(d) not in (8, 12, 13, 14) or len(set(d)) == 1:
+        return False
+    total = sum(int(x) * (3 if i % 2 == 0 else 1) for i, x in enumerate(reversed(d[:-1])))
+    return (10 - total % 10) % 10 == int(d[-1])
+
+
+def pick_gtin(row) -> str:
+    """First VALID of GTIN / CODBARRA; otherwise whatever was typed (the Hub classifies it INVALID_GTIN/MISSING_GTIN without calling any provider)."""
+    candidates = [_text(row.get("GTIN")), _text(row.get("CODBARRA"))]
+    return next((c for c in candidates if gtin_check_ok(c)), next((c for c in candidates if c), ""))
+
+
 def read_codes(conn) -> set[str]:
     cur = conn.cursor()
     cur.execute("SELECT CODIGO FROM C000025")
@@ -202,8 +217,20 @@ class EnrichmentWorker:
         self._last_diff = 0.0
         self._last_override = 0.0
         self._auth_cache = (None, 0.0)
+        self._count_ms = 0.0
+        self._wconn = None
         self.metrics = {"detected": 0, "resolved": 0, "fields_written": 0, "detect_latency_ms": []}
 
+    def _open(self, factory, attempts=3):
+        """Firebird can drop a connection under load ('connection lost'): retry briefly instead of aborting the product."""
+        for attempt in range(attempts):
+            try:
+                return factory()
+            except Exception:  # noqa: BLE001
+                if attempt == attempts - 1:
+                    raise
+                self.sleep(0.3 * (attempt + 1))
+
     # ---- feature gate ------------------------------------------------------------------------------------------------------------------------------
     def features(self, max_age=10.0) -> dict:
         if self._features is None or self.clock() - self._features_at > max_age:
@@ -234,7 +261,7 @@ class EnrichmentWorker:
 
     # ---- detection ---------------------------------------------------------------------------------------------------------------------------------
     def detect(self) -> list[str]:
-        conn = self.read_conn()
+        conn = self._open(self.read_conn)
         try:
             if self.store.meta("baseline_done") != "1":
                 codes = read_codes(conn)
@@ -245,7 +272,9 @@ class EnrichmentWorker:
                 self._last_count, self._last_diff = len(codes), self.clock()
                 return []
             started = self.clock()
+            t0 = time.monotonic()
             count = read_count(conn)
+            self._count_ms = (time.monotonic() - t0) * 1000
             moved = self._last_count is None or count != self._last_count
             if not moved and self.clock() - self._last_diff < SAFETY_DIFF_SECONDS:
                 return []
@@ -265,7 +294,7 @@ class EnrichmentWorker:
     # ---- one product ---------------------------------------------------------------------------------------------------------------------------------
     def process(self, item: dict):
         code, mode = item["code"], self.mode()
-        conn = self.read_conn()
+        conn = self._open(self.read_conn)
         try:
             rows = read_rows(conn, [code])
         finally:
@@ -274,14 +303,14 @@ class EnrichmentWorker:
         if row is None:
             self.store.update(code, state="GONE")
             return
-        snapshot = f"{_text(row['GTIN'])}|{_text(row['PRODUTO'])}"
+        snapshot = f"{pick_gtin(row)}|{_text(row['PRODUTO'])}"
         if mode not in ("DRY_RUN", "ENABLED"):
             self.store.update(code, state="SKIPPED_FEATURE", snapshot=snapshot)
             return
         attempts = int(item["attempts"] or 0) + 1
         current = {f: row[f] for f in ALLOWED_FIELDS}
         try:
-            result = self.hub.resolve_product({"product_code": code, "gtin": _text(row["GTIN"]), "description": _text(row["PRODUTO"]), "current": {k: (None if v is None else str(v)) for k, v in current.items()}})
+            result = self.hub.resolve_product({"product_code": code, "gtin": pick_gtin(row), "description": _text(row["PRODUTO"]), "current": {k: (None if v is None else str(v)) for k, v in current.items()}})
         except Exception as error:  # noqa: BLE001 - Hub offline / slow: the product stays pending, the SGV is unaffected
             delay = RETRY_BACKOFF[min(attempts - 1, len(RETRY_BACKOFF) - 1)]
             self.store.update(code, state="RETRY", attempts=attempts, next_attempt_at=self.clock() + delay, last_state=f"HUB_ERROR:{type(error).__name__}")
@@ -322,6 +351,8 @@ class EnrichmentWorker:
             if outcome["outcome"] == "APPLIED":
                 written[field] = value
                 self.metrics["fields_written"] += 1
+                self.store.update(code, written=written)                # remembered at once: a crash right after a write still leaves it protected/tracked
+        self._close_write()
         self.store.update(code, state="DRY_RUN_DONE" if mode == "DRY_RUN" else "DONE", attempts=attempts, snapshot=snapshot, written=written, last_state=state, next_attempt_at=None)
         if reports:
             try:
@@ -335,29 +366,31 @@ class EnrichmentWorker:
                 logger.warning("reprocess request for %s failed", code)
 
     def _write_safely(self, code, field, value) -> dict:
+        """Per field: fresh central authorization (fail closed) + guarded typed UPDATE. write_field re-reads the field INSIDE its own transaction and refuses when anything is there."""
         if not self._fresh_enabled():
             return {"outcome": "ABORTED", "reason": "AUTHORIZATION_CHANGED"}
-        conn = self.read_conn()
         try:
-            fresh = read_rows(conn, [code]).get(code)
-        finally:
-            conn.close()
-        if fresh is None:
-            return {"outcome": "ABORTED", "reason": "PRODUCT_NOT_FOUND"}
-        if _text(fresh[field]):
-            return {"outcome": "ABORTED", "reason": "VALUE_ALREADY_SET_NEVER_OVERWRITTEN"}
-        wconn = self.write_conn()
-        try:
-            return write_field(wconn, code, field, value)
-        finally:
-            wconn.close()
+            if self._wconn is None:
+                self._wconn = self._open(self.write_conn)
+            return write_field(self._wconn, code, field, value)
+        except Exception as error:  # noqa: BLE001 - one field failing never breaks the others or the SGV
+            self._close_write()
+            return {"outcome": "FAILED", "reason": type(error).__name__}
+
+    def _close_write(self):
+        conn, self._wconn = self._wconn, None
+        if conn is not None:
+            try:
+                conn.close()
+            except Exception:  # noqa: BLE001
+                pass
 
     # ---- human override watch --------------------------------------------------------------------------------------------------------------------------
     def check_overrides(self):
         tracked = self.store.tracked_written()
         if not tracked:
             return
-        conn = self.read_conn()
+        conn = self._open(self.read_conn)
         try:
             rows = read_rows(conn, [t["code"] for t in tracked])
         finally:
@@ -388,17 +421,23 @@ class EnrichmentWorker:
         self.detect()
         done = []
         for item in self.store.due(self.clock()):
-            self.process(item)
+            try:
+                self.process(item)
+            except Exception as error:  # noqa: BLE001 - keep the product retryable, never lose the batch
+                self._close_write()
+                attempts = int(item["attempts"] or 0) + 1
+                self.store.update(item["code"], state="RETRY", attempts=attempts, next_attempt_at=self.clock() + RETRY_BACKOFF[min(attempts - 1, len(RETRY_BACKOFF) - 1)], last_state=f"ERROR:{type(error).__name__}")
+                logger.warning("enrichment of %s failed: %s", item["code"], type(error).__name__)
             done.append(item["code"])
         for item in self.store.watching(self.clock()):                      # operator completes the record (EAN typed a moment later): re-resolve once per change
             rows = None
-            conn = self.read_conn()
+            conn = self._open(self.read_conn)
             try:
                 rows = read_rows(conn, [item["code"]])
             finally:
                 conn.close()
             row = (rows or {}).get(item["code"])
-            if row and f"{_text(row['GTIN'])}|{_text(row['PRODUTO'])}" != item.get("snapshot"):
+            if row and f"{pick_gtin(row)}|{_text(row['PRODUTO'])}" != item.get("snapshot"):
                 self.process(item)
                 done.append(item["code"])
         if self.clock() - self._last_override >= OVERRIDE_CHECK_SECONDS:
@@ -406,10 +445,14 @@ class EnrichmentWorker:
             self.check_overrides()
         return done
 
+    def poll_seconds(self) -> float:
+        """Adaptive: the SGV is never loaded by more than ~2.5% of one query-time. 3 s on small catalogs, longer when COUNT(*) is slow (huge catalogs); never above 30 s."""
+        return min(30.0, max(POLL_SECONDS, 40 * self._count_ms / 1000))
+
     def run_forever(self, stopper):
         while not stopper.stop:
             try:
                 self.step()
             except Exception:  # noqa: BLE001 - the enrichment thread must never die or affect the SGV
                 logger.exception("enrichment step failed")
-            self.sleep(POLL_SECONDS)
+            self.sleep(self.poll_seconds())
diff --git a/agent/src/gti_fiscal_agent/preflight.py b/agent/src/gti_fiscal_agent/preflight.py
index f84217d..e065b42 100644
--- a/agent/src/gti_fiscal_agent/preflight.py
+++ b/agent/src/gti_fiscal_agent/preflight.py
@@ -9,7 +9,7 @@ import os
 import sqlite3
 from pathlib import Path
 
