Paired performance and cleanup

Run yarn test:package first. Set ARTPLAYER_BROWSER_ARTIFACTS to the resulting refactor/.cache/packages/run-*/browser-artifacts.json, then run yarn test:performance. The command starts its own local server; do not run it alongside other browser tests or CPU-heavy builds. All three engines run with one worker, no trace/video recording and the same media and viewport within a pair.

Each engine runs three alternating published/candidate pairs. Each side retains one warm-up and five timed samples for both core and core + chapter, followed by separate instrumented resource probes. scripts/performance-fixture.mjs adapts the frozen BASE-06 fixture's script URLs, report transport and enforcement of its minimum observation window. The frozen files and all timed samples remain unchanged, and their existing baseline tests still run.

The candidate must come from an isolated package installation. Artifact hashes, source snapshots, package inputs and build tools are checked before measurement; an old candidate cannot silently stand in for changed source. Raw, gzip 9 and Brotli 6 compare matching modern, legacy and ESM files from the same installation. Timing uses the exact JS files selected by the artifact map.

Reports are in refactor/.cache/performance/run-<engine>-*/report.json and summary.md. Playwright evidence is in refactor/.cache/performance/browser/. CI uploads both and appends summaries to the workflow run. Failed or incomplete runs remain recorded and are not promoted to a passing baseline.

A successful test proves the measurement contract and candidate resource cleanup. review-required means size or timing thresholds were exceeded; it is not release approval. CORE-22 and release reviews must resolve the corresponding risks before publishing. These counters do not measure native heap/GPU memory, real mobile hardware or screen-reader behavior. See refactor/coverage-performance.md.

Observation window

The current performance fixture adapter enforces the existing 350 ms teardown observation using the measured clock. It re-arms the probe's original timer if the nominal delay returns before that minimum, without rounding up reported time or relaxing validation. This applies equally to both variants and leaves the frozen BASE-06 fixture and all timed samples unchanged. See test/performance-report.test.js and refactor/changes/2026-09-14-ENG-13-observation-window.md.