feat: cross-encoder reranker (deferred)

[Lucas-Bur]

What to build

Add a cross-encoder reranker to the query pipeline. After RRF fusion (#81), run a small cross-encoder model on the top 20-50 results to re-rank them. Cross-encoders see (query, chunk) as a pair, enabling interaction-aware scoring that dense embeddings miss.

Decision: Defer implementation

• Now: Skip (C). RRF + header-lex boost provides strong baseline.
• Later: Evaluate tiny-reranker-v1 (upcoming, ~70-90 MB ONNX INT8, distilled from Qwen3-8B). If benchmarks show >10% quality gain over RRF alone, integrate.
• Alternative: Qwen3-Reranker-0.6B (573 MB) is available now but too large for MVP.

Acceptance criteria (for future implementation)

• Reranker runs on top N results after RRF fusion
• Configurable via config.reranker.enabled (default: false)
• Model auto-downloads on first use
• Latency impact documented and acceptable
• Adapter test verifies reranking improves score order vs RRF-only baseline

Blocked by

• feat: hybrid search with BM25 + RRF fusion #81 (Hybrid search with BM25 + RRF fusion)
• Release of tiny-reranker-v1 or equivalent lightweight model