-AGENT_VERSION_DEFAULT = "1.2.5"
+AGENT_VERSION_DEFAULT = "1.3.0"
 REQUIRED_ENV = ("GTI_HUB_URL", "AGENT_TOKEN", "FIREBIRD_DATABASE_PATH", "FB_USER", "FB_PASSWORD")
 REQUIRED_CAPABILITIES = ("supports_ncm", "supports_ipi", "supports_cest")
 FISCAL_MODES = ("DRY_RUN", "AUTO_APPLY")
diff --git a/agent/tests/test_enrichment.py b/agent/tests/test_enrichment.py
index cfdc8cf..1df92e3 100644
--- a/agent/tests/test_enrichment.py
+++ b/agent/tests/test_enrichment.py
@@ -66,7 +66,7 @@ DECISIONS = {"CLASSIFICACAO_FISCAL": {"status": "AUTO_FILL_SAFE", "value": "2202
 @pytest.fixture()
 def env(tmp_path):
     db = sqlite3.connect(":memory:", check_same_thread=False)
-    db.execute("CREATE TABLE C000025(CODIGO TEXT PRIMARY KEY, PRODUTO TEXT, GTIN TEXT, CLASSIFICACAO_FISCAL TEXT, CEST TEXT, IPI REAL, OTHER TEXT)")
+    db.execute("CREATE TABLE C000025(CODIGO TEXT PRIMARY KEY, PRODUTO TEXT, GTIN TEXT, CLASSIFICACAO_FISCAL TEXT, CEST TEXT, IPI REAL, OTHER TEXT, CODBARRA TEXT)")
     db.executemany("INSERT INTO C000025(CODIGO,PRODUTO,GTIN,OTHER) VALUES(?,?,?,?)", [(f"H{i}", f"HIST {i}", "", "keep") for i in range(50)])
     db.commit()
     hub = Hub(dict(ENABLED), DECISIONS)
@@ -199,3 +199,13 @@ def test_hub_offline_keeps_product_pending_and_retries(env):
     hub.resolve_product = good
     now[0] += 40
     assert worker.step() == ["N6"] and row(db, "N6")[0] == "22021000"
+
+
+def test_gtin_check_parity_with_the_hub_and_codbarra_is_used():
+    sys.path.insert(0, str(ROOT))
+    from central.app.enrichment.gtin import normalize_gtin
+    for value in ("7891000100103", "7891000100104", "", "0", "12345678", "40170725", "036000291452", "11111111", "78910001001030", "abc", "2000000000008"):
+        assert E.gtin_check_ok(value) == (normalize_gtin(value)["status"] in ("VALID_GTIN", "RESTRICTED_GTIN")), value
+    assert E.pick_gtin({"GTIN": "", "CODBARRA": "7891000100103"}) == "7891000100103"
+    assert E.pick_gtin({"GTIN": "7891000100103", "CODBARRA": "999"}) == "7891000100103"
+    assert E.pick_gtin({"GTIN": "", "CODBARRA": "123"}) == "123" and E.pick_gtin({}) == ""
diff --git a/central/app/enrichment/observability.py b/central/app/enrichment/observability.py
new file mode 100644
index 0000000..adb815c
--- /dev/null
+++ b/central/app/enrichment/observability.py
@@ -0,0 +1,82 @@
+"""Read-only views for the panel: product sources status, official dataset freshness and enrichment coverage metrics. Never exposes tokens."""
+from __future__ import annotations
+
+from datetime import datetime, timezone
+
+from sqlalchemy import func, select
+
+from ..config import settings
+from ..datasets.updater import last_runs
+from ..models import Dataset, GtinLookupCache, ProductEnrichment, ProviderState, ProviderUsageDaily
+from .providers import build_providers
+
+STATUS_LABELS = {"ONLINE": "Online", "DEGRADED": "Instável", "RATE_LIMITED": "Limite atingido", "AUTH_ERROR": "Credencial recusada", "OFFLINE": "Indisponível", "NOT_CONFIGURED": "Não configurado"}
+OFFICIAL_KINDS = (("NCM", "NCM (Siscomex)"), ("TIPI_IPI", "TIPI / IPI"), ("CEST", "CEST (CONFAZ)"), ("BA_ST_MVA", "ST/MVA Bahia"))
+
+
+def _aware(dt):
+    return dt.replace(tzinfo=timezone.utc) if dt is not None and dt.tzinfo is None else dt
+
+
+def provider_sources(session, providers=None) -> list[dict]:
+    providers = providers or build_providers(settings)
+    today = datetime.now(timezone.utc).strftime("%Y-%m-%d")
+    rows = []
+    for provider in providers:
+        state = session.get(ProviderState, provider.name)
+        usage = session.scalar(select(ProviderUsageDaily).where(ProviderUsageDaily.provider == provider.name, ProviderUsageDaily.day == today))
+        requests = usage.requests if usage else 0
+        lookups = (usage.cache_hits + usage.cache_misses) if usage else 0
+        status = ("NOT_CONFIGURED" if not provider.configured else (state.status if state else "ONLINE"))
+        rows.append({
+            "provider": provider.name, "configured": bool(provider.configured), "status": status, "status_label": STATUS_LABELS.get(status, status),
+            "used": requests, "limit": provider.daily_limit or (state.daily_limit if state else None), "rate_per_minute": provider.rate_per_minute,
+            "remaining": state.remaining if state and state.remaining is not None else ((provider.daily_limit - requests) if provider.daily_limit else None),
+            "quota_kind": state.quota_kind if state else ("LOCAL_ESTIMATE" if provider.daily_limit else "DOCUMENTED_RATE"),
+            "reset_at": _aware(state.reset_at) if state and state.reset_at else None,
+            "last_success_at": _aware(state.last_success_at) if state and state.last_success_at else None,
+            "last_error": state.last_error if state else None, "last_error_at": _aware(state.last_error_at) if state and state.last_error_at else None,
+            "latency_ms": state.last_latency_ms if state else None,
+            "cache_hit_rate": round(100 * usage.cache_hits / lookups) if usage and lookups else None,
+            "errors": usage.errors if usage else 0, "rate_limited": usage.rate_limited if usage else 0,
+        })
+    return rows
+
+
+def official_sources(session) -> list[dict]:
+    out = []
+    for kind, label in OFFICIAL_KINDS:
+        ds = session.scalar(select(Dataset).where(Dataset.kind == kind, Dataset.status == "ACTIVE").order_by(Dataset.id.desc()).limit(1))
+        item = {"kind": kind, "label": label, "active": ds is not None, "records": ds.record_count if ds else None, "version": ds.version_label if ds else None,
+                "sha256": (ds.sha256[:12] if ds else None), "last_update": _aware(ds.downloaded_at) if ds else None, "status": "ATUALIZADA" if ds else "SEM DADOS", "error": None, "auto_refresh": kind == "NCM"}
+        if kind == "NCM":
+            recent = last_runs(session, "NCM", 1)
+            if recent:
+                last = recent[0]
+                item["last_check"], item["last_run_status"] = _aware(last.finished_at or last.started_at), last.status
+                if last.status == "FAILED":
+                    item["status"], item["error"] = "FALHA NA ÚLTIMA ATUALIZAÇÃO (dados anteriores mantidos)", last.error
+        out.append(item)
+    return out
+
+
+def coverage_metrics(session) -> dict:
+    total = session.scalar(select(func.count()).select_from(ProductEnrichment)) or 0
+    by = dict(session.execute(select(ProductEnrichment.status, func.count()).group_by(ProductEnrichment.status)).all())
+    with_gtin = session.scalar(select(func.count()).select_from(ProductEnrichment).where(ProductEnrichment.gtin.is_not(None))) or 0
+    rows = [r.decisions or {} for r in session.scalars(select(ProductEnrichment).limit(5000))]
+    states = {}
+    for d in rows:
+        states[d.get("state")] = states.get(d.get("state"), 0) + 1
+
+    def rate(n):
+        return round(100 * n / total, 1) if total else None
+    auto_ncm = sum(1 for d in rows if ((d.get("fields") or {}).get("CLASSIFICACAO_FISCAL") or {}).get("status") == "AUTO_FILL_SAFE")
+    usage = session.execute(select(func.sum(ProviderUsageDaily.requests), func.sum(ProviderUsageDaily.cache_hits), func.sum(ProviderUsageDaily.cache_misses), func.sum(ProviderUsageDaily.rate_limited))).one()
+    requests, hits, misses, limited = (int(x or 0) for x in usage)
+    return {"tracked": total, "GTIN_VALID_RATE": rate(with_gtin), "KB_HIT_RATE": rate(states.get("EXACT_GTIN_VERIFIED", 0) + states.get("EXACT_KB_NAME_VERIFIED", 0)),
+            "EXTERNAL_LOOKUP_RATE": round(100 * requests / (requests + hits), 1) if (requests + hits) else None,
+            "PRODUCT_FOUND_RATE": rate(total - states.get("NOT_FOUND", 0) - states.get("INVALID_GTIN", 0) - states.get("AMBIGUOUS", 0)),
+            "NCM_AUTO_FILL_SAFE_RATE": rate(auto_ncm), "NAME_ONLY_RESOLUTION_RATE": rate(states.get("EXACT_KB_NAME_VERIFIED", 0)), "AMBIGUOUS_RATE": rate(states.get("AMBIGUOUS", 0)),
+            "CONFLICT_RATE": rate(states.get("CONFLICT", 0)), "PROVIDER_QUOTA_BLOCK_RATE": round(100 * limited / requests, 1) if requests else None, "states": states, "by_status": by,
+            "provider_cache_entries": session.scalar(select(func.count()).select_from(GtinLookupCache)) or 0, "provider_cache_hit_rate": round(100 * hits / (hits + misses), 1) if (hits + misses) else None}
diff --git a/central/app/enrichment/observations.py b/central/app/enrichment/observations.py
new file mode 100644
index 0000000..6853b63
--- /dev/null
+++ b/central/app/enrichment/observations.py
@@ -0,0 +1,27 @@
+"""ERP observations -> KB evidence. A client observation is EVIDENCE of family CLIENT_OBSERVATION; it never becomes VERIFIED_KB by itself
+(needs the official dataset AND a second independent family, or >= 2 distinct clients + another family - see kb.evaluate)."""
+from __future__ import annotations
+
+from ..classification import normalize_code
+from . import kb
+from .gtin import normalize_gtin
+
+
+def observe(session, rows, client_label: str, is_official) -> dict:
+    """rows: iterable of {"GTIN","PRODUTO","NCM"}. Only VALID_GTIN rows with an OFFICIAL 8-digit NCM are recorded. Returns counters."""
+    stats = {"rows": 0, "valid_gtin": 0, "with_official_ncm": 0, "recorded": 0}
+    for row in rows:
+        stats["rows"] += 1
+        g = normalize_gtin(row.get("GTIN"))
+        if g["status"] != "VALID_GTIN":
+            continue
+        stats["valid_gtin"] += 1
+        ncm = normalize_code(row.get("NCM"))
+        if len(ncm) != 8 or not is_official(ncm):
+            continue
+        stats["with_official_ncm"] += 1
+        product = kb.upsert_product(session, g["key"], row.get("PRODUTO"))
+        kb.add_evidence(session, product, "ncm", ncm, "client:observation", family=kb.CLIENT_FAMILY, client_label=client_label)
+        stats["recorded"] += 1
+    session.flush()
+    return stats
diff --git a/central/app/enrichment/providers.py b/central/app/enrichment/providers.py
index 20d73b8..9f9c3cd 100644
--- a/central/app/enrichment/providers.py
+++ b/central/app/enrichment/providers.py
@@ -138,9 +138,10 @@ class OpenFoodFacts(Provider):
     name = "openfoodfacts"
     supports_search = True
     rate_per_minute = 15
+    host = "world.openfoodfacts.org"
 
     def lookup_gtin(self, gtin: str) -> ProviderProduct:
-        response = self._get(f"https://world.openfoodfacts.org/api/v2/product/{gtin}", params={"fields": "code,product_name,brands,quantity,categories,generic_name"})
+        response = self._get(f"https://{self.host}/api/v2/product/{gtin}", params={"fields": "code,product_name,brands,quantity,categories,generic_name"})
         headers = {k.lower(): v for k, v in response.headers.items()}
         if response.status_code == 404:
             return ProviderProduct(self.name, gtin, False, latency_ms=getattr(response, "latency_ms", 0), headers=headers)
@@ -166,6 +167,19 @@ class OpenFoodFacts(Provider):
                 for p in (response.json().get("products") or [])[:5] if p.get("product_name")]
 
 
+class OpenProductsFacts(OpenFoodFacts):
+    """Same open (ODbL) platform and API as Open Food Facts, for non-food goods. Identity only (no NCM)."""
+    name = "openproductsfacts"
+    host = "world.openproductsfacts.org"
+    supports_search = False
+
+
+class OpenBeautyFacts(OpenFoodFacts):
+    name = "openbeautyfacts"
+    host = "world.openbeautyfacts.org"
+    supports_search = False
+
+
 class Cosmos(Provider):
     name = "cosmos"
     supports_search = True
@@ -212,8 +226,8 @@ class Cosmos(Provider):
 
 
 def build_providers(settings, client: httpx.Client | None = None) -> list[Provider]:
-    """Fixed order chosen from the benchmark (docs/product_enrichment.md): Cosmos (fiscal-capable) first when configured, then UPCitemdb, then Open Food Facts. Paid providers do not exist here."""
-    providers: list[Provider] = [Cosmos(client, getattr(settings, "cosmos_token", None)), UpcItemDb(client), OpenFoodFacts(client)]
+    """Fixed order chosen from the benchmark (docs/product_enrichment.md): Cosmos (fiscal-capable) first when configured, then UPCitemdb, then the Open *Facts family (food, general products, beauty; identity only). Paid providers do not exist here."""
+    providers: list[Provider] = [Cosmos(client, getattr(settings, "cosmos_token", None)), UpcItemDb(client), OpenFoodFacts(client), OpenProductsFacts(client), OpenBeautyFacts(client)]
     if any(p.paid for p in providers) and not getattr(settings, "allow_paid_providers", False):
         raise RuntimeError("paid providers are not allowed (ALLOW_PAID_PROVIDERS=false)")
     return providers
diff --git a/central/app/enrichment/resolver.py b/central/app/enrichment/resolver.py
index d29153e..689b1bb 100644
--- a/central/app/enrichment/resolver.py
+++ b/central/app/enrichment/resolver.py
@@ -45,6 +45,17 @@ def ncm_is_official(session, code: str, today: date | None = None) -> bool:
     return False
 
 
+class _ScopedDatasets(DatasetService):
+    """The Classifier only needs the ONE exact NCM row here; loading the whole 15k-row catalogue on every product would cost ~300 ms of the 5 s budget."""
+
+    def __init__(self, session, code: str):
+        super().__init__(session)
+        self.code = code
+
+    def active_entries(self):
+        return self.session.scalars(select(NcmEntry).join(Dataset).where(Dataset.kind == "NCM", Dataset.status == "ACTIVE", NcmEntry.code == self.code)).all()
+
+
 def dataset_versions(session) -> dict:
     return {d.kind: {"sha256": (d.sha256 or "")[:12], "version": d.version_label} for d in session.scalars(select(Dataset).where(Dataset.status == "ACTIVE"))}
 
@@ -75,11 +86,13 @@ def _fiscal_decisions(session, ncm_eval: dict, gtin_kb_verified: bool, current:
         for name in ("CEST", "IPI"):
             out[name] = _field(ALREADY if not _empty(current, name) else INSUFFICIENT, reason="depende de um NCM oficial válido")
         return out
-    decision = Classifier(DatasetService(session)).classify({"current_ncm": ncm_for_engine, "description": description})
+    # CEST / IPI are deterministic functions of an OFFICIAL NCM: the NCM we are about to write (AUTO) or one already in the product (validated above against the official dataset)
+    ncm_solid = ncm["status"] == AUTO or (ncm["status"] == ALREADY and ncm_is_official(session, ncm_for_engine))
+    decision = Classifier(_ScopedDatasets(session, ncm_for_engine)).classify({"current_ncm": ncm_for_engine, "description": description})
     cest = (decision["classification"].get("cest") or [])
     if not _empty(current, "CEST"):
         out["CEST"] = _field(ALREADY, reason="campo já preenchido")
-    elif decision["cest_status"] == "CEST_CONFIRMED" and len(cest) == 1 and ncm["status"] == AUTO:
+    elif decision["cest_status"] == "CEST_CONFIRMED" and len(cest) == 1 and ncm_solid:
         out["CEST"] = _field(AUTO, normalize_code(cest[0]), "CEST_OFFICIAL_UNIQUE_MAPPING", ncm["kb_status"], [f"NCM {ncm_for_engine}"])
     elif decision["cest_status"] == "NO_CEST_MAPPING_FOUND":
         out["CEST"] = _field(NA, reason="NCM sem mapeamento de CEST no dataset oficial")
@@ -88,7 +101,7 @@ def _fiscal_decisions(session, ncm_eval: dict, gtin_kb_verified: bool, current:
     ipi = decision["classification"].get("ipi") or {}
     if not _empty(current, "IPI"):
         out["IPI"] = _field(ALREADY, reason="campo já preenchido")
-    elif decision["ipi_status"] == "IPI_RESOLVED" and ipi.get("rate_kind") == "NUMERIC" and ncm["status"] == AUTO:
+    elif decision["ipi_status"] == "IPI_RESOLVED" and ipi.get("rate_kind") == "NUMERIC" and ncm_solid:
         out["IPI"] = _field(AUTO, str(ipi["rate_numeric"]), "IPI_TIPI_NUMERIC", ncm["kb_status"], [f"TIPI NCM {ncm_for_engine}"])
     elif decision["ipi_status"] in {"IPI_RESOLVED"}:
         out["IPI"] = _field(NA, reason="TIPI sem alíquota numérica (NT/branca)")
diff --git a/central/app/version.py b/central/app/version.py
index 3997907..40155f6 100644
--- a/central/app/version.py
+++ b/central/app/version.py
@@ -4,5 +4,5 @@ kept identical to this) and what an AgentRelease row's `version` column holds
 once published — never a second, independent version number. Fiscal semantics
 are versioned separately (ENGINE_VERSION/DECISION_VERSION in classification.py,
 DECISION_ENGINE_VERSION etc.) and are NOT changed by a release bump."""
-GTI_FISCAL_VERSION = "1.2.5"
+GTI_FISCAL_VERSION = "1.3.0"
 API_COMPAT_MAJOR_MINOR = "1.0"  # an Agent 1.0.x/1.1.x talks to any Hub 1.0.x
diff --git a/central/app/web.py b/central/app/web.py
index 59fa081..e76ed37 100644
--- a/central/app/web.py
+++ b/central/app/web.py
@@ -22,6 +22,7 @@ from .panel_presentation import format_ui_datetime
 from .models import AdminUser, Agent, AgentEnrollment, AgentRelease, AgentSnapshot, BlockedProduct, FiscalContextEntry, Installation, Job, ProductEnrichment, ProductFieldAudit, RtcCandidate, RtcWriteAudit, Tenant
 from .jobs import create_job
 from .audit import record_audit
+from .enrichment.observability import coverage_metrics, official_sources, provider_sources
 
