fix(retrieval): bound map-nav snapshot memory and SQL loads

[suguanYang]

Summary

Hotfix the production retrieval timeout/memory failure for namespaces around 60,000 chunks.

• Load map-nav chunk metadata/index in bounded 10,000-row keyset pages instead of one oversized snapshot query.
• Lazy-load section payloads during scoring and release each section after use.
• Stream corpus BM25/RRF scoring without retaining all raw chunk payloads.
• Preserve the existing BM25 formulas, channel weights, RRF fusion, ordering, tie-breaking, and hydrated retrieval inputs.

Verification

• Focused retrieval contracts: 3 passed.
• Full repository contracts: 651 passed, 14 warnings.
• Ruff and Pyright passed.
• Real local 60k PostgreSQL corpus: lazy retrieval completed successfully with substantially lower peak RSS; eager and lazy map/unit scores and selected results were identical.
• Real local 600k PostgreSQL corpus: bounded loading completed without SQL timeout or OOM; this measured 111.5s loading, 64.1s scoring, 175.6s total, and 645.7 MiB peak RSS. The larger-scale optimization is tracked separately in Optimize map-nav retrieval for 600k+ chunk namespaces #340.

Quality guarantee

No ranking or candidate-selection behavior is intentionally changed. Exact eager/lazy parity is covered by contract tests.