[yashaskn8]

Source-verified causes addressed

• V6 previously kept only aggregate retrieval metrics; equal totals obscured query-level regressions.
• Hit@1 under head-five reranking is bounded by the fraction of queries whose gold evidence is in the first five frozen candidates.
• Development-only comparisons cannot prove production accuracy or independent holdout performance.

Changes

• Add pure CPU-only paired diagnostics with candidate-recall/top-one oracle ceilings, query-level movement, and strict frozen/gold/rank checks.
• Save per-query reports and SHA-256 fingerprints for V5/V6 comparisons to existing approved private Colab output folder.
• Run counterfactual pool-size 5/8/10 analysis using existing scores with no changes to serving.
• Add regression tests and CI gate, plus a no-retraining historical replay runbook.

Boundaries

No model training, no private weights/data uploaded, no automatic model promotion, no change to production retrieval/authorization. CI is synthetic. The previously completed experiment requires a fresh read-only replay to obtain the new per-query artifacts.

Required verification

CI backend suite (MySQL) and frontend build/lint; private V5/V6 reproducible inference, candidate-set hashes and eventual independent holdout are not verified in CI.