Mejoras de sesion: estabilizacion, IA, Nexus Yonke y marketplace B2B

- F0 blindaje: Redis/Celery, pg_hba sin trust LAN, secretos a env 600, backups, webhook auth, jobs muertos
- F1 caja: bugs POS, reembolsos al corte, apartados, CFDI, topes de descuento en cotizacion/apartado
- F2/F7 IA de catalogo: autocategorizacion, dedup de SKUs con embeddings, OCR de facturas, descripciones, claves SAT, limpieza de fitment
- F3 IA en el flujo: vision Yonke, preguntas MeLi, venta cruzada, anomalias, sinonimos, busqueda semantica (rerank/embeddings)
- F4/F8 escala y cartera: /metrics+Prometheus+Grafana, pool, score de credito, churn, metricas multi-tenant
- Nexus Yonke (M1-M6): modelo de datos 1-de-1, alta por donador, vitrina publica, MeLi usado, consignacion, garantia por dias, reportes, etiquetas
- Marketplace B2B entre yonkes (catalogo + pedidos cross-tenant)
- Seguridad: auth en /api/admin/* (35 endpoints), alertas por tenant
- Migraciones v4.27/v4.28; timers expire-demos y rollup-metrics

pos/tenant_db.py

@@ -26,8 +26,11 @@ def _get_tenant_pool(db_name):
     global _tenant_pools
     if db_name not in _tenant_pools:
         dsn = TENANT_DB_URL_TEMPLATE.format(db_name=db_name)
+        # Pool is per-process: with N gunicorn workers there are N pools per
+        # tenant. Keep minconn low to bound idle connections (the real fix at
+        # scale is PgBouncer; see docs/plans/2026-10-11-auditoria-completa-mejoras-ia.md).
         _tenant_pools[db_name] = pool.ThreadedConnectionPool(
-            minconn=2, maxconn=20, dsn=dsn
+            minconn=1, maxconn=20, dsn=dsn
         )
     return _tenant_pools[db_name]
 
@@ -94,6 +97,30 @@ class _PooledConnection:
         self.close()
 
 
+def _get_live_conn(db_pool, attempts=3):
+    """Get a connection from the pool, verifying it is alive.
+
+    Idle pooled connections get killed by PostgreSQL/SSL timeouts; a dead
+    socket surfaces later as 'SSL connection has been closed unexpectedly'
+    mid-query. Probe with SELECT 1 and discard dead connections.
+    """
+    last_exc = None
+    for _ in range(attempts):
+        conn = db_pool.getconn()
+        try:
+            cur = conn.cursor()
+            cur.execute("SELECT 1")
+            cur.close()
+            return conn
+        except psycopg2.OperationalError as e:
+            last_exc = e
+            try:
+                db_pool.putconn(conn, close=True)  # discard dead connection
+            except Exception:
+                pass
+    raise last_exc
+
+
 # ─── Public API ────────────────────────────────
 
 def get_master_conn():
@@ -111,10 +138,10 @@ def get_tenant_conn(tenant_id):
         raise ValueError(f"Tenant {tenant_id} not found or inactive")
 
     p = _get_tenant_pool(db_name)
-    return _PooledConnection(p.getconn(), p)
+    return _PooledConnection(_get_live_conn(p), p)
 
 
 def get_tenant_conn_by_dbname(db_name):
     """Get a pooled connection to a tenant DB directly by name."""
     p = _get_tenant_pool(db_name)
-    return _PooledConnection(p.getconn(), p)
+    return _PooledConnection(_get_live_conn(p), p)