Measure your target.

What is measured?

Each workload warms up for 20 frames, then records the requested sample count. Static HUD and responsive layout use 96 cards, telemetry uses 48 cards updated every six frames, and inventory scroll uses 400 retained rows with deterministic focus targets. Two root-local atlas registrations exercise an ASCII primary face and Latin-1 fallback. The responsive workload alternates a 360/960 CSS-pixel width, constrained by the available container. Five scene creation/teardown cycles follow.

CPU frame time is elapsed time around ui.frame(), including layout, painting, accessibility synchronization and WebGL submission. Frame intervals are requestAnimationFrame timestamp differences. Render-to-commit records scheduling latency through a React layout effect. These are distinct measurements; none measures GPU execution or confirms presentation. Raw samples use milliseconds and summaries use nearest-rank percentiles.

The initial setup duration uses the browser's existing cache state. No cold-cache claim is made. The harness does not measure heap or GPU memory; removing every accessibility mirror is only a teardown check. A hidden tab or lost context stops the run. Set budgets for your device and workload before running; absent budgets are reported as not-set.

Reports identify repository source, which can differ from the npm release with the same package version. A revision ending in -dirty means local changes were present at build time. CI uses short software-rendered smoke runs to check the report contract and cleanup; timing thresholds are left unset.

Recorded desktop baseline

A Chrome 156 / Apple M3 run acquired 120 samples per workload and five teardown cycles. CPU frame-call p95 was 9.5 ms for the static HUD and 26.9 ms for the 400-row inventory. These UI-only results motivate profiling large collections; they do not establish the cause of the difference. Budgets were unset. Read the environment, source revision and qualification status, or download the original JSON and checksum before comparing results.

Complete device qualification

Also exercise both playable integrations and the component catalogue, including touch, rotation, background/resume, context recovery, screen readers and a 30-minute session. Follow the qualification protocol and measurement instructions. Physical iOS/Android and assistive-technology qualification remain incomplete until actual results are acquired.