[nichinichisou0609]

Problem

A whole-chart cache loses useful work when unrelated master metadata or a skill catalog changes. It also cannot retain an expensive completed subprogram if a later measurement fails.

Changes

• Add source-versioned, content-addressed caches for compiled expectations, deterministic runs and lottery propagation curves.
• Identify actual initialized interpreter inputs and complete play/rank schedules. Normalize naming-only skill/effect IDs while preserving alias relationships, resolved predicates, native execution order and lottery proof dependencies.
• Keep current input validation, metadata, kind/shape catalogs, check decks and final statistics assembly outside the cache. New kinds, new shapes and catalog index shifts can reuse existing programs.
• Persist raw integer results and binary64 endpoint bits with schema/key/source checks and a payload checksum. Commit each completed program atomically; repair corrupt records on demand and retry temporary-file collisions after an interrupted process.
• Expose ChartStatsCache, chart_stats_with_cache, document_with_cache and --stats-cache, with work counters on stderr. Existing uncached APIs and chart-stats/3 output remain unchanged.
• Start a new cache namespace without importing older whole-chart caches.

Validation

• cargo fmt --all -- --check
• Both CI clippy variants, including native-fixtures, with -D warnings
• 10 program identity tests, 6 storage/recovery tests and all 22 chartstats integration tests
• Four native incremental cases: metadata/growth changes, a new ordinary kind, changed real skill values, and a new member shape shifting support shape IDs
• Three real charts covering Just, Combo and Luck: uncached, cold and warm output is byte-identical to each other and to the pre-change main baseline. The warm run records 1,023 hits, zero computations and zero writes.

Full regression and snapshot validation

Full workspace cargo test --release: 813 passed, zero failed, 5 ignored. The ignored tests are the repository's existing manual fixture/corpus exports. The validation build used a temporary single-codegen-unit setting; repository build profiles are unchanged.

A complete real snapshot update grows from 348 charts and 37 aptitude shapes to 352 charts and 38 shapes, including a changed shape and shifted catalog indices:

Run	Cache hits	Computed units	Writes
Old snapshot, cold cache	14,690	96,562	96,562
New snapshot, reuse old cache	95,940	6,644	6,644
New snapshot, warm cache	100,676	0	0

The incremental, warm and uncached new-snapshot outputs are byte-identical: 12,668,775 bytes, SHA-256 7b48dda69ee9e828ac78bdc383f466fa572c733de3d4a1908d86c4d0dd266448. Counters represent operations across the three cache tiers, not chart counts; an expectation hit avoids lower-tier requests.

The actual Python extension in StarMoe-org/nnnotes#21 also passed warm reuse, seed-prefix/provenance-only changes, invalid-original-input rejection, interrupted-build recovery and one-record corruption recovery.

Integration

The coordinated producer is StarMoe-org/nnnotes#21. Nominal expectation consumers are empty-sekai/ournotes-player#18 and nichinichisou0609/moenotes#7. Production replay/recommendation bundles require a release with matching model commit and source identities; this PR does not publish a release.

See docs/chart-stats-cache.md for the identity boundary, APIs, counter meanings and storage lifecycle. The cache is disposable and has no automatic garbage collection; obsolete source namespaces can be removed independently.