perf: bound large retrieval snapshot reads

[suguanYang]

Summary

• Bound the small-corpus probe to top_k + 1 rows so it never downloads every serving manifest just to decide routing.
• Use revision-pinned normalized section/chunk reads for large namespaces instead of transferring hundreds of compressed manifests over asyncpg.
• Keep manifest loading for small namespaces and batch manifest reads for bounded payloads.

Validation

• Focused retrieval contracts: 12 passed.
• Ruff: passed.
• Pyright: 0 errors.
• No production writes; production read-only reproduction was used only for diagnosis.

This addresses the timeout observed before map-nav/LLM execution when a namespace contains hundreds of active revisions.