 
 def router(session_factory):
@@ -160,7 +161,7 @@ def router(session_factory):
     def dashboard(request: Request):
         session, auth_data, redirect = page_auth(request)
         if redirect: return redirect
-        try: return render(request, "dashboard.html", {"summary": dashboard_summary(session), "auth": auth_data})
+        try: return render(request, "dashboard.html", {"summary": dashboard_summary(session), "auth": auth_data, "sources": provider_sources(session), "official": official_sources(session), "enrich_metrics": coverage_metrics(session)})
         finally: session.close()
 
     @routes.get("/agents", response_class=HTMLResponse)
diff --git a/central/pyproject.toml b/central/pyproject.toml
index 857c5a1..aa31279 100644
--- a/central/pyproject.toml
+++ b/central/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "gti-fiscal-hub"
-version = "1.2.5"
+version = "1.3.0"
 requires-python = ">=3.10"
 dependencies = [
   "fastapi>=0.115,<1",
diff --git a/central/templates/dashboard.html b/central/templates/dashboard.html
index 744f9f4..48eaabb 100644
--- a/central/templates/dashboard.html
+++ b/central/templates/dashboard.html
@@ -12,6 +12,23 @@
   </a>
   {% endfor %}
 </section>
+<section class="split overview-split" id="fontes">
+  <article class="panel">
+    <div class="panel-title"><div><h2>Fontes de produtos</h2><p class="muted">Consultas feitas somente pelo Hub, dentro do limite gratuito.</p></div></div>
+    {% for src in sources %}<div class="list-row"><span><b>{{ src.provider }}</b> <em class="badge {{ '' if src.status == 'ONLINE' else 'critical' }}">{{ src.status_label }}</em>
+      <small>{% if src.limit %}{{ src.used }} / {{ src.limit }} consultas hoje · restam {{ src.remaining if src.remaining is not none else "—" }}{% if src.reset_at %} · renova {{ src.reset_at.strftime("%d/%m %H:%M") }} UTC{% endif %}{% elif src.rate_per_minute %}{{ src.rate_per_minute }} consultas/min (limite documentado) · {{ src.used }} hoje{% else %}{{ src.used }} consultas hoje{% endif %}
+      {% if src.latency_ms %} · {{ src.latency_ms }} ms{% endif %}{% if src.cache_hit_rate is not none %} · cache {{ src.cache_hit_rate }}%{% endif %}
+      {% if src.last_success_at %} · último sucesso {{ src.last_success_at.strftime("%d/%m %H:%M") }}{% endif %}{% if src.last_error %} · último erro: {{ src.last_error }}{% endif %}
+      · configurado: {{ "sim" if src.configured else "não" }}</small></span></div>{% endfor %}
+  </article>
+  <article class="panel">
+    <div class="panel-title"><div><h2>Fontes oficiais</h2><p class="muted">Bases fiscais usadas para validar qualquer preenchimento.</p></div></div>
+    {% for o in official %}<div class="list-row"><span><b>{{ o.label }}</b> <em class="badge {{ '' if o.active and not o.error else 'critical' }}">{{ o.status }}</em>
+      <small>{% if o.active %}{{ o.records }} registros · {{ o.version }} · {{ o.sha256 }}{% if o.last_update %} · baixada {{ o.last_update.strftime("%d/%m/%Y") }}{% endif %}{% endif %}{% if o.auto_refresh %} · atualização automática diária{% endif %}{% if o.error %} · {{ o.error }}{% endif %}</small></span></div>{% endfor %}
+  </article>
+</section>
+<section class="panel" id="enriquecimento-resumo"><div class="panel-title"><div><h2>Enriquecimento de produtos novos</h2><p class="muted">Cobertura medida sobre os produtos novos acompanhados (o catálogo histórico não entra).</p></div></div>
+{% if enrich_metrics.tracked %}<div class="coverage">{% for key,label in [("GTIN_VALID_RATE","Com GTIN válido"),("KB_HIT_RATE","Resolvidos pela base"),("PRODUCT_FOUND_RATE","Produto identificado"),("NCM_AUTO_FILL_SAFE_RATE","NCM preenchível com segurança"),("AMBIGUOUS_RATE","Ambíguos"),("CONFLICT_RATE","Em conflito")] %}<div class="coverage-item"><strong>{{ enrich_metrics[key] if enrich_metrics[key] is not none else "—" }}{% if enrich_metrics[key] is not none %}%{% endif %}</strong><span>{{ label }}</span></div>{% endfor %}</div>{% else %}<p class="empty-state">Nenhum produto novo acompanhado ainda.</p>{% endif %}</section>
 <section class="split overview-split">
   <article class="panel">
     <div class="panel-title"><h2>Cobertura dos produtos</h2><a href="/agents?scope=active">Ver servidores</a></div>
diff --git a/central/tests/test_health.py b/central/tests/test_health.py
index ad8085a..e5b6d79 100644
--- a/central/tests/test_health.py
+++ b/central/tests/test_health.py
@@ -17,7 +17,7 @@ def test_health_is_unavailable_when_database_is_down():
 
 def test_version_endpoint_reports_release_and_fiscal_versions():
     body = TestClient(create_app(database_probe=lambda: True)).get("/version").json()
-    assert body["version"] == "1.2.5" and body["api_compat"] == "1.0" and body["engine_version"] and body["decision_version"]
+    assert body["version"] == "1.3.0" and body["api_compat"] == "1.0" and body["engine_version"] and body["decision_version"]
 
 
 def test_production_refuses_the_dev_token(monkeypatch):
diff --git a/central/tests/test_product_enrichment_api.py b/central/tests/test_product_enrichment_api.py
index 0b938ce..074c58a 100644
--- a/central/tests/test_product_enrichment_api.py
+++ b/central/tests/test_product_enrichment_api.py
@@ -94,3 +94,67 @@ def test_features_route_is_admin_only_csrf_and_audited():
         assert "INSTALLATION_FEATURES" in [e.action for e in s.scalars(select(AuditEvent))]
     page = client.get(f"/agents/{agent_id}")
     assert "Funcionalidades" in page.text and "Enriquecimento" in page.text
+
+
+def test_dashboard_shows_sources_without_tokens_and_official_freshness(monkeypatch):
+    from central.app.enrichment import observability, providers
+    monkeypatch.setattr(observability, "build_providers", lambda _s: providers.build_providers(type("S", (), {"cosmos_token": "SEGREDO-NAO-VAZAR", "allow_paid_providers": False})()))
+    client, agent_id, _ = _panel_client()
+    page = client.get("/").text
+    assert "Fontes de produtos" in page and "cosmos" in page and "configurado: sim" in page and "upcitemdb" in page and "openfoodfacts" in page
+    assert "SEGREDO-NAO-VAZAR" not in page and "Fontes oficiais" in page and "NCM (Siscomex)" in page
+
+
+def _second_agent(engine, mode="ENABLED", corrector=True):
+    from central.app.auth import token_hash
+    from central.app.models import Agent, Tenant
+    with Session(engine) as s:
+        first = s.scalars(select(Installation)).first()
+        agent = Agent(tenant_id=first.tenant_id, name="second", token_hash=token_hash("second-token"), enabled=True)
+        s.add(agent)
+        s.flush()
+        s.add(Installation(agent_id=agent.id, tenant_id=first.tenant_id, installation_id="00000000-0000-4000-8000-0000000000bb", display_name="B", features_configured=True, corrector_enabled=corrector, enrichment_mode=mode))
+        s.commit()
+    return {"Authorization": "Bearer second-token"}
+
+
+def test_same_product_in_another_installation_costs_zero_extra_provider_requests(monkeypatch):
+    client, engine = setup(monkeypatch)
+    h2 = _second_agent(engine)
+    providers = enrich_api.PROVIDERS_OVERRIDE
+    first = resolve(client)
+    calls_after_first = [p.calls for p in providers]
+    second = client.post("/v1/products/resolve", json={"product_code": "X9", "gtin": G1, "description": "REFRIGERANTE COLA LATA 350ML", "current": {}}, headers=h2).json()
+    assert [p.calls for p in providers] == calls_after_first                       # other installation: served by KB, zero external calls
+    assert second["state"] == "EXACT_GTIN_VERIFIED" and second["fields"]["CLASSIFICACAO_FISCAL"]["value"] == first["fields"]["CLASSIFICACAO_FISCAL"]["value"]
+    with Session(engine) as s:
+        assert {r.installation_id for r in s.scalars(select(ProductEnrichment))} == {1, 2}
+
+
+def test_installations_are_isolated_by_their_own_feature_flags(monkeypatch):
+    client, engine = setup(monkeypatch, "DISABLED")
+    h2 = _second_agent(engine, "ENABLED")
+    assert resolve(client)["state"] == "FEATURE_DISABLED"
+    body = client.post("/v1/products/resolve", json={"product_code": "N1", "gtin": G1, "description": "REFRIGERANTE COLA LATA 350ML", "current": {}}, headers=h2).json()
+    assert body["state"] != "FEATURE_DISABLED" and body["fields"]["CLASSIFICACAO_FISCAL"]["status"] == "AUTO_FILL_SAFE"
+
+
+def test_quota_exhausted_gives_pending_state_and_retry_time_through_the_api(monkeypatch):
+    client, engine = setup(monkeypatch)
+    monkeypatch.setattr(enrich_api, "PROVIDERS_OVERRIDE", [Fake("cosmos", COSMOS(), True, daily_limit=1)])
+    resolve(client, "A1", G1)
+    body = resolve(client, "A2", G4, "OUTRO PRODUTO")
+    assert body["state"] == "PENDING_PROVIDER_QUOTA" and body["retry_at"]
+    with Session(engine) as s:
+        row = s.scalars(select(ProductEnrichment).where(ProductEnrichment.product_code == "A2")).one()
+        assert row.status == "PENDING" and row.next_attempt_at is not None
+        assert s.scalars(select(BlockedProduct).where(BlockedProduct.reason_code == "PENDING_PROVIDER_QUOTA")).one().field == "ENRICHMENT"
+
+
+def test_corrector_switch_off_stops_jobs_but_not_enrichment(monkeypatch):
+    client, engine = setup(monkeypatch)
+    with Session(engine) as s:
+        s.scalars(select(Installation)).one().corrector_enabled = False
+        s.commit()
+    assert client.get("/v1/jobs/next", headers=H).json() is None
+    assert resolve(client)["state"] != "FEATURE_DISABLED"
diff --git a/central/tests/test_product_enrichment_resolver.py b/central/tests/test_product_enrichment_resolver.py
index d793446..33d5ee8 100644
--- a/central/tests/test_product_enrichment_resolver.py
+++ b/central/tests/test_product_enrichment_resolver.py
@@ -223,7 +223,7 @@ def test_cosmos_without_token_is_not_configured_and_costs_nothing(env):
 
 def test_paid_provider_guard():
     class S: allow_paid_providers = False; cosmos_token = None
-    assert [p.name for p in build_providers(S())] == ["cosmos", "upcitemdb", "openfoodfacts"]
+    assert [p.name for p in build_providers(S())] == ["cosmos", "upcitemdb", "openfoodfacts", "openproductsfacts", "openbeautyfacts"]
 
