perf: skip hashes for ineligible configuration stores

[yeahjack]

Current-base specification disposition

Current main: 4283de1599c7da914138f82a99405d67c2c861ec. This PR remains Draft and is not applicable to that base.

• Goal: skip configuration hashes only when their values cannot be consumed
• Constraints: preserve current persistence policy, change detection, debounce, retries and serialization
• Assessment: the current store implementation consumes both hashes on each eligible update, after its existing early-return guard. The secondary-pet and blocked-store branches targeted by this patch are no longer present
• Non-goals: restoring removed store flags/policy or substituting an unrelated optimization
• Acceptance outcome: the original optimization has no valid current-base target. No replacement implementation is proposed; the branch is retained as a draft reference

The implementation, validation and performance results below apply only to the historical 666999650f8afb405fa34dbf1ee0b98cd7145117 base. They are not claims about current main.

Summary

Move update-time configuration hashes inside their existing store-eligibility branches. This avoids computing hashes whose results cannot be used. Hash inputs, change detection, 300 ms debounce, 1 s retry, serialization, loading, and flush behavior are unchanged.

This is an independent change based on 666999650f8afb405fa34dbf1ee0b98cd7145117; it does not depend on the other performance PRs.

Correctness checks

• Full Linux Release diagnostic-backend build, warnings as errors
• All 6 CTests pass, including a new differential test against the frozen upstream implementation
• 1,425 deterministic clock-driven steps, 260 save attempts, and 3,196,535 serialized JSON bytes compared exactly
• Covers all eight secondary/settings-blocked/session-blocked combinations, defaults and full-capacity configurations, edits/reverts, exact debounce/retry boundaries, independent failures, validation changes, restored eligibility, flush, and mixed sequences
• Focused ASan/UBSan passes; core/SDL archives were not instrumented and LeakSanitizer was disabled because the executor's ptrace prevents it
• Repository line, localization, platform-runtime and Cubism-user-model checks pass

Measured component cost

GCC 14.2, -Os and native IPO/LTO, Debian 13 x86_64 cloud VM, Xeon Platinum 8573C, pinned logical CPU 0 (not a dedicated host). Old and new implementations use the same compiler and real production hash functions. Three independent runs each use warmup/calibration and 31 alternating paired samples, with at least a 30 ms calibration target per batch. Process CPU and monotonic wall time are recorded separately.

Pooled wall-time median / p95 of batch-average cost:

Scenario	Before	After
Saturated, session store blocked	616.932 / 1,046.383 µs	390.059 / 783.880 µs
Saturated, both stores blocked	233.395 µs median	4.478 ns median
Default, both stores blocked	344.644 ns median	4.460 ns median

The saturated session-blocked median falls by 36.77%; its three separate runs show 33.57–38.81% reductions. No allocation, cache, persistent field, or structure-size change is introduced.

Important limits: GCC already skips some unused settings hashing in the baseline. Healthy primary/secondary and settings-only-blocked controls show noisy results, including slower samples; this PR does not claim a common-path gain. The full raw data, control ranges, methodology and reproduction commands are in tests/runtime/README.md and tests/runtime/config_store_measurements.csv.

These are steady-state configuration-update microbenchmarks, not input-to-render latency, disk-save latency, FPS, or full-app CPU usage. Cubism SDK is absent, so full Live2D rendering was not built. Native window smoke is blocked by the cloud display-socket restriction; Windows/macOS results must come from CI and are not local runtime measurements.