Phase D follow-up: formal shadow evaluation + relevance acceptance for hybrid retrieval

[SweetSophia]

Split from #261.

Context

Hybrid retrieval is merged (#261 capability), off by default (ADR), and live on the reference deployment since 2026-09-03 (llama.cpp local embeddings, coverage 99.3%, recall serving with zero lexical fallback). The rollout went coverage-gate → serve directly.

What did not run

Phase D specified two acceptance gates that were skipped, not passed:

1. Shadow evaluation — run hybrid and keyword rankings side-by-side on real query traffic (or a sampled set), log both, compare while keyword remains the user-facing path.
2. Relevance/quality acceptance — a scored query set (gold answers or graded relevance judgments) with an agreed threshold that hybrid must meet or beat keyword.

Proposed acceptance

• Shadow harness: dual-rank a query corpus without serving hybrid results; persist per-query rankings + metrics
• Scored query set: ≥ N queries with graded relevance, versioned in-repo
• Metric report: recall@k / nDCG or equivalent, hybrid vs keyword, published as evidence
• Threshold decision documented (accept / tune / reject), covering the keyword-only fallback path for over-context articles
• Results recorded for the reference deployment profile (nomic-embed-text-v1.5 Q8_0, 768-d)

Non-goals: changing the default-off posture; remote embedding providers; Ollama (excluded by policy).