 
 # ---- provider contract tests against recorded fixtures (httpx.MockTransport, no network) ----------------------------------------------------------
diff --git a/deploy/agent-windows/agent.env.template b/deploy/agent-windows/agent.env.template
index c351d3f..620f416 100644
--- a/deploy/agent-windows/agent.env.template
+++ b/deploy/agent-windows/agent.env.template
@@ -19,4 +19,4 @@ FB_PASSWORD=
 FISCAL_MODE=DRY_RUN
 FIREBIRD_WRITES=0
 POLL_INTERVAL_SECONDS=5
-AGENT_VERSION=1.2.4
+AGENT_VERSION=1.3.0
diff --git a/deploy/agent-windows/gti-agent.ps1 b/deploy/agent-windows/gti-agent.ps1
index 9ab8820..fb9546a 100644
--- a/deploy/agent-windows/gti-agent.ps1
+++ b/deploy/agent-windows/gti-agent.ps1
@@ -164,7 +164,7 @@ switch ($Command) {
                 "GTI_HUB_URL=$hub", "AGENT_TOKEN=$token", "AGENT_ID=$agentId",
                 "COMPANY_CODE=$code", "COMPANY_CNPJ=$cnpj", "COMPANY_CRT_SOURCE=$crtSource",
                 "FIREBIRD_DATABASE_PATH=$db", "FB_USER=$fbUser", "FB_PASSWORD=$fbPass",
-                'FISCAL_MODE=DRY_RUN', 'FIREBIRD_WRITES=0', 'POLL_INTERVAL_SECONDS=5', 'AGENT_VERSION=1.2.4'
+                'FISCAL_MODE=DRY_RUN', 'FIREBIRD_WRITES=0', 'POLL_INTERVAL_SECONDS=5', 'AGENT_VERSION=1.3.0'
             )
             Set-Content -Path $EnvFile -Value $lines -Encoding UTF8
         }
diff --git a/deploy/agent-windows/pilot-construcao.env.example b/deploy/agent-windows/pilot-construcao.env.example
index 9854140..55f8b42 100644
--- a/deploy/agent-windows/pilot-construcao.env.example
+++ b/deploy/agent-windows/pilot-construcao.env.example
@@ -19,4 +19,4 @@ FB_PASSWORD=
 FISCAL_MODE=DRY_RUN
 FIREBIRD_WRITES=0
 POLL_INTERVAL_SECONDS=5
-AGENT_VERSION=1.2.4
+AGENT_VERSION=1.3.0
diff --git a/deploy/agent-windows/profiles/mmmodas.env.example b/deploy/agent-windows/profiles/mmmodas.env.example
index d0c5f1a..2989cd6 100644
--- a/deploy/agent-windows/profiles/mmmodas.env.example
+++ b/deploy/agent-windows/profiles/mmmodas.env.example
@@ -19,4 +19,4 @@ FB_PASSWORD=
 FISCAL_MODE=DRY_RUN
 FIREBIRD_WRITES=0
 POLL_INTERVAL_SECONDS=5
-AGENT_VERSION=1.2.4
+AGENT_VERSION=1.3.0
diff --git a/deploy/agent-windows/profiles/rincao.env.example b/deploy/agent-windows/profiles/rincao.env.example
index b8d3e38..4309ad1 100644
--- a/deploy/agent-windows/profiles/rincao.env.example
+++ b/deploy/agent-windows/profiles/rincao.env.example
@@ -19,4 +19,4 @@ FB_PASSWORD=
 FISCAL_MODE=DRY_RUN
 FIREBIRD_WRITES=0
 POLL_INTERVAL_SECONDS=5
-AGENT_VERSION=1.2.4
+AGENT_VERSION=1.3.0
diff --git a/deploy/agent-windows/profiles/uniao.env.example b/deploy/agent-windows/profiles/uniao.env.example
index bce7734..32f99fa 100644
--- a/deploy/agent-windows/profiles/uniao.env.example
+++ b/deploy/agent-windows/profiles/uniao.env.example
@@ -19,4 +19,4 @@ FB_PASSWORD=
 FISCAL_MODE=DRY_RUN
 FIREBIRD_WRITES=0
 POLL_INTERVAL_SECONDS=5
-AGENT_VERSION=1.2.4
+AGENT_VERSION=1.3.0
diff --git a/deploy/hub/.env.example b/deploy/hub/.env.example
index 4ff0a0f..9550e08 100644
--- a/deploy/hub/.env.example
+++ b/deploy/hub/.env.example
@@ -1,5 +1,5 @@
 # GTI Fiscal Hub — production environment (copy to .env, never commit .env)
-GTI_FISCAL_VERSION=1.2.4
+GTI_FISCAL_VERSION=1.3.0
 
