Epic: Hybrid retrieval (embeddings, RRF, recency) #15

Keyword search (FTS5) misses paraphrases ("auth" vs "login"). Add optional embeddings without breaking the zero-LLM default.

Scope (to be broken into issues after the MVP)

• Optional embedding providers: Ollama / any OpenAI-compatible endpoint, configured by env vars; none by default.
• Vector index with sqlite-vec in the existing SQLite db (rebuildable).
• Reciprocal Rank Fusion of FTS + vector results; recency boost.

Evidence

• sqlite-vec 0.1.9 loads in better-sqlite3 13.0.3 (verified 2026-10-30); pre-v1, expect breaking changes: https://github.com/asg017/sqlite-vec
• RRF (Cormack et al., SIGIR 2009): https://dl.acm.org/doi/10.1145/1571941.1572114
• ai-memory uses RRF over FTS5 + embeddings: https://github.com/akitaonrails/ai-memory