perf(scan): skip manifest lists on exact snapshot cache hits

[wangyong9999]

Purpose

Closes #375.

Repeated scans of the same snapshot and bucket decode the manifest lists before checking the snapshot live-entry cache. Check that cache after selecting the snapshot, and bypass manifest-list reads and decoding on an exact hit. Misses still rebuild the complete bucket; each scan still applies its request-specific filters.

Store the table file count with the cached entries to preserve skipped-file metrics. An exact hit now reports zero scanned manifests. This targets repeated point-read planning without introducing another cache or changing snapshot discovery.

Tests

• Local C++17 Release build of the identical patch on the downstream main branch succeeded without compiler warnings.
• 88 focused scan, snapshot, predicate, serialization, and bucket-pruning tests passed on that downstream build. Regression coverage checks no manifest-cache access on an exact hit, changed predicates, eviction/rebuild, and retained file counts; existing tests cover snapshot changes.
• pre-commit run --all-files and git diff --check passed on this upstream checkout.
• The downstream CI suite passed. SCM test packages built successfully for both x86_64 and aarch64.
• Upstream workflows require maintainer approval (action_required); they have not run. No standalone end-to-end performance uplift is claimed for this individual planner optimization.

API and Format

No public API, table storage format, or protocol change. Only the transient serialized snapshot-cache payload changes; its magic is versioned so old cache entries fall back to rebuilding. Snapshot selection and cache eligibility remain unchanged.

Documentation

Updated the manifest-entry cache guide to describe bypassing manifest lists and exact-hit scan metrics.

Generative AI tooling

Generated-by: Codex based on GPT-6.