 # REQUIRED. URL-safe secret (used inside a connection URL): openssl rand -hex 24
 POSTGRES_PASSWORD=
diff --git a/deploy/hub/docker-compose.prod.yml b/deploy/hub/docker-compose.prod.yml
index c361929..bf386dc 100644
--- a/deploy/hub/docker-compose.prod.yml
+++ b/deploy/hub/docker-compose.prod.yml
@@ -36,7 +36,7 @@ services:
     build:
       context: ../..
       dockerfile: central/Dockerfile
-    image: gti-fiscal-hub:${GTI_FISCAL_VERSION:-1.2.4}
+    image: gti-fiscal-hub:${GTI_FISCAL_VERSION:-1.3.0}
     restart: unless-stopped
     depends_on:
       postgres:
diff --git a/docs/PRODUCT_NEW_DETECTION_DISCOVERY_V1.json b/docs/PRODUCT_NEW_DETECTION_DISCOVERY_V1.json
new file mode 100644
index 0000000..21d9e62
--- /dev/null
+++ b/docs/PRODUCT_NEW_DETECTION_DISCOVERY_V1.json
@@ -0,0 +1,119 @@
+{
+ "artifact": "PRODUCT_NEW_DETECTION_DISCOVERY_V1",
+ "table": "C000025",
+ "decision": "SELECT COUNT(*) poll (adaptive interval) + CODIGO-set diff only when the count moves (+ 60 s safety diff). No trigger/table/procedure is created in the SGV database.",
+ "why": "DATA_CADASTRO/ULTIMA_ALTERACAO are NOT indexed (a filter would scan the table like COUNT does); CODIGO is the PK; the ERP already owns triggers C000025_INS/UPD/DEL that must not be touched.",
+ "databases": {
+  "CONSTRUCAO": {
+   "rows": 3712,
+   "count_ms_p50": 2.2,
+   "code_list_ms": 62.9,
+   "max_codigo_ms": 5.5,
+   "pk": "PK_C000025(CODIGO)",
+   "indexed_columns": [
+    "CCLASSTRIB",
+    "CODIGO",
+    "PRODUTO"
+   ],
+   "sgv_triggers_present": [
+    "C000025_DEL",
+    "C000025_INS",
+    "C000025_UPD",
+    "CODBARRA_CODPRO"
+   ],
+   "date_columns_unindexed": [
+    "ULTIMA_ALTERACAO",
+    "DATA_CADASTRO"
+   ]
+  },
+  "MMMODAS": {
+   "rows": 20997,
+   "count_ms_p50": 9.0,
+   "code_list_ms": 241.1,
+   "max_codigo_ms": 24.3,
+   "pk": "PK_C000025(CODIGO)",
+   "indexed_columns": [
+    "CCLASSTRIB",
+    "CODIGO",
+    "PRODUTO"
+   ],
+   "sgv_triggers_present": [
+    "C000025_DEL",
+    "C000025_INS",
+    "C000025_UPD",
+    "CODBARRA_CODPRO"
+   ],
+   "date_columns_unindexed": [
+    "ULTIMA_ALTERACAO",
+    "DATA_CADASTRO"
+   ]
+  },
+  "RINCAO": {
+   "rows": 894,
+   "count_ms_p50": 1.0,
+   "code_list_ms": 15.6,
+   "max_codigo_ms": 4.0,
+   "pk": "PK_C000025(CODIGO)",
+   "indexed_columns": [
+    "CCLASSTRIB",
+    "CODIGO",
+    "PRODUTO"
+   ],
+   "sgv_triggers_present": [
+    "C000025_DEL",
+    "C000025_INS",
+    "C000025_UPD",
+    "CODBARRA_CODPRO"
+   ],
+   "date_columns_unindexed": [
+    "ULTIMA_ALTERACAO",
+    "DATA_CADASTRO"
+   ]
+  },
+  "UNIAO": {
+   "rows": 8942,
+   "count_ms_p50": 4.3,
+   "code_list_ms": 119.2,
+   "max_codigo_ms": 11.4,
+   "pk": "PK_C000025(CODIGO)",
+   "indexed_columns": [
+    "CCLASSTRIB",
+    "CODBARRA",
+    "CODIGO",
+    "PRODUTO"
+   ],
+   "sgv_triggers_present": [
+    "C000025_DEL",
+    "C000025_INS",
+    "C000025_UPD",
+    "CODBARRA_CODPRO"
+   ],
+   "date_columns_unindexed": [
+    "ULTIMA_ALTERACAO",
+    "DATA_CADASTRO"
+   ]
+  },
+  "TESTESOFICIAIS": {
+   "rows": 521,
+   "count_ms_p50": 0.9,
+   "code_list_ms": 8.0,
+   "max_codigo_ms": 2.2,
+   "pk": "PK_C000025(CODIGO)",
+   "indexed_columns": [
+    "CCLASSTRIB",
+    "CODIGO",
+    "PRODUTO"
+   ],
+   "sgv_triggers_present": [
+    "C000025_DEL",
+    "C000025_INS",
+    "C000025_UPD",
+    "CODBARRA_CODPRO"
+   ],
+   "date_columns_unindexed": [
+    "ULTIMA_ALTERACAO",
+    "DATA_CADASTRO"
+   ]
+  }
+ }
+}
\ No newline at end of file
diff --git a/docs/PRODUCT_PROVIDER_BENCHMARK_V1.json b/docs/PRODUCT_PROVIDER_BENCHMARK_V1.json
new file mode 100644
index 0000000..b41c995
--- /dev/null
+++ b/docs/PRODUCT_PROVIDER_BENCHMARK_V1.json
@@ -0,0 +1,96 @@
+{
+ "measured_at": "2026-09-25T21:53:01Z",
+ "sample_source": "lab ERP databases (read-only), valid GTINs only",
+ "limit_per_provider": 30,
+ "providers": [
+  {
+   "provider": "cosmos",
+   "configured": false,
+   "sample_size": 0,
+   "found": 0,
+   "not_found": 0,
+   "with_description": 0,
+   "with_brand": 0,
+   "with_ncm": 0,
+   "with_cest": 0,
+   "errors": 0,
+   "error_kinds": {},
+   "latencies_ms": [],
+   "status": "NOT_CONFIGURED"
+  },
+  {
+   "provider": "upcitemdb",
+   "configured": true,
+   "sample_size": 6,
+   "found": 1,
+   "not_found": 5,
+   "with_description": 1,
+   "with_brand": 1,
+   "with_ncm": 0,
+   "with_cest": 0,
+   "errors": 1,
+   "error_kinds": {
+    "RATE_LIMITED": 1
+   },
+   "error_rate": 0.143,
+   "latency_p50_ms": 139.5,
+   "latency_p95_ms": 140,
+   "found_rate": 0.167,
+   "status": "MEASURED"
+  },
+  {
+   "provider": "openfoodfacts",
+   "configured": true,
+   "sample_size": 30,
+   "found": 4,
+   "not_found": 26,
+   "with_description": 4,
+   "with_brand": 3,
+   "with_ncm": 0,
+   "with_cest": 0,
+   "errors": 0,
+   "error_kinds": {},
+   "error_rate": 0.0,
+   "latency_p50_ms": 655.0,
+   "latency_p95_ms": 827,
+   "found_rate": 0.133,
+   "status": "MEASURED"
+  },
+  {
+   "provider": "openproductsfacts",
+   "configured": true,
+   "sample_size": 30,
+   "found": 0,
+   "not_found": 30,
+   "with_description": 0,
+   "with_brand": 0,
+   "with_ncm": 0,
+   "with_cest": 0,
+   "errors": 0,
+   "error_kinds": {},
+   "error_rate": 0.0,
+   "latency_p50_ms": 640.0,
+   "latency_p95_ms": 969,
+   "found_rate": 0.0,
+   "status": "MEASURED"
+  },
+  {
+   "provider": "openbeautyfacts",
+   "configured": true,
+   "sample_size": 30,
+   "found": 0,
+   "not_found": 30,
+   "with_description": 0,
+   "with_brand": 0,
+   "with_ncm": 0,
+   "with_cest": 0,
+   "errors": 0,
+   "error_kinds": {},
+   "error_rate": 0.0,
+   "latency_p50_ms": 655.5,
+   "latency_p95_ms": 1311,
+   "found_rate": 0.0,
+   "status": "MEASURED"
+  }
+ ]
+}
\ No newline at end of file
diff --git a/docs/product_enrichment.md b/docs/product_enrichment.md
new file mode 100644
index 0000000..2f81fc6
--- /dev/null
+++ b/docs/product_enrichment.md
@@ -0,0 +1,67 @@
+# GTI Fiscal — Product Enrichment V1 (produto novo → EAN/nome → perfil fiscal → preenchimento seguro)
+
+Living document. Same Hub / FastAPI / PostgreSQL / Agent as the rest of GTI Fiscal: **no new server, service, panel or product**, no change in the SGV, no Firebird trigger/table/procedure.
+
+## Hard rules (verified by tests)
+- `AI_ENABLED=false`, `LLM_ENABLED=false`, `EMBEDDINGS_ENABLED=false`. Everything here is deterministic.
+- External cost R$ 0: only free providers; `ALLOW_PAID_PROVIDERS=false` (a paid provider makes `build_providers` raise). **Agents never call providers – only the Hub.**
+- Per-installation functions live in the Hub (Agent detail → *Funcionalidades*): **Corretor fiscal** on/off and **Enriquecimento** `DISABLED | DRY_RUN | ENABLED`. ADMIN only, CSRF, audited (`INSTALLATION_FEATURES`),
+  typed confirmation `PREENCHER` for ENABLED. A **new** installation is `WAITING_FEATURE_CONFIGURATION` (`features_configured=false`): it heartbeats but `claim_next_job`/`schedule_automatic_sync`
+  hand it nothing and the enrichment worker does zero Firebird work, until features are chosen (at creation in *Nova instalação*, or later). Existing installations keep the previous behaviour (migration 0029).
+- Scope = **new products only**. When the function is first active the Agent stores a baseline of the existing `C000025.CODIGO` set (local SQLite); historical products are never sent anywhere.
+
+## Flow
+1. **Detection (Agent, ≤ 5 s)** – `SELECT COUNT(*) FROM C000025` every 2 s (adaptive up to 30 s if the count is slow); code-set diff only when the count moved (+ a 60 s safety diff). No trigger. Discovery on the five
+   lab databases (`docs/PRODUCT_NEW_DETECTION_DISCOVERY_V1.json`): PK `CODIGO` char(6), `DATA_CADASTRO/ULTIMA_ALTERACAO` **not indexed**, count 1–9 ms, full code list 8–241 ms, the ERP already owns triggers on the table.
+   The EAN typed by the operator is in **`CODBARRA`** (fiscal `GTIN` is mostly empty): the first VALID of `GTIN`/`CODBARRA` is sent.
+2. **Resolve (Hub)** `POST /v1/products/resolve {product_code, gtin, description, current}` → per-field decisions. Hub time budget 4.5 s; beyond it (or on quota/offline) the state is `PENDING_*` and the Agent retries with backoff (30 s … 30 min); results are cached, so the retry is usually free.
+3. **Write (Agent)** typed capability `PRODUCT_ENRICHMENT_WRITE_V1`, allowlist `CLASSIFICACAO_FISCAL, CEST, IPI`; fixed statement `UPDATE C000025 SET f=? WHERE CODIGO=? AND (f IS NULL OR TRIM(f)='')`, one short transaction per field,
+   guarded re-select inside the transaction, post-write verification. Before every field: fresh central authorization (fail closed). `COD_ORIGEM_MERCADORIA` is never written; cClassTrib stays with the existing RTC gates
+   (the Agent only requests `REPROCESS_PRODUCTS` for the product; nothing is decided here).
+4. **Audit** `POST /v1/products/enrichment/applied` → `product_field_audits` (tenant, agent, installation, code, GTIN, description snapshot, field, old/new, rule, KB status, evidence, dataset versions, authorization origin, job, result). No secrets.
+
+## Identity and states
+- GTIN: 8/12/13/14 digits + check digit → `VALID_GTIN | INVALID_GTIN | MISSING_GTIN` (+ `RESTRICTED_GTIN` for store-internal prefixes 02/04/20–29). **GTIN has absolute priority over the description.** An invalid GTIN never reaches a provider.
+- Resolution states (discrete, no scores): `EXACT_GTIN_VERIFIED, EXACT_KB_NAME_VERIFIED, MULTI_SOURCE_MATCH, SINGLE_WEAK_SOURCE, AMBIGUOUS, CONFLICT, NOT_FOUND` (+ operational `INVALID_GTIN, PENDING_PROVIDER_QUOTA, PENDING_PROVIDERS_OFFLINE, FEATURE_DISABLED`).
+- Field decisions: `AUTO_FILL_SAFE | SUGGESTION_ONLY | CONFLICT | INSUFFICIENT_EVIDENCE | ALREADY_SET | NOT_APPLICABLE | HUMAN_OVERRIDDEN`. A field that is not resolved never blocks the others.
+
+## Knowledge base (same PostgreSQL: `kb_products`, `kb_knowledge` with per-field `status`)
+`UNVERIFIED / EVIDENCE_ONLY / VERIFIED_KB / CONFLICT / STALE`. NCM becomes `VERIFIED_KB` only when it is **valid and vigente in the official NCM dataset**, has **≥ 2 independent evidence families** (each provider = 1 family; client observations = 1 family and only when ≥ 2 distinct clients)
+and **no conflicting value**. Client consensus alone is never authority (measured on the lab databases: 241 GTINs shared by ≥ 2 clients, **93 with different NCMs**). Company-specific decisions (CSOSN, CFOP, …) never enter the KB.
+**Human edits never train the KB.** `NCM` auto-fill: only `VERIFIED_KB`; a single non-official source is `SUGGESTION_ONLY`; conflicts write nothing.
+CEST and IPI are deterministic functions of an official NCM through the existing engine (`Classifier`): they are filled when the NCM is being written (AUTO) **or is already present and official**, only for a unique official CEST mapping / a numeric TIPI rate.
+Name-only: only an exact fingerprint (sorted normalized tokens, quantity/unit normalized) of a **specific** description (≥ 2 informative tokens + quantity, or ≥ 4) that maps to a unique `VERIFIED_KB` product may fill; generic names never do.
+ERP observations (`scripts/enrichment_lab_kb.py`, read-only) enter as `CLIENT_OBSERVATION` evidence only.
+
+## Human override protection
+After writing, the Agent watches the value. Any value different from what it wrote — **including a deleted value** — marks the field `HUMAN_OVERRIDDEN` (local SQLite + reported to the Hub). The Hub then answers `HUMAN_OVERRIDDEN` for that field, the enrichment never rewrites it,
+and `enrichment.is_human_overridden(state_path, code, field)` is the hook every automatic writer must call. No write loops: second run = 0 writes.
+
+## Providers (Hub only; `central/app/enrichment/providers.py`, single interface `lookup_gtin / search_description / health`)
+| Provider | Access | Limit used | Gives |
+|---|---|---|---|
+| Cosmos (Bluesoft) | `COSMOS_TOKEN` in the Hub env (optional) | 25/day | description, brand, **NCM, CEST** — `NOT_CONFIGURED` without a token (not an error) |
+| UPCitemdb (trial) | no key | 100/day, burst 15 / 30 s (headers `X-RateLimit-*`) | title, brand, category |
+| Open Food Facts / Open Products Facts / Open Beauty Facts | open (ODbL), identifiable `User-Agent` | 15 req/min | title, brand, quantity |
+Quota is central (`ProviderState`, `ProviderUsageDaily`, `quota_kind = REMOTE_QUOTA | LOCAL_ESTIMATE | DOCUMENTED_RATE`), with positive (30 d) and negative (24 h) `GtinLookupCache`, single-flight per (provider, GTIN), circuit breaker
+(3 failures → 60 s doubling up to 15 min; 429 → until reset; 401/403 → 1 h). Quota exhausted → `PENDING_PROVIDER_QUOTA` + retry after reset.
+**Benchmark on real lab GTINs** (`docs/PRODUCT_PROVIDER_BENCHMARK_V1.json`): identity found rate is low for Brazilian goods (Open Food Facts 4/30, UPCitemdb rate-limited after 6 calls) and **only Cosmos returns NCM**. Consequently
+NCM auto-fill depends on the KB (Cosmos + client observations) and the product delivers CEST/IPI immediately whenever the operator already typed a valid NCM. Cosmos without a token is not a blocker.
+
+## Official data
+`NCM` (Siscomex JSON) is refreshed daily by `central/app/datasets/updater.py` (staging in memory → schema validation → plausibility ≥ 90 % of the active count → atomic promotion by `DatasetService.import_payload`; on any failure the last valid dataset stays).
+Every attempt is a `source_refresh_runs` row (url, time, sha256, version, count, status, error). TIPI/CEST/CFOP/CST/CSOSN/PIS-COFINS/RTC datasets are reused as they are. The dashboard shows *Fontes oficiais* (records, version, sha, last update, status).
+
+## Panel
