Phase 2 oracle: 20-query recall@10 with real SPECTER2 + BGE-M3

[deOliveira-R]

Manual acceptance test required before tagging v1.0.

Per plan §6 Phase 2: the 20-query recall@10 oracle and <500ms cold-start oracle need the real library + real models loaded. `tools/phase2_search_oracle.py` already exists; it needs a curated queries file.

Steps

1. Full Zotero migration first: `grimoire migrate zotero` (~8 min for 15k items).
2. `grimoire index` to populate SPECTER2 item embeddings (~8 min) and BGE-M3 chunks (overnight, CPU-only).
3. Draft 20 queries + gold answers (item IDs or DOIs) from your research notes — the ones you'd actually ask.
4. Wire the queries file path into `tools/phase2_search_oracle.py` if not already supported.
5. Run the oracle against the full corpus; compare keyword / semantic / hybrid.

Acceptance

• 20 curated queries + gold-answer item IDs captured (commit `tools/phase2_queries.jsonl` or similar).
• Hybrid recall@10 ≥ 0.90 per plan target.
• Cold-start semantic query latency < 500ms on the TrueNAS host.
• Regression baseline stored in `project_oracle_results.md` memory or repo docs for future regressions.

Notes

• Small queries file (20 lines) is fine; format as `{query, gold_item_ids: [..], notes: "..."}`.
• BGE-M3 indexing is the slow part — schedule overnight. SPECTER2 alone is ~8 min.