feat: rerank stage fused as a weighted RRF leg

[Verdenroz]

Description

The fused top candidates already contain most of the misses: on the golden sets, 8 to 12 targets per set sat at ranks 11 to 44. A cross-encoder can see them, but replacing the order with the reranker's lost as many correct top-10 hits as it gained, so the reranker joins the fusion instead as a third reciprocal-rank leg. It is optional config, off when absent, and the search never gets worse when the reranker is slow or broken.

Changes

• Added an optional [rerank] section: base_url, api_key (optional, env: indirection), model, top_n (default 20), weight (default 1.0).
• Added scry-core/src/rerank.rs: a client for a Jina-style POST {base_url}/rerank and fuse(), which adds the reranker's ranking as a weighted RRF leg over the pool search_with_vector returns, sharing RRF_K with the other legs. Repeated or out-of-range indices are rejected.
• The search route asks for max(limit, top_n) candidates, reranks after the store call, and falls back to the fused order on any error or after six seconds, logging at warn. Scores stay the dense cosine.
• Added rerank to SearchRequest (default true) and --no-rerank to scry search and scry eval.
• Added eval/simulate_rerank.py, which scores replacement, fusion weights, and rerank pool sizes offline from a dump of one run.
• Documented the section in docs/search.md and deploy/config.example.toml. The example keeps [rerank] commented out, so nothing changes for a user who does not add the section.

Testing

• cargo test --workspace: 46 core tests plus the integration suite pass; cargo clippy --workspace --all-targets -- -D warnings clean. New tests: config parsing, fuse() promotion and zero-weight identity, and an integration test with a mock /v1/rerank endpoint checking the fused order and the rerank: false bypass.
• Offline simulation from one dump (bge-reranker-v2-m3 over the top 50, CPU build, copy of the live index at 22,604 chunks), summed recall@10 / MRR over three sets: no rerank 1.934 / 1.290, replacement 1.954 / 1.234, RRF weight 1 top 20 2.064 / 1.383. A natural-language gate lowered MRR everywhere and was dropped.
• Confirmation with this branch's binary and the 6s budget, bge served on the Radeon 890M, serial runs on the same copy, --no-rerank pair per set: scry 0.667 / 0.426 to 0.730 / 0.444, finance-query 0.700 / 0.556 to 0.717 / 0.575, soothfast 0.667 / 0.401 to 0.700 / 0.436. Zero fallbacks in 183 reranked queries.
• The live gain (+0.213 / +0.165 summed) is larger than the simulator's prediction (+0.130 / +0.093). Same direction on every set; the dump's scores came from the CPU build and the confirmation from the Vulkan build, and backend numerics reorder near-tied pairs. The simulator is a direction tool, not ground truth.
• p50 3.2-4.0s and p95 5.9-7.5s were measured while a CPU dump saturated the machine and include HyDE and embedding, so they are ceilings. On a quiet box the rerank call alone takes 1.85s for 20 documents on the iGPU and 9.5s on CPU, where it would fall back.
• Documents truncated to 1200 characters keep most of the gain at half the reranker cost; no knob was added since the iGPU fits full-length documents in budget.
• CI is check-only for this PR: gate.yml has no rerank bench and the route is async.