+Dashboard: *Fontes de produtos* (status, used/limit/remaining/reset, last success/error, latency, cache-hit rate, `configurado: sim/não` — never a token), *Fontes oficiais*, small *Enriquecimento* coverage.
+Agent detail: *Funcionalidades* (admin) and *Enriquecimento* (counters + product lookup with per-field provenance). Problems appear through the existing blockers (`field=ENRICHMENT`, PT-BR reasons).
+
+## Coverage metrics (`observability.coverage_metrics`)
+`GTIN_VALID_RATE, KB_HIT_RATE, EXTERNAL_LOOKUP_RATE, PRODUCT_FOUND_RATE, NCM_AUTO_FILL_SAFE_RATE, NAME_ONLY_RESOLUTION_RATE, AMBIGUOUS_RATE, CONFLICT_RATE, PROVIDER_QUOTA_BLOCK_RATE` (+ cache stats).
+
+## Evidence
+- Offline tests: `central/tests/test_product_enrichment_resolver.py` (12 resolution scenarios, quota/circuit/cache, provider contract fixtures), `test_product_enrichment_api.py` (feature gates, audit, override, blockers, multi-installation), `test_official_refresh.py`, `agent/tests/test_enrichment.py`.
+- Real Firebird **copy-only** E2E `scripts/enrichment_copy_e2e.py` → `docs/runs/enrichment_copy_e2e.json`: detect→written P50 2.8 s / P95 3.9 s / max 4.1 s; only the 3 allowed columns; other tables changed only by the SGV's own triggers for the test rows;
+  second run 0 writes; DRY_RUN 0 writes; WAITING/DISABLED 0 Firebird activity and no baseline; human override incl. deleted value; Hub offline → pending → recovers.
+- Migration 0029: fresh-vs-upgrade `pg_dump -s` identical; upgrade → downgrade → upgrade clean.
+- Live smoke `scripts/enrichment_live_smoke.py` (1 lookup/provider, real internet, separate from the suite).
diff --git a/docs/runs/enrichment_copy_e2e.json b/docs/runs/enrichment_copy_e2e.json
new file mode 100644
index 0000000..24d9c22
--- /dev/null
+++ b/docs/runs/enrichment_copy_e2e.json
@@ -0,0 +1,47 @@
+{
+ "copy_only": true,
+ "products_enabled": 12,
+ "detect_to_written_seconds": {
+  "p50": 2.835,
+  "p95": 3.89,
+  "max": 4.14,
+  "all": [
+   1.66,
+   2.08,
+   3.89,
+   2.83,
+   2.52,
+   2.84,
+   3.05,
+   2.58,
+   3.31,
+   2.45,
+   3.89,
+   4.14
+  ]
+ },
+ "hub_resolve_calls": 15,
+ "tables_changed": [
+  "C000025",
+  "C000058",
+  "C000100"
+ ],
+ "checks": {
+  "WAITING_NO_BASELINE_NO_CALLS": true,
+  "DISABLED_NO_BASELINE_NO_CALLS": true,
+  "BASELINE_TAKEN_ON_ACTIVATION": true,
+  "DRY_RUN_ZERO_WRITES": true,
+  "FIELDS_WRITTEN": true,
+  "SECOND_RUN_ZERO_WRITES": true,
+  "HISTORICAL_KNOWN_COUNT": true,
+  "OVERRIDE_DELETED_VALUE": true,
+  "OVERRIDE_REPORTED_TO_HUB": true,
+  "OVERRIDE_NEVER_REWRITTEN": true,
+  "HUB_OFFLINE_PENDING_NO_WRITE": true,
+  "HUB_BACK_RECOVERS": true,
+  "OTHER_TABLES_ONLY_ERP_TRIGGER_SIDE_EFFECTS_OF_TEST_ROWS": true,
+  "ORIGIN_NEVER_WRITTEN": true
+ },
+ "ALL_PASS": true,
+ "poll_seconds": 2.0
+}
\ No newline at end of file
diff --git a/docs/runs/enrichment_local_hub_e2e.json b/docs/runs/enrichment_local_hub_e2e.json
new file mode 100644
index 0000000..82f5882
--- /dev/null
+++ b/docs/runs/enrichment_local_hub_e2e.json
@@ -0,0 +1,21 @@
+{
+ "copy_only": true,
+ "hub": "http://127.0.0.1:8000",
+ "detect_to_written_seconds_enabled": 4.22,
+ "checks": {
+  "REGISTERED_DEFAULT_DISABLED": true,
+  "DISABLED_NO_ACTIVITY": true,
+  "BASELINE_ON_ACTIVATION": true,
+  "DRY_RUN_ZERO_WRITES": true,
+  "DRY_RUN_AUDITED_IN_HUB": true,
+  "ENABLED_NCM_WRITTEN": true,
+  "WRITTEN_FIELDS_MATCH_HUB_AUDIT": true,
+  "EARLIER_DRY_RUN_PRODUCT_STILL_UNTOUCHED": true,
+  "UNKNOWN_GTIN_NO_WRITE": true,
+  "UNKNOWN_GTIN_BLOCKER_PT": true,
+  "INVALID_GTIN_NO_PROVIDER_CALL": true,
+  "DISABLED_AGAIN_ZERO_WRITES": true,
+  "CORRECTOR_OFF_NO_JOBS": true
+ },
+ "ALL_PASS": true
+}
\ No newline at end of file
diff --git a/release/RELEASE_MANIFEST.json b/release/RELEASE_MANIFEST.json
index 4f9b5b6..6a34e7e 100644
--- a/release/RELEASE_MANIFEST.json
+++ b/release/RELEASE_MANIFEST.json
@@ -1,14 +1,14 @@
 {
  "artifact": "RELEASE_MANIFEST",
- "GTI_FISCAL_VERSION": "1.2.5",
+ "GTI_FISCAL_VERSION": "1.3.0",
  "generated_at": "2026-09-25T15:43:40+00:00",
  "hub": {
-  "version": "1.2.5",
+  "version": "1.3.0",
   "api_compat": "1.0",
   "engine_version": "0.3.0",
   "decision_version": "mission-04c-ipi-ex-split"
  },
- "agent_version": "1.2.5",
+ "agent_version": "1.3.0",
  "fiscal_engine": {
   "ENGINE_VERSION": "0.3.0",
   "DECISION_VERSION": "mission-04c-ipi-ex-split",
@@ -17,10 +17,10 @@
   "ICMS_CONTEXT_ENGINE_VERSION": "icms-context-v1"
  },
  "migrations": {
-  "alembic_head_in_repo": "0028_rtc_write_modes",
-  "applied_revision": "0028_rtc_write_modes",
-  "applied_revision_source": "production Hub 10.0.10.117 (gti-fiscal-hub-fiscal-hub-1), verified 2026-09-25",
-  "count": 28
+  "alembic_head_in_repo": "0029_product_enrichment",
+  "applied_revision": "0029_product_enrichment",
+  "applied_revision_source": "fresh + upgrade(0028->0029) + downgrade/upgrade verified on the dev PostgreSQL (schema parity by pg_dump -s); production applies 0029 at deploy (alembic upgrade head at container start)",
+  "count": 29
  },
  "datasets_active": [
   {
@@ -90,4 +90,4 @@
   }
  ],
  "note": "A release bump does not change fiscal semantics; engine/decision versions are versioned independently."
-}
+}
\ No newline at end of file
diff --git a/release/VERSION b/release/VERSION
index c813fe1..f0bb29e 100644
--- a/release/VERSION
+++ b/release/VERSION
@@ -1 +1 @@
-1.2.5
+1.3.0
diff --git a/scripts/enrichment_benchmark.py b/scripts/enrichment_benchmark.py
new file mode 100644
index 0000000..bdeb484
--- /dev/null
+++ b/scripts/enrichment_benchmark.py
@@ -0,0 +1,79 @@
+"""Conservative provider benchmark on REAL GTINs taken (read-only) from the lab ERP databases. Free quotas only, paced below documented limits.
+
+    python scripts/enrichment_benchmark.py --sample scratch_enrich/benchmark_gtins.json --out docs/PRODUCT_PROVIDER_BENCHMARK_V1.json [--limit 30]
+
+Each provider gets at most --limit lookups (UPCitemdb: also <= 12 per 30 s burst window; Open Food Facts: <= 12/min). Cosmos is measured only when COSMOS_TOKEN is set.
+Reports sample_size / found / not_found / description / brand / NCM / CEST / latency / error_rate per provider. Secrets are never printed.
+"""
+import argparse
+import json
+import os
+import statistics
+import sys
+import time
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+import httpx  # noqa: E402
+
+from central.app.enrichment.providers import Cosmos, OpenBeautyFacts, OpenFoodFacts, OpenProductsFacts, ProviderError, UpcItemDb, USER_AGENT  # noqa: E402
+
+PACE = {"upcitemdb": 2.6, "openfoodfacts": 5.2, "openproductsfacts": 5.2, "openbeautyfacts": 5.2, "cosmos": 1.0}      # seconds between calls: comfortably under the documented burst / per-minute limits
+
+
+def bench(provider, items):
+    out = {"provider": provider.name, "configured": provider.configured, "sample_size": 0, "found": 0, "not_found": 0, "with_description": 0, "with_brand": 0, "with_ncm": 0, "with_cest": 0, "errors": 0, "error_kinds": {}, "latencies_ms": []}
+    if not provider.configured:
+        out["status"] = "NOT_CONFIGURED"
+        return out
+    for item in items:
+        started = time.monotonic()
+        try:
+            product = provider.lookup_gtin(item["gtin"])
+        except ProviderError as error:
+            out["errors"] += 1
+            out["error_kinds"][error.kind] = out["error_kinds"].get(error.kind, 0) + 1
+            if error.kind in ("RATE_LIMITED", "AUTH_ERROR"):
+                break                                                            # never insist against a limit
+            time.sleep(PACE[provider.name])
+            continue
+        out["sample_size"] += 1
+        out["latencies_ms"].append(int((time.monotonic() - started) * 1000))
+        if product.found:
+            out["found"] += 1
+            out["with_description"] += bool(product.title)
+            out["with_brand"] += bool(product.brand)
+            out["with_ncm"] += bool(product.ncm)
+            out["with_cest"] += bool(product.cest)
+        else:
+            out["not_found"] += 1
+        time.sleep(PACE[provider.name])
+    n = out["sample_size"] + out["errors"]
+    lat = sorted(out.pop("latencies_ms"))
+    out.update({"error_rate": round(out["errors"] / n, 3) if n else None, "latency_p50_ms": statistics.median(lat) if lat else None, "latency_p95_ms": lat[int(0.95 * (len(lat) - 1))] if lat else None,
+                "found_rate": round(out["found"] / out["sample_size"], 3) if out["sample_size"] else None, "status": "MEASURED"})
+    return out
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--sample", required=True)
+    ap.add_argument("--out", required=True)
+    ap.add_argument("--limit", type=int, default=30)
+    ap.add_argument("--only", default="", help="comma list of provider names; results are MERGED into an existing --out")
+    args = ap.parse_args()
+    items = json.loads(Path(args.sample).read_text(encoding="utf-8"))[: args.limit]
+    client = httpx.Client(timeout=8.0, headers={"User-Agent": USER_AGENT})
+    providers = [Cosmos(client, os.environ.get("COSMOS_TOKEN")), UpcItemDb(client), OpenFoodFacts(client), OpenProductsFacts(client), OpenBeautyFacts(client)]
+    only = {n for n in args.only.split(",") if n}
+    out = Path(args.out)
+    result = json.loads(out.read_text(encoding="utf-8")) if only and out.exists() else {"measured_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()), "sample_source": "lab ERP databases (read-only), valid GTINs only", "limit_per_provider": args.limit, "providers": []}
+    measured = [bench(p, items) for p in providers if not only or p.name in only]
+    result["providers"] = [p for p in result["providers"] if p["provider"] not in {m["provider"] for m in measured}] + measured
+    Path(args.out).write_text(json.dumps(result, ensure_ascii=False, indent=1), encoding="utf-8")
+    print(json.dumps(result, ensure_ascii=False, indent=1))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/enrichment_copy_e2e.py b/scripts/enrichment_copy_e2e.py
new file mode 100644
index 0000000..f0da06f
--- /dev/null
+++ b/scripts/enrichment_copy_e2e.py
@@ -0,0 +1,237 @@
+"""COPY-ONLY E2E of the new-product enrichment on a REAL Firebird (disposable copy restored from its own .fbk).
+
+    python scripts/enrichment_copy_e2e.py [--products 12]
+
+The Agent code under test is the real one (gti_fiscal_agent.enrichment: real fdb connections, real typed UPDATE, real threads/timing). The Hub side is the real resolver run in-process on
+SQLite with the official NCM dataset and fake providers (no internet). Proves on a real Firebird:
+  detection latency P50/P95 (product INSERTed by an "operator" -> fields written), only-empty fields written, other columns + every other table untouched, historical rows untouched,
+  second run = 0 writes, DRY_RUN = 0 writes, DISABLED / WAITING = 0 Firebird activity, human override incl. deleted value, Hub offline -> pending -> recovers, authorization revoked -> 0 writes.
+"""
+import json
+import os
+import statistics
+import sys
+import threading
+import time
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT)); sys.path.insert(0, str(ROOT / "scripts")); sys.path.insert(0, str(ROOT / "agent" / "src"))
+from dotenv import load_dotenv  # noqa: E402
+
+load_dotenv(ROOT / ".env")
+import rtc_copy_write_e2e as base  # noqa: E402
+from gti_fiscal_agent import enrichment as E  # noqa: E402
+from sqlalchemy import create_engine  # noqa: E402
+from sqlalchemy.orm import Session  # noqa: E402
+from sqlalchemy.pool import StaticPool  # noqa: E402
+
+from central.app.datasets.ncm import parse_ncm_payload  # noqa: E402
+from central.app.datasets.service import DatasetService  # noqa: E402
+from central.app.enrichment import kb  # noqa: E402
+from central.app.enrichment.providers import Provider, ProviderProduct  # noqa: E402
+from central.app.enrichment.quota import ProviderGateway  # noqa: E402
+from central.app.enrichment.resolver import resolve  # noqa: E402
+from central.app.models import Base, CestEntry, CestNcmEntry, Dataset, TipiEntry  # noqa: E402
+from datetime import datetime, timezone  # noqa: E402
+import httpx  # noqa: E402
+
+
+def gtin(body12):
+    total = sum(int(d) * (3 if i % 2 == 0 else 1) for i, d in enumerate(reversed(body12)))
+    return body12 + str((10 - total % 10) % 10)
+
+
+class Fake(Provider):
+    def __init__(self, name, ncm):
+        super().__init__(client=httpx.Client())
+        self.name, self.ncm, self.provides_ncm = name, ncm, True
+
+    def lookup_gtin(self, g):
+        return ProviderProduct(self.name, g, True, title="PRODUTO DE TESTE ESPECIFICO 350ML", ncm=self.ncm)
+
+
+class InProcessHub:
+    """The real Hub resolver, in-process. `online=False` simulates the Hub being unreachable."""
+
+    def __init__(self):
+        engine = create_engine("sqlite://", connect_args={"check_same_thread": False}, poolclass=StaticPool)
+        Base.metadata.create_all(engine)
+        self.session = Session(engine)
+        DatasetService(self.session).import_payload(parse_ncm_payload((ROOT / "scratch_enrich" / "ncm_siscomex.json").read_bytes()), "official", datetime.now(timezone.utc))
+        tipi = Dataset(kind="TIPI_IPI", provider="RFB", sha256="b" * 64, version_label="v", act="a", source_url="u", status="ACTIVE", record_count=1)
+        cest = Dataset(kind="CEST", provider="CONFAZ", sha256="c" * 64, version_label="v", act="a", source_url="u", status="ACTIVE", record_count=1)
+        self.session.add_all([tipi, cest]); self.session.flush()
+        self.session.add(TipiEntry(dataset_id=tipi.id, ncm="22021000", ex=None, description="x", ipi_rate=0.0, ipi_rate_text="0", source_row=1))
