[user-github-me]

discardWarmupCapture() restores scalar performance counters but shallow-copies capturePerf.frameMs. The inner capture appends to that shared array, so discarded samples remain in the p50/p95/p99 summary on both success and failure.

Copy the timing samples by value before capture. Extend the existing restoration tests to use real sample arrays and verify the full performance summary, including an injected capture that replaces sample contents.

Refs #4435 and the accounting review on #2489. This fixes the exported helper's accounting contract. The helper still has no production call sites; this does not fix blank worker-boundary frames or add a warm-up capture.

Validation:

• Four restoration regressions fail on main with leaked or replaced samples.
• Warm-up, percentile, frame-capture and parallel-coordinator suites: 129 tests passed.
• Engine typecheck/build, repository lint, changed-file formatting and applicable pre-commit gates passed.
• No capture-mode, seek, compositor timestamp or rendering changes.

[jrusso1020]

Good catch. The snapshot copied capturePerf shallowly, so the inner capture pushed its timing into the same frameMs array, and the discarded sample stayed in the percentile summary. Copying the array by value fixes that. The new tests compare the full perf summary on both the success and failure paths, so they would catch a regression. Small and correct, thanks.

— Rames