Add real retrieval baseline benchmark

[princerm06]

Summary

• freeze 25 natural-language retrieval queries against the current 21-page Synapse dataset
• add a live runner using the existing embedding + pgvector + 4B deduplicated retrieval pipeline
• preserve the first measured baseline: Hit Rate@5 0.800, Recall@5 0.800, MRR 0.800
• diagnose that pages 1–4 have zero chunk rows, explaining the benchmark misses
• add a transactional, explicit page reprocessing utility that rebuilds chunks from already-stored canonical page content
• default the repair command to dry-run; require --apply before any DB mutation
• add unit tests and CI coverage for the reprocessing service

Methodology

Ground-truth labels were frozen before running Synapse. The benchmark itself remains unchanged after discovering the legacy-data inconsistency, so the post-repair run can be compared directly with the initial baseline.

Safety

No credentials, migrations, deployment changes, or automatic production/database writes. The live benchmark and reprocessing command remain local. Reprocessing is limited to page IDs explicitly supplied by the user and commits transactionally.

First measured baseline

• Queries: 25
• Hit Rate@5: 0.800
• Recall@5: 0.800
• MRR: 0.800

Pages 1–4 were then verified to have zero chunks. The next local step is to dry-run and then rebuild only pages 1–4, followed by rerunning the exact frozen benchmark.