perf(retrieval): batch map-nav payload loads and add stage timing

[suguanYang]

Summary

Optimize default map-nav retrieval execution while preserving exact retrieval quality.

Changes

• Batch lazy payload loading per document with one ordered SQL query instead of one query per section.
• Stream each document through the existing BM25/RRF scorer and release payloads before the next document.
• Preserve BM25 formulas, channel weights, RRF ordering, tie-breaking, asset remounting, duplicate chunk semantics, and hydration references.
• Add stage timing logs for snapshot load, episode, and hydration to identify remaining production latency.
• Keep the psycopg2 DSN normalization fix that resolved the production-only 500 from the original incident.

Verification

• Full repository suite: 679 passed, 14 warnings.
• Ruff and Pyright passed.
• Real local 60k PostgreSQL corpus: about 10 seconds total; exact eager/lazy map and unit score parity.
• Real local 600k PostgreSQL corpus: about 107 seconds total and 672 MiB peak RSS; exact top-result parity. This is materially better than the previous roughly 176-second run, but still above the 60-second ALB target; further scale work remains tracked in Optimize map-nav retrieval for 600k+ chunk namespaces #340.

Quality guarantee

No ranking, candidate-selection, evidence-budget, or result-assembly policy changes are included. The implementation changes only loading and release mechanics, with exact parity contracts.