+        e = CestEntry(dataset_id=cest.id, code="03.007.00", description="PRODUTO DE TESTE ESPECIFICO 350ML", annex="I")
+        self.session.add(e); self.session.flush(); self.session.add(CestNcmEntry(cest_entry_id=e.id, ncm_code="22021000")); self.session.commit()
+        self.providers = [Fake("cosmos", "22021000"), Fake("second", "22021000")]
+        self.features = {"features_configured": True, "corrector_enabled": True, "enrichment_mode": "ENABLED"}
+        self.online, self.reports, self.overridden, self.calls = True, [], {}, 0
+        self.lock = threading.Lock()
+
+    def authorizations(self):
+        if not self.online:
+            raise ConnectionError("hub offline")
+        return {"features": dict(self.features)}
+
+    def resolve_product(self, payload):
+        if not self.online:
+            raise ConnectionError("hub offline")
+        with self.lock:
+            self.calls += 1
+            r = resolve(self.session, ProviderGateway(self.session, self.providers), gtin_raw=payload["gtin"], description=payload["description"], current=payload["current"])
+            for f in self.overridden.get(payload["product_code"], ()):
+                r["fields"][f] = {"status": "HUMAN_OVERRIDDEN", "value": None}
+            r["retry_at"] = None
+            return r
+
+    def report_enrichment(self, payload):
+        with self.lock:
+            self.reports.append(payload)
+            for it in payload["fields"]:
+                if it["outcome"] == "HUMAN_OVERRIDDEN":
+                    self.overridden.setdefault(payload["product_code"], set()).add(it["field"])
+        return {"ok": True}
+
+
+_orig_connect = base.connect
+
+
+def _connect_retry(rw):
+    """Same brief retry the Agent applies: local XNET connections can report 'connection lost' when several threads connect at the same instant."""
+    for attempt in range(6):
+        try:
+            return _orig_connect(rw)
+        except Exception:  # noqa: BLE001
+            if attempt == 5:
+                raise
+            time.sleep(0.2 * (attempt + 1))
+
+
+base.connect = _connect_retry
+
+
+def insert_product(code, name, ean):
+    conn = base.connect(True); cur = conn.cursor()
+    cur.execute("INSERT INTO C000025 (CODIGO, PRODUTO, CODBARRA) VALUES (?, ?, ?)", (code, name, ean))
+    conn.commit(); conn.close()
+
+
+def row(code):
+    conn = base.connect(False); cur = conn.cursor()
+    cur.execute("SELECT CLASSIFICACAO_FISCAL, CEST, IPI FROM C000025 WHERE CODIGO=?", (code,))
+    r = cur.fetchone(); conn.close()
+    return tuple(None if v is None else (str(v).strip() if not isinstance(v, float) else v) for v in r) if r else None
+
+
+def side_hashes():
+    """C000058 (ERP change journal) and C000100 (stock) are maintained by the SGV's OWN triggers on C000025 INSERT/UPDATE. Only rows of the test products (code 'Z...') may differ."""
+    import hashlib
+    conn = base.connect(False); cur = conn.cursor(); out = {}
+    for table, col in (("C000058", "CODREGISTRO"), ("C000100", "CODPRODUTO")):
+        cur.execute(f"SELECT * FROM {table} WHERE {col} NOT STARTING WITH 'Z'")
+        h = hashlib.sha256()
+        n = 0
+        for r in cur.fetchall():
+            h.update(repr(r).encode("utf-8", "replace")); n += 1
+        out[table] = f"{n}:{h.hexdigest()[:16]}"
+    conn.close()
+    return out
+
+
+def main():
+    n = int(sys.argv[sys.argv.index("--products") + 1]) if "--products" in sys.argv else 12
+    checks = {}
+    base.restore_copy()
+    before_tables, _ = base.baseline()
+    before_side = side_hashes()
+    hub = InProcessHub()
+    for suffix in ("", "-wal", "-shm"):
+        try:
+            (ROOT / "scratch_enrich" / f"enrich_e2e.db{suffix}").unlink()
+        except FileNotFoundError:
+            pass
+    store = E.EnrichmentStore(ROOT / "scratch_enrich" / "enrich_e2e.db")
+    worker = E.EnrichmentWorker(hub=hub, store=store, read_conn=lambda: base.connect(False), write_conn=lambda: base.connect(True),
+                                create_reprocess=lambda codes: None)
+
+    class Stop:
+        stop = False
+    stopper = Stop()
+    thread = threading.Thread(target=worker.run_forever, args=(stopper,), daemon=True)
+
+    # --- WAITING / DISABLED: zero Firebird activity, baseline NOT taken
+    hub.features = {"features_configured": False}
+    thread.start(); time.sleep(4)
+    checks["WAITING_NO_BASELINE_NO_CALLS"] = store.meta("baseline_done") is None and hub.calls == 0
+    hub.features = {"features_configured": True, "corrector_enabled": True, "enrichment_mode": "DISABLED"}; worker._features_at = -999; time.sleep(4)
+    checks["DISABLED_NO_BASELINE_NO_CALLS"] = store.meta("baseline_done") is None and hub.calls == 0
+
+    # --- DRY_RUN: baseline, detect, resolve, report, zero writes
+    hub.features = {"features_configured": True, "corrector_enabled": True, "enrichment_mode": "DRY_RUN"}; worker._features_at = -999; time.sleep(5)
+    checks["BASELINE_TAKEN_ON_ACTIVATION"] = store.meta("baseline_done") == "1"
+    hist = len(store.known())
+    insert_product("ZD001", "PRODUTO DE TESTE ESPECIFICO 350ML", gtin("789200000001")); time.sleep(8)
+    checks["DRY_RUN_ZERO_WRITES"] = row("ZD001") == (None, None, None) and hub.reports and hub.reports[-1]["fields"][0]["outcome"] == "DRY_RUN"
+
+    # --- ENABLED: N products inserted like an operator would, latency measured insert -> written
+    hub.features = {"features_configured": True, "corrector_enabled": True, "enrichment_mode": "ENABLED"}; worker._features_at = -999; worker._auth_cache = (None, 0.0)
+    latencies = []
+    for i in range(n):
+        code, ean = f"ZE{i:03d}", gtin(f"78930000{i:04d}")
+        t0 = time.monotonic()
+        insert_product(code, "PRODUTO DE TESTE ESPECIFICO 350ML", ean)
+        while time.monotonic() - t0 < 20 and row(code) == (None, None, None):
+            time.sleep(0.1)
+        latencies.append(round(time.monotonic() - t0, 2))
+        time.sleep(0.5)
+    ok_rows = [row(f"ZE{i:03d}") for i in range(n)]
+    checks["FIELDS_WRITTEN"] = all(r and r[0] == "22021000" and r[1] == "0300700" and r[2] == 0.0 for r in ok_rows)
+    time.sleep(6)
+    writes_before = sum(1 for rep in hub.reports for f in rep["fields"] if f["outcome"] == "APPLIED")
+    calls_before = hub.calls
+    time.sleep(8)                                                                # a "second run": nothing pending, nothing re-written
+    checks["SECOND_RUN_ZERO_WRITES"] = writes_before == sum(1 for rep in hub.reports for f in rep["fields"] if f["outcome"] == "APPLIED") and hub.calls - calls_before <= 0
+    checks["HISTORICAL_KNOWN_COUNT"] = hist > 100
+
+    # --- human override incl. deleted value
+    conn = base.connect(True); cur = conn.cursor()
+    cur.execute("UPDATE C000025 SET CEST = NULL WHERE CODIGO = 'ZE000'"); cur.execute("UPDATE C000025 SET CLASSIFICACAO_FISCAL = '19059090' WHERE CODIGO = 'ZE001'"); conn.commit(); conn.close()
+    time.sleep(E.OVERRIDE_CHECK_SECONDS + 6)
+    checks["OVERRIDE_DELETED_VALUE"] = store.is_overridden("ZE000", "CEST") and store.is_overridden("ZE001", "CLASSIFICACAO_FISCAL")
+    checks["OVERRIDE_REPORTED_TO_HUB"] = "CEST" in hub.overridden.get("ZE000", set())
+    # ... and a later description edit does NOT bring the value back
+    conn = base.connect(True); cur = conn.cursor(); cur.execute("UPDATE C000025 SET PRODUTO = 'PRODUTO DE TESTE ESPECIFICO 350ML ZERO' WHERE CODIGO = 'ZE000'"); conn.commit(); conn.close()
+    time.sleep(10)
+    checks["OVERRIDE_NEVER_REWRITTEN"] = row("ZE000")[1] in (None, "") and row("ZE001")[0] == "19059090"
+
+    # --- Hub offline -> pending -> recovers ; authorization revoked -> 0 writes
+    hub.online = False
+    insert_product("ZO001", "PRODUTO DE TESTE ESPECIFICO 350ML", gtin("789400000001")); time.sleep(8)
+    checks["HUB_OFFLINE_PENDING_NO_WRITE"] = row("ZO001") == (None, None, None) and store.get("ZO001")["state"] in ("RETRY", "PENDING")
+    hub.online = True
+    with store.db:
+        store.db.execute("UPDATE enrich_products SET next_attempt_at=0 WHERE code='ZO001'")
+    time.sleep(8)
+    checks["HUB_BACK_RECOVERS"] = row("ZO001") == ("22021000", "0300700", 0.0)
+    stopper.stop = True
+    thread.join(timeout=10)
+
+    # --- invariants on the real Firebird copy: only the 3 allowed columns of the new rows changed
+    after_tables, _ = base.baseline()
+    changed = base.diff_tables(before_tables, after_tables)
+    checks["OTHER_TABLES_ONLY_ERP_TRIGGER_SIDE_EFFECTS_OF_TEST_ROWS"] = set(changed) <= {"C000025", "C000058", "C000100"} and side_hashes() == before_side
+    conn = base.connect(False); cur = conn.cursor()
+    cur.execute("SELECT COUNT(*) FROM C000025 WHERE CODIGO STARTING WITH 'Z' AND COD_ORIGEM_MERCADORIA IS NOT NULL AND TRIM(COD_ORIGEM_MERCADORIA) <> ''"); origin = cur.fetchone()[0]; conn.close()
+    checks["ORIGIN_NEVER_WRITTEN"] = origin == 0
+    result = {"copy_only": True, "products_enabled": n, "detect_to_written_seconds": {"p50": statistics.median(latencies), "p95": sorted(latencies)[int(0.95 * (len(latencies) - 1))], "max": max(latencies), "all": latencies},
+              "hub_resolve_calls": hub.calls, "tables_changed": changed, "checks": checks, "ALL_PASS": all(checks.values()), "poll_seconds": worker.poll_seconds()}
+    Path(ROOT / "docs" / "runs").mkdir(parents=True, exist_ok=True)
+    (ROOT / "docs" / "runs" / "enrichment_copy_e2e.json").write_text(json.dumps(result, indent=1), encoding="utf-8")
+    print(json.dumps(result, indent=1))
+    sys.exit(0 if result["ALL_PASS"] else 1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/enrichment_lab_kb.py b/scripts/enrichment_lab_kb.py
new file mode 100644
index 0000000..3809b53
--- /dev/null
+++ b/scripts/enrichment_lab_kb.py
@@ -0,0 +1,103 @@
+"""READ-ONLY study of the five lab ERP databases for the product enrichment: GTIN quality, NCM coverage, cross-database agreement, and KB seeding.
+
+    python scripts/enrichment_lab_kb.py [--seed-url sqlite:///kb.db] [--sample 40 --sample-out scratch/benchmark_gtins.json]
+
+Firebird is opened READ-ONLY (no write path exists here). The KB seed goes ONLY to the database given by --seed-url (default: an in-memory SQLite, i.e. statistics only) and stores
+CLIENT_OBSERVATION evidence, never VERIFIED_KB by itself. Company-specific fiscal fields are not read at all.
+"""
+import argparse
+import json
+import os
+import sys
+from collections import Counter, defaultdict
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+sys.path.insert(0, str(ROOT / "scripts"))
+from dotenv import load_dotenv  # noqa: E402
+
+load_dotenv(ROOT / ".env")
+import local_completion as lc  # noqa: E402
+from sqlalchemy import create_engine, select  # noqa: E402
+from sqlalchemy.orm import Session  # noqa: E402
+
+from central.app.enrichment import kb  # noqa: E402
+from central.app.enrichment.gtin import normalize_gtin  # noqa: E402
+from central.app.enrichment.observations import observe  # noqa: E402
+from central.app.enrichment.resolver import ncm_is_official  # noqa: E402
+from central.app.models import Base, Dataset, KbKnowledge, NcmEntry  # noqa: E402
+
+
+def read_lab(name, path):
+    conn = lc.connect(path)
+    try:
+        cur = conn.cursor()
+        cur.execute("SELECT CODIGO, PRODUTO, GTIN, CLASSIFICACAO_FISCAL, CODBARRA FROM C000025")
+        # the SGV operator types the EAN in CODBARRA; GTIN (fiscal) is mostly empty. The first VALID one wins (same rule as the Agent).
+        return [{"CODIGO": r[0], "PRODUTO": (r[1] or "").strip(), "GTIN": next((x for x in (r[2], r[4]) if normalize_gtin(x)["status"] == "VALID_GTIN"), r[2] or r[4]), "NCM": r[3]} for r in cur.fetchall()]
+    finally:
+        conn.close()
+
+
+def load_ncm_official(session, ncm_json: Path | None):
+    """Official NCM into the throw-away/seed session when it has none (parsed from the saved official JSON)."""
+    if session.scalar(select(Dataset).where(Dataset.kind == "NCM", Dataset.status == "ACTIVE")):
+        return
+    from central.app.datasets.ncm import parse_ncm_payload
+    from central.app.datasets.service import DatasetService
+    from datetime import datetime, timezone
+    payload = parse_ncm_payload(ncm_json.read_bytes())
+    DatasetService(session).import_payload(payload, "official-json", datetime.now(timezone.utc))
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--seed-url", default="sqlite://")
+    ap.add_argument("--ncm-json", default=None, help="official Siscomex JSON (used only if the seed DB has no NCM dataset)")
+    ap.add_argument("--sample", type=int, default=0)
+    ap.add_argument("--sample-out", default=None)
+    args = ap.parse_args()
+    engine = create_engine(args.seed_url)
+    Base.metadata.create_all(engine)
+    dbs = dict(lc.allowlist())
+    dbs["TESTESOFICIAIS"] = os.environ.get("FIREBIRD_DATABASE_PATH", str(ROOT / "TESTESOFICIAIS.FDB"))
+    report, all_rows = {}, {}
+    with Session(engine) as session:
+        load_ncm_official(session, Path(args.ncm_json) if args.ncm_json else None)
+        official = lambda v: ncm_is_official(session, v)  # noqa: E731
+        for name, path in dbs.items():
+            if not Path(path).exists():
+                report[name] = {"error": "missing"}
+                continue
+            rows = read_lab(name, path)
+            all_rows[name] = rows
+            status = Counter(normalize_gtin(r["GTIN"])["status"] for r in rows)
+            stats = observe(session, rows, name, official)
+            ncm_ok = sum(1 for r in rows if len("".join(c for c in str(r["NCM"] or "") if c.isdigit())) == 8 and official("".join(c for c in str(r["NCM"]) if c.isdigit())))
