Brain search ranking (brain_search)

Local hits use SQLite FTS5 bm25 (brain_search.search_local): lower score means more relevant.

When the MCP tool passes prefer_stack (list of strings, e.g. python, react):

1. The local path may return up to min(limit × 5, 100) candidates so a strong stack match is not dropped before reranking.
2. After merging with optional Notion fallback hits, results are sorted by effective score
   effective = bm25_score − stack_boost, where stack_boost grows when a row’s stack JSON overlaps a preferred label (exact or substring match). Boost is capped so bm25 still dominates for distant matches.
3. The final list is truncated to limit.

Empty or whitespace-only query: handlers return an immediate message; search_with_fallback returns no results and a warning explaining that the query is empty.