feat(gateway): stadio rerank con catena di fallback resiliente (frigate→brain)

- gateway/rerank.py: catena da RERANK_CHAIN (JSON, per-nodo key+timeout),
  cooldown 60s sui nodi falliti, score sigmoide [0,1], degrada con grazia
  all'ordine di fusione se tutti i nodi sono giù
- routes: /v1/memories:search applica il rerank post-fusione (fetch esteso a
  RERANK_CANDIDATES), risposta con rerank{used,backend,took_ms}, flag
  per-query rerank=false; /v1/version espone lo stato rerank
- store: search() accetta limit esteso; models: SearchIn.rerank
- metrics: qmem_rerank_calls_total + durata per backend
- test: 10 nuovi (fallback, cooldown, degradazione, integrazione) — 46 pass

gateway/store.py

@@ -2,7 +2,7 @@
 from __future__ import annotations
 
 import hashlib
-from typing import Any
+from typing import Any, Optional
 
 from qdrant_client.http import models as qm
 
@@ -27,25 +27,27 @@ def search_filter(body: SearchIn) -> qm.Filter | None:
     return None
 
 
-def search(qdrant: Any, collection: str, body: SearchIn, vector: list[float], sparse: Any) -> list[Any]:
+def search(qdrant: Any, collection: str, body: SearchIn, vector: list[float], sparse: Any, limit: Optional[int] = None) -> list[Any]:
+    """Ricerca ibrida o densa. Con reranking attivo limit > top_k per dare candidati extra allo stadio di rerank."""
+    eff_limit = limit if limit is not None else body.top_k
     qfilter = search_filter(body)
     if body.hybrid and sparse is not None:
         return qdrant.query_points(
             collection_name=collection,
             prefetch=[
-                qm.Prefetch(query=vector, using="", limit=body.top_k * 4, score_threshold=body.min_score),
-                qm.Prefetch(query=sparse, using=SPARSE_VECTOR_NAME, limit=body.top_k * 4),
+                qm.Prefetch(query=vector, using="", limit=max(body.top_k * 4, eff_limit), score_threshold=body.min_score),
+                qm.Prefetch(query=sparse, using=SPARSE_VECTOR_NAME, limit=max(body.top_k * 4, eff_limit)),
             ],
             query=qm.FusionQuery(fusion=qm.Fusion.RRF),
             query_filter=qfilter,
-            limit=body.top_k,
+            limit=eff_limit,
             with_payload=True,
         ).points
     return qdrant.query_points(
         collection_name=collection,
         query=vector,
         query_filter=qfilter,
-        limit=body.top_k,
+        limit=eff_limit,
         score_threshold=body.min_score,
         with_payload=True,
     ).points