feat: add retrieval quality benchmark

[davidrobertson]

Summary

• run production session search against a fixed nine-session corpus
• report Recall@1, Recall@5, mean reciprocal rank, category scores, and per-query ranks
• cover exact, cross-session, temporal, and paraphrase retrieval without new dependencies

Baseline

• Recall@1: 0.875
• Recall@5: 0.875
• mean reciprocal rank: 0.875
• paraphrase Recall@5: 0.5

Validation

• npm test (183/183)
• npm run typecheck
• npm run smoke
• npm pack --dry-run
• independent VoltFlow review