feat: Phase 3 — Brain Mode hybrid search (usearch + RRF)

[ajianaz]

Summary

Phase 3: Brain Mode — hybrid code search combining 3 signals with RRF fusion.

What's new

Component	Description
CodeVectorIndex	usearch HNSW vector index with fs2 file locking, key↔symbol mapping
brain_search()	FTS5 + vector KNN + graph BFS → RRF (k=60) fusion
Schema v4	embedding_tier, embedding_dims, embedding_model, last_embedded_at on projects
cora brain <query>	CLI command with --json and --limit flags
cora.brain_search	MCP tool for LLM-driven code search
Index-time embedding	Static token embeddings (256d) computed during cora index

Architecture

cora index
  → walk files, parse symbols, build FTS5 + graph
  → embed_project(): load all symbols, embed via static tokens, insert to usearch

cora brain "error handling"
  → FTS5 keyword search (top 50)
  → usearch KNN cosine (top 50)
  → graph BFS depth-2 from top-5 FTS hits
  → RRF fusion: score += 1/(60+rank) per signal
  → sorted results with provenance (fts, vector, graph)

Files

• src/index/vector.rs (384 lines) — CodeVectorIndex: usearch wrapper with CRUD, persistence, locking
• src/index/brain.rs (257 lines) — Hybrid search engine + index-time embedding
• src/index/schema.rs — Schema v4 migration
• src/index/mod.rs — Wire embed_project into index pipeline
• src/main.rs — Brain CLI command
• src/mcp/tools.rs — MCP tool registration + handler
• Cargo.toml — usearch v2, fs2 v0.4

Test results

• 692 tests pass, 0 failures
• Clippy: 0 errors (16 dead_code warnings — expected for Phase 3 scope)
• Manual test: cora index → 1734 symbols embedded, cora brain "TokenEmbedding" → top result = TokenEmbedding struct with fts,vector signals

Closes #359