feat: add Voyage AI embedding search

[BYK]

Phase 5: Vector embedding search (depends on #49)

Adds semantic vector search using Voyage AI's voyage-code-3 model, layered on top of the existing BM25 + RRF fusion pipeline.

How it works

1. On knowledge create/update: fire-and-forget embedding via Voyage AI API → stored as Float32Array BLOB in SQLite
2. On recall search: embed the query → brute-force cosine similarity over all knowledge BLOBs → feed vector results into existing RRF fusion as another ranked list

Why Voyage AI?

• voyage-code-3 is code-optimized — best-in-class for technical/code text retrieval
• 200M free tokens — effectively free forever for <100 knowledge entries
• Single API key (VOYAGE_API_KEY), OpenAI-compatible response format
• Supports `input_type: "document"" vs "query"" for retrieval-optimized embeddings

Why pure-JS cosine, not libSQL/sqlite-vec?

Tested both. bun:sqlite is standard SQLite — no vector_distance_cos() or DiskANN. The @libsql/client-wasm WASM package also lacks vector functions. Native libsql works but adds a ~15MB native dependency — overkill for <100 entries where brute-force cosine takes microseconds.

Graceful degradation

• Gated behind search.embeddings.enabled (default: false) + VOYAGE_API_KEY env var
• If API key missing or API fails: FTS-only search continues working normally
• embedKnowledgeEntry() is fire-and-forget — embedding failures are logged, never thrown
• Entries without embeddings are simply excluded from vector search

RRF integration

Vector results use the same k: key prefix as BM25 knowledge results → RRF merges rather than duplicates. An entry found by both BM25 and vector search gets a higher combined RRF score.

Config

{
  "search": {
    "embeddings": {
      "enabled": true,
      "model": "voyage-code-3",
      "dimensions": 1024
    }
  }
}

Test coverage

• 15 new tests: cosine similarity (identical/opposite/orthogonal/zero vectors), BLOB round-trip (single/empty/1024-dim), vectorSearch (sorting/limit/null embeddings/confidence filter), isAvailable, config schema
• 281 total tests, 0 failures