Phase 3: Brain search enrichment for review & scan — inject semantic context into LLM prompt #458

Goal

Inject brain search results (FTS5 + vector + graph RRF fusion) as context enrichment into both cora review and cora scan LLM prompts.

Context

• cora scan currently has ZERO context enrichment — only sends raw file content to LLM
• cora review has context chain (regex-based) but no brain/semantic context
• brain_search() in src/index/brain.rs provides hybrid search (FTS5 + usearch vector + graph BFS → RRF fusion) — only used in CLI and MCP, never in review/scan

Scope

1. Add enrich_with_brain_search() in new file src/engine/brain_enrichment.rs
2. Extract key symbols from diff (review) or file batch (scan)
3. Query brain_search for each symbol, deduplicate, format under token budget
4. Inject as "## Code Intelligence (brain search)" section in LLM prompt
5. Config: brain_search.enabled, max_results_per_symbol, max_tokens

Files

• NEW: src/engine/brain_enrichment.rs
• src/engine/review.rs — inject brain context before LLM call
• src/commands/scan.rs — inject brain context before LLM call

Benefits

• Review gets semantic context (not just syntactic)
• Scan gets context it never had before
• LLM can assess risk more accurately with related code visibility

Dependencies

• Requires Wire vendored nomic-embed-code (pretrained-embed) to brain search — replace hashing-trick 256d #455 (better embedding → better brain search quality)
• Requires Phase 1: Wire symbol index to context chain — replace regex-based resolution with SQLite queries #456 (IndexBridge)

Effort: 4-5 days