refactor(vector/search): extract RerankStage/RerankPipeline, port HNSW Stage-2 (#650 PR-1)

[mosuka]

Summary

PR-1 of the #650 rerank-pipeline campaign: extracts Stage-2 rerank (Issue #481) from its inline match arm in the HNSW searcher into a shared, composable abstraction — with zero behavior change, proven bit-for-bit by a frozen oracle.

• New vector/search/rerank.rs (crate-private): the object-safe RerankStage trait (a stage's per-query preparation coincides with its single per-query invocation, so it folds into rescore — the object-safety deviation from the perf(vector/search): two-stage rerank exists only for HNSW; pipeline is hard-coded #650 sketch, recorded on refactor(vector/search): RerankStage/RerankPipeline abstraction, port HNSW Stage-2 (#650 PR-1) #931), the SoA RerankCandidates buffer (parallel doc_ids/distances, crate-wide distance-ascending + doc-id-tiebreak sort per bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927/bug(vector/search): per-searcher similarity sort underflows at long range — top-k membership becomes arbitrary #933), and the RerankPipeline driver: stage i consumes the top top_k * factors[i] of the previous basis, and run reports whether the FINAL stage applied — exactly the score_basis = "f32-rerank" condition the multi-segment fan-out keys on (bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927).
• F32SidecarStage: the former inline arm as a stage — field position index resolved at construction, query prepared once, rescoring against pool.f32_slice_at, docs absent from the sidecar dropped (pre-refactor continue semantics).
• HNSW port: finalize_graph_results builds a one-stage pipeline from (rerank_factor, reader.rerank_storage()), feeds it the ascending int8 candidates (matching the pre-refactor into_sorted_vec().take(widened) prefix), takes rerank_applied from the pipeline's return, and the inline arm is deleted. No public API change.

Zero-behavior-change evidence

• Frozen oracle (rerank_pipeline_oracle_test.rs, captured against the pre-refactor implementation on the post-bug(vector/search): per-searcher similarity sort underflows at long range — top-k membership becomes arbitrary #933 baseline): rerank-on results equal an independently-computed exact-f32 expectation bit-for-bit (doc ids + similarity bits); rerank-off stays exact-ordered and bit-deterministic; single-segment and multi-segment-with-stale-duplicate fixtures. Green after the port. (Writing this oracle is what exposed bug(vector/search): per-searcher similarity sort underflows at long range — top-k membership becomes arbitrary #933, fixed first in PR fix(vector/search): select per-searcher top-k by distance, not underflow-prone similarity (#933) #934.)
• Existing vector_rerank_engine_test (Stage-2 applied + silent Stage-1 fallback), the bug(vector/search): cross-segment scores are not comparable — out-of-range queries clamp to similarity 1.0 per segment #927 comparability suite, and the bug(vector/search): per-searcher similarity sort underflows at long range — top-k membership becomes arbitrary #933 distance-sort suite: green unchanged.
• 3 unit tests for pipeline mechanics (tiebreak sort, per-stage narrowing + last-stage reporting, absent-stage pass-through).

Verification

• cargo fmt --check and cargo clippy --all-targets -- -D warnings clean on stable and 1.97.0
• cargo test -p laurus --lib: 1261 passed; --tests: all 66 binaries ok
• wasm32 cargo check -p laurus-wasm: clean; cargo doc --no-deps -p laurus: 73 warnings (baseline)

Closes #931. Part of #650; unblocks #932 (Flat/IVF adoption) and keeps #673's 3-stage chain expressible.