Search quality phase 2: semantic rerank + measured query eval

[anand-testcompare]

Why

Current lexical + facet search works, but relevance quality is not consistently strong for natural cable queries.

Outcome

Improve ranking quality with measurable gains and keep results explainable.

In Scope

• Add semantic rerank over top lexical candidates.
• Preserve deterministic facet filtering behavior.
• Add a stable benchmark query set and report precision@k + p95 latency.
• Normalize/validate price values used by filtering and ranking.

Out of Scope

• Full retrieval architecture rewrite.
• New external search infrastructure.
• UI redesign beyond minimal benchmark/report visibility.

Implementation Plan

1. Add rerank stage after lexical retrieval (top-N candidate window).
2. Expose ranking debug signals so ordering is inspectable.
3. Add benchmark runner (fixed query set + golden expectations).
4. Add price normalization guards and explicit unknown handling.
5. Add regression tests for ranking and filter behavior.

Test Plan

• convex: deterministic ranking fixture tests for lexical-only vs rerank.
• convex: normalization tests for malformed/missing/unknown prices.
• manual: run benchmark script and record precision@k + p95 latency deltas.

Acceptance Criteria

• Stable ranked output for benchmark queries.
• Benchmark report includes precision@k and p95 latency for lexical-only and rerank modes.
• Price ranking/filtering uses normalized values with explicit unknown semantics.
• CI contains at least one deterministic search-quality regression test.