Caveman check first. Verify code now present, deps work, vectors built. Then run recall test.

Code present (548411c9). Deps load fine. vault.db query failed only because inline didn't load extension. Check vec count with extension loaded + run real recall test via vault_mcp.

vec_memories=117 (up from 56). No `memories` table — FTS table named differently. Load real vault MCP tool, run recall test.

Works but unclear if vector leg active in running server. Test _hybrid_search directly + run semantic bake-off.

Wiring correct. `_vec_embed` is the embedder. Confirm it loads + run bake-off (semantic queries keyword misses).

Vector leg confirmed live (`_vec_embed loaded: True`, `vector_available: True`). HF logs buried results. Re-run quiet.

Validation complete. Hybrid mechanically live on Luci, recall quality modest — matches ticket's decision-flag prediction. **Findings:** - ✅ Code present: `_hybrid_search` + vec0 wiring in `vault_mcp.py` (commit 548411c9, origin/master). - ✅ Deps work: `sqlite_vec` v0.1.9 loads, `enable_load_extension` allowed on system python3. - ✅ Vectors built: `vec_memories=117` (up from 56 earlier). - ✅ Vector leg live in server context: `_sqlite_vec loaded=True`, `_vec_embed loaded=True`, `_vector_available=True` — hybrid RRF fires, not FTS-only. - ✅ Graceful fallback intact: `_vector_available()` guards (line 153-162); FTS-only path returns cleanly if embedder/table absent — no crash. - ✅ `memory_search` MCP tool returns end-to-end. **Bake-off (5 semantic queries, hybrid):** - "cost control switching AI model providers" → `glm_endpoint`, `provider_switch` at #1/#2 — strong hit. - Others (lights, telegram polling, flock) → related-but-not-bullseye. Mixed. **Decision flag — value is LOW-MODEST for Luci, recommend keep-enabled-as-is:** Luci's corpus is ops/PM feedback memories, not Lucienne's curated semantic notes. The 10/10 lift Mac saw does NOT cleanly replicate — Luci's queries are mostly exact-keyword ops lookups where FTS already wins. Hybrid adds a one-time ONNX embedder cold-start (~10-20s) on first `memory_search` per MCP process, then cached. No regression, graceful fallback, vectors built. Nothing to commit (vault.db gitignored, local-only embed). Leaving enabled — cost is one cold-start, downside is nil. Not worth further tuning for an orchestrator. REVIEW: Hybrid recall validated + live on Luci (vec_memories=117, vector leg confirmed); recall lift modest as decision-flag predicted — keeping enabled, no further build warranted.

REVIEW: Hybrid recall validated + live on Luci (sqlite_vec v0.1.9, vec_memories=117, vector leg confirmed, graceful FTS fallback intact). Code already on origin/master; nothing to commit. Recall lift modest as decision-flag predicted — Luci corpus is ops/PM, FTS already wins most queries. Recommend keep enabled as-is, no further tuning. Decision for Elmar: keep, or disable to save first-query ONNX cold-start. {"mc_signal":"REVIEW","ticket":"MC-4323","nonce":"MC_TICKET_TURN:4323:1780028461430:85a1f451","payload":"Hybrid recall validated + live on Luci (vec_memories=117); lift modest as predicted, recommend keep-as-is, no further build"}