perf+calitate sugestii k-NN: matvec numpy, vot top-5 cu prag calibrat, indicatori import

- embeddings: corpus ca matrice numpy cu norme precalculate; suggest_nearest
  = un matvec (~0.6ms/query fata de ~500ms cosine pur-Python la 17k vectori)
- enrich_suggestions: vot ponderat cu similaritatea pe top-5 vecini (NUL =
  eticheta proprie); prag 0.5 -> 0.88, calibrat LOO pe corpusul SILVER
  (tools/mapare-llm/knn_calibrate.py): precizie 90.5% -> 93.1%, cod gresit
  preselectat 7.2% -> 4.7%; sub prag abtinere -> preselectie fuzzy
- UI: codul sugerat de sistem afisat explicit cu sursa si scorul, separat de
  lista fuzzy ("potrivire text"); indicator de progres reparat pe upload
  (display:inline anula .htmx-indicator) si adaugat pe pasii 2->3 si
  "Salveaza maparile"

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

tests/test_mapare_integrare_l14.py

@@ -278,7 +278,7 @@ def test_embeddings_functional_cand_flag_activ(conn, monkeypatch):
     conn.execute(
         "INSERT OR REPLACE INTO mapping_suggestions "
         "(denumire_normalizata, cod_prestatie, is_nul, source, confidence) VALUES (?, ?, ?, ?, ?)",
-        ("Schimb ulei", "UL-1", 0, "llm", 0.95),
+        ("Schimb ulei motor", "UL-1", 0, "llm", 0.95),
     )
     conn.execute(
         "INSERT OR REPLACE INTO mapping_suggestions "
@@ -292,7 +292,8 @@ def test_embeddings_functional_cand_flag_activ(conn, monkeypatch):
     ensure_embeddings_corpus(conn)
     assert emb_mod.has_corpus(), "corpusul trebuie indexat cand flagul e activ"
 
-    # "schimbat uleiul motor" -> vector [1,1,0] -> cel mai apropiat = UL-1 (Schimb ulei).
+    # "schimbat uleiul motor" -> vector [1,1,0] -> identic cu "Schimb ulei motor"
+    # (cosine 1.0, peste EMB_MIN_SIMILARITATE calibrat) -> UL-1.
     result = enrich_suggestions(conn, "schimbat uleiul motor", include_embeddings=True)
     assert result["surse"]["embedding"] == "UL-1", (
         f"embeddings trebuie sa sugereze UL-1, got {result['surse']}"