+            report[name] = {"products": len(rows), "gtin_status": dict(status), "valid_gtin_rate": round(100 * status["VALID_GTIN"] / max(1, len(rows)), 1), "official_ncm_rate": round(100 * ncm_ok / max(1, len(rows)), 1), **stats}
+        session.commit()
+        by_gtin = defaultdict(set)
+        for name, rows in all_rows.items():
+            for r in rows:
+                g = normalize_gtin(r["GTIN"])
+                if g["status"] == "VALID_GTIN":
+                    by_gtin[g["key"]].add(name)
+        shared = {k: v for k, v in by_gtin.items() if len(v) >= 2}
+        conflicts = 0
+        for kp in session.scalars(select(kb.KbProduct)):
+            ncms = {r.value for r in session.scalars(select(KbKnowledge).where(KbKnowledge.kb_product_id == kp.id, KbKnowledge.knowledge_key == "ncm"))}
+            conflicts += len(ncms) > 1
+        summary = {"databases": report, "distinct_valid_gtin": len(by_gtin), "gtin_in_2plus_databases": len(shared), "kb_products": session.scalar(select(kb.KbProduct.id).limit(1)) is not None and session.query(kb.KbProduct).count(),
+                   "kb_ncm_conflicts_between_clients": conflicts}
+        if args.sample and args.sample_out:
+            picks = sorted(by_gtin)[:: max(1, len(by_gtin) // args.sample)][: args.sample]
+            names = {normalize_gtin(r["GTIN"])["key"]: r["PRODUTO"] for rows in all_rows.values() for r in rows if normalize_gtin(r["GTIN"])["status"] == "VALID_GTIN"}
+            Path(args.sample_out).parent.mkdir(parents=True, exist_ok=True)
+            Path(args.sample_out).write_text(json.dumps([{"gtin": k[1:] if k.startswith("0") else k, "erp_description": names[k]} for k in picks], ensure_ascii=False, indent=1), encoding="utf-8")
+            summary["sample_written"] = len(picks)
+    print(json.dumps(summary, ensure_ascii=False, indent=1))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/enrichment_live_smoke.py b/scripts/enrichment_live_smoke.py
new file mode 100644
index 0000000..b3fb858
--- /dev/null
+++ b/scripts/enrichment_live_smoke.py
@@ -0,0 +1,44 @@
+"""Conservative LIVE smoke of the free providers (separate from the offline test suite): 1 lookup per provider on a famous, stable GTIN.
+
+    python scripts/enrichment_live_smoke.py
+
+Goes through the real ProviderGateway (cache + quota + circuit breaker) on an in-memory database; tokens are never printed. Exit 0 when every CONFIGURED provider answered
+without an infrastructure error (NOT_FOUND is a valid answer; a rate limit is reported, not treated as a failure of our code).
+"""
+import json
+import os
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+from sqlalchemy import create_engine  # noqa: E402
+from sqlalchemy.orm import Session  # noqa: E402
+from sqlalchemy.pool import StaticPool  # noqa: E402
+
+from central.app.enrichment.providers import build_providers  # noqa: E402
+from central.app.enrichment.quota import ProviderGateway  # noqa: E402
+from central.app.models import Base  # noqa: E402
+
+GTIN = "7894900011517"        # Coca-Cola 350 ml (BR): stable, widely catalogued
+
+
+def main():
+    engine = create_engine("sqlite://", connect_args={"check_same_thread": False}, poolclass=StaticPool)
+    Base.metadata.create_all(engine)
+    settings = type("S", (), {"cosmos_token": os.environ.get("COSMOS_TOKEN", ""), "allow_paid_providers": False})()
+    out, ok = [], True
+    with Session(engine) as session:
+        gateway = ProviderGateway(session, build_providers(settings))
+        for provider in gateway.providers:
+            result = gateway.lookup(provider, GTIN)
+            row = {"provider": provider.name, "configured": provider.configured, "state": result.state, "reason": result.reason, "title": result.product.title if result.product else None}
+            out.append(row)
+            if provider.configured and result.state == "ERROR" and result.reason not in ("RATE_LIMITED",):
+                ok = False
+    print(json.dumps({"gtin": GTIN, "results": out, "OK": ok}, ensure_ascii=False, indent=1))
+    sys.exit(0 if ok else 1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/enrichment_local_hub_e2e.py b/scripts/enrichment_local_hub_e2e.py
new file mode 100644
index 0000000..8880990
--- /dev/null
+++ b/scripts/enrichment_local_hub_e2e.py
@@ -0,0 +1,137 @@
+"""E2E of the enrichment across REAL HTTP: the real Agent worker (real Firebird COPY) <-> the real Hub container (dev PostgreSQL, real datasets, real providers).
+
+    docker compose up -d --build fiscal-hub   # Hub at http://localhost:8000
+    python scripts/enrichment_local_hub_e2e.py
+
+Registers a throw-away dev Agent (AGENT_DEV_TOKEN), flips its per-installation features straight in the dev database (the panel route is covered by tests), seeds the KB of the dev Hub with a VERIFIED NCM for
+one test GTIN (two independent provider families), then walks: DISABLED -> DRY_RUN -> ENABLED -> DISABLED, an unknown GTIN (real providers, internet) and an invalid GTIN (no provider call).
+Writes go ONLY to scratch_rtc_e2e/TESTESOFICIAIS_RTC_COPY.FDB (guarded by rtc_copy_write_e2e.guard). The dev database receives test rows only.
+"""
+import json
+import os
+import subprocess
+import sys
+import threading
+import time
+import uuid
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT)); sys.path.insert(0, str(ROOT / "scripts")); sys.path.insert(0, str(ROOT / "agent" / "src"))
+from dotenv import load_dotenv  # noqa: E402
+
+load_dotenv(ROOT / ".env")
+import enrichment_copy_e2e as ce  # noqa: E402  (also patches base.connect with the brief retry)
+from gti_fiscal_agent import enrichment as E  # noqa: E402
+from gti_fiscal_agent.client import HubClient  # noqa: E402
+
+base = ce.base
+HUB = "http://127.0.0.1:8000"          # NOT "localhost": on Windows it resolves ::1 first and adds ~500 ms per request
+TOKEN = os.environ["AGENT_DEV_TOKEN"]
+CONTAINER = "projeto-postgres-1"
+
+
+def sql(query):
+    done = subprocess.run(["docker", "exec", CONTAINER, "psql", "-U", "gti", "-d", "gti_fiscal", "-At", "-c", query], capture_output=True, text=True, timeout=60)
+    if done.returncode != 0:
+        raise RuntimeError(done.stderr[-300:])
+    return done.stdout.strip()
+
+
+def seed_kb(gtin14, description, ncm):
+    code = f'''
+from central.app.db import SessionLocal
+from central.app.enrichment import kb
+with SessionLocal() as s:
+    p = kb.upsert_product(s, "{gtin14}", "{description}")
+    kb.add_evidence(s, p, "ncm", "{ncm}", "provider:cosmos", family="PROVIDER:cosmos")
+    kb.add_evidence(s, p, "ncm", "{ncm}", "provider:other", family="PROVIDER:other")
+    s.commit()
+print("seeded")
+'''
+    done = subprocess.run(["docker", "exec", "-i", "projeto-fiscal-hub-1", "python", "-"], input=code, capture_output=True, text=True, timeout=120, cwd=str(ROOT))
+    assert "seeded" in done.stdout, done.stderr[-400:]
+
+
+def set_features(installation_id, mode, corrector=True):
+    sql(f"UPDATE installations SET enrichment_mode='{mode}', corrector_enabled={'true' if corrector else 'false'}, features_configured=true WHERE installation_id='{installation_id}'")
+
+
+def main():
+    checks = {}
+    base.restore_copy()
+    client = HubClient(HUB, TOKEN)
+    try:
+        installation_id = client.identity().get("installation_id") or str(uuid.uuid4())        # the dev Agent may already own an installation
+    except Exception:  # noqa: BLE001
+        installation_id = str(uuid.uuid4())
+    client.heartbeat_metadata({"installation_id": installation_id, "status": "IDLE", "agent_version": "1.3.0", "display_name": "E2E-ENRICH", "capabilities": {"PRODUCT_ENRICHMENT_WRITE_V1": True}})
+    set_features(installation_id, "DISABLED")                                   # start every run from the default
+    checks["REGISTERED_DEFAULT_DISABLED"] = client.authorizations()["features"]["enrichment_mode"] == "DISABLED"
+    gtin_ok = ce.gtin("789500000001")
+    seed_kb("0" + gtin_ok, "REFRIGERANTE COLA LATA 350ML", "22021000")
+    for suffix in ("", "-wal", "-shm"):
+        try:
+            (ROOT / "scratch_enrich" / f"enrich_hub_e2e.db{suffix}").unlink()
+        except FileNotFoundError:
+            pass
+    store = E.EnrichmentStore(ROOT / "scratch_enrich" / "enrich_hub_e2e.db")
+    worker = E.EnrichmentWorker(hub=client, store=store, read_conn=lambda: base.connect(False), write_conn=lambda: base.connect(True), create_reprocess=lambda codes: client.create_job("REPROCESS_PRODUCTS", {"codigos": list(codes), "reason": "ENRICHMENT"}))
+
+    class Stop:
+        stop = False
+    stopper = Stop()
+    threading.Thread(target=worker.run_forever, args=(stopper,), daemon=True).start()
+
+    def refresh():
+        worker._features_at = -999
+        worker._auth_cache = (None, 0.0)
+
+    time.sleep(4)
+    checks["DISABLED_NO_ACTIVITY"] = store.meta("baseline_done") is None
+    set_features(installation_id, "DRY_RUN"); refresh(); time.sleep(6)
+    checks["BASELINE_ON_ACTIVATION"] = store.meta("baseline_done") == "1"
+    ce.insert_product("HD001", "REFRIGERANTE COLA LATA 350ML", gtin_ok); time.sleep(8)
+    checks["DRY_RUN_ZERO_WRITES"] = ce.row("HD001") == (None, None, None)
+    audits = int(sql(f"SELECT count(*) FROM product_field_audits a JOIN installations i ON i.id=a.installation_id WHERE i.installation_id='{installation_id}' AND a.outcome='DRY_RUN' AND a.product_code='HD001'"))
+    checks["DRY_RUN_AUDITED_IN_HUB"] = audits >= 1
+    set_features(installation_id, "ENABLED"); refresh()
+    ce.insert_product("HE001", "REFRIGERANTE COLA LATA 350ML", gtin_ok)
+    t0 = time.monotonic()
+    while time.monotonic() - t0 < 25 and ce.row("HE001")[0] is None:
+        time.sleep(0.2)
+    latency = round(time.monotonic() - t0, 2)
+    r = ce.row("HE001")
+    checks["ENABLED_NCM_WRITTEN"] = r[0] == "22021000"
+    time.sleep(3)                                                              # the Agent reports the audit right after the write
+    applied = sql(f"SELECT field||':'||coalesce(new_value,'') FROM product_field_audits a JOIN installations i ON i.id=a.installation_id WHERE i.installation_id='{installation_id}' AND a.outcome='APPLIED' AND a.product_code='HE001' ORDER BY a.id")
+    checks["WRITTEN_FIELDS_MATCH_HUB_AUDIT"] = ("CLASSIFICACAO_FISCAL:22021000" in applied) and all(v in applied for v in ([f"CEST:{r[1]}"] if r[1] else []) + ([f"IPI:{r[2]}"] if r[2] not in (None,) else []))
+    time.sleep(6)
+    hd = ce.row("HD001")
+    checks["EARLIER_DRY_RUN_PRODUCT_STILL_UNTOUCHED"] = hd == (None, None, None)
+    # unknown GTIN goes to the REAL providers (internet); nothing may be written, a human blocker must appear
+    unknown = ce.gtin("789600000009")
+    ce.insert_product("HE002", "PRODUTO INEXISTENTE ESPECIAL XYZ 123", unknown); time.sleep(14)
+    checks["UNKNOWN_GTIN_NO_WRITE"] = ce.row("HE002") == (None, None, None)
+    blockers = sql("SELECT reason_code FROM blocked_products WHERE field='ENRICHMENT' AND product_code='HE002'")
+    checks["UNKNOWN_GTIN_BLOCKER_PT"] = blockers in ("NOT_FOUND", "SINGLE_WEAK_SOURCE", "PENDING_PROVIDER_QUOTA", "PENDING_PROVIDERS_OFFLINE") or blockers.startswith("NOT_FOUND")
+    bad = unknown[:-1] + str((int(unknown[-1]) + 1) % 10)
+    before_req = sql("SELECT coalesce(sum(requests),0) FROM provider_usage_daily")
+    ce.insert_product("HE003", "PRODUTO GTIN INVALIDO", bad); time.sleep(8)
+    checks["INVALID_GTIN_NO_PROVIDER_CALL"] = sql("SELECT coalesce(sum(requests),0) FROM provider_usage_daily") == before_req and ce.row("HE003") == (None, None, None)
+    # feature switched off centrally: the next new product is not touched
+    set_features(installation_id, "DISABLED"); refresh(); time.sleep(4)
+    ce.insert_product("HE004", "REFRIGERANTE COLA LATA 350ML", gtin_ok); time.sleep(8)
+    checks["DISABLED_AGAIN_ZERO_WRITES"] = ce.row("HE004") == (None, None, None)
+    # corrector off + enrichment on: independent switches
+    set_features(installation_id, "ENABLED", corrector=False); refresh(); time.sleep(3)
+    checks["CORRECTOR_OFF_NO_JOBS"] = client.next_job() is None
+    stopper.stop = True
+    result = {"copy_only": True, "hub": HUB, "detect_to_written_seconds_enabled": latency, "checks": checks, "ALL_PASS": all(checks.values())}
+    (ROOT / "docs" / "runs" / "enrichment_local_hub_e2e.json").write_text(json.dumps(result, indent=1), encoding="utf-8")
+    print(json.dumps(result, indent=1))
+    sys.exit(0 if result["ALL_PASS"] else 1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/test_release_candidate.py b/tests/test_release_candidate.py
index 576b083..98c293b 100644
--- a/tests/test_release_candidate.py
+++ b/tests/test_release_candidate.py
@@ -14,11 +14,11 @@ sys.path.insert(0, str(ROOT))
 def test_versions_are_coherent_across_hub_agent_and_packages():
     from central.app.version import GTI_FISCAL_VERSION
     v = (ROOT / "release" / "VERSION").read_text(encoding="utf-8").strip()
-    assert v == GTI_FISCAL_VERSION == "1.2.5"
+    assert v == GTI_FISCAL_VERSION == "1.3.0"
     pep = v.replace("-rc", "rc")
     for pp in ("central/pyproject.toml", "agent/pyproject.toml"):
         assert f'version = "{pep}"' in (ROOT / pp).read_text(encoding="utf-8")
-    assert '"AGENT_VERSION", "1.2.5"' in (ROOT / "agent" / "scripts" / "run_agent_service.py").read_text(encoding="utf-8")
+    assert '"AGENT_VERSION", "1.3.0"' in (ROOT / "agent" / "scripts" / "run_agent_service.py").read_text(encoding="utf-8")
     manifest = ROOT / "release" / "RELEASE_MANIFEST.json"
     if manifest.exists():
         m = json.loads(manifest.read_text(encoding="utf-8"))