feat(embeddings): cache persistent de vectori in SQLite + warmup in fundal

Vectorii corpusului k-NN persista in tabela embedding_cache (PK model+text_hash,
blob float32 LE); la warmup se vectorizeaza doar textele lipsa din cache, deci
restartul cu corpus neschimbat nu mai plateste ~1-2 min de embed (embed=0).

- app/embedding_cache.py: serializare array('f'), load/save/purge chunk 500 cu
  BEGIN/COMMIT explicit (conexiuni autocommit), validare dimensiune la scriere
  si citire, orchestrare sync_corpus_vectors cu embed_fn injectat
- index_corpus(vectors=): vectori precalculati cu validare aliniere; mismatch
  -> fallback embed complet
- ensure_embeddings_corpus: warmup in thread la startup (block=True), calea de
  request ne-blocanta (acquire non-blocking pe lock; warmup in curs -> return
  imediat); purjare orfane + modele vechi doar dupa indexare reusita
- log warmup: cache=N embed=M in Xs
- 31 teste noi (cold/warm/incremental, model schimbat, concurenta, ranking
  exact, echivalenta float32); suita completa 1596 passed

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

TODOS.md

@@ -2,6 +2,21 @@
 
 Elemente deferate din review-uri. Negrupte de un PRD curent; de promovat cand devin prioritare.
 
+## Din /autoplan PRD 5.21 (2026-07-06)
+
+- [ ] **CLI administrare cache embeddings (`python3 -m tools.embcache stats|clear|rebuild`)** — v1 foloseste
+  `DELETE FROM embedding_cache` manual (decizie Open Q1). De construit cand operarea manuala devine
+  frecventa sau cand apare al doilea operator. Effort: S. (CEO, low.)
+- [ ] **Precompute vectori la build (artefact in imaginea Docker)** — modelul e pinned si corpusul seed e
+  comis, deci vectorii pot fi precalculati la build; ar face si PRIMUL start pe volum proaspat ieftin
+  (azi doar al doilea beneficiaza). Effort: M (pipeline build + fallback la runtime). (Outside voice CEO, medium.)
+- [ ] **Graceful reload / zero-downtime restart** — procesul vechi serveste pana cel nou e cald; ar elimina
+  COMPLET fereastra post-restart, inclusiv incarcarea modelului (~10-20s) pe care cache-ul n-o poate evita.
+  Infra (Dokploy/compose), nu app. Effort: M. (Outside voice CEO, medium.)
+- [ ] **Latenta per-query `suggest_nearest` (cosine pur Python)** — bucla Python peste 17k x 384 la fiecare
+  cautare; de MASURAT intai; daca >100ms, matrice numpy + dot product (efort mic). Designul curent e validat
+  pana la ~50k randuri de corpus (prag documentat in PRD 5.21 A12). (Eng F5, medium.)
+
 ## Din hardening 80/20 (/autoplan, 2026-07-03)
 
 - [ ] **Traefik IP-allowlist pe /v1 + /metrics in fereastra de lansare** (T4) — cu zero clienti API,