Skip to content

performance-and-budgets.md

Latest commit

 

History

History
50 lines (41 loc) · 2.72 KB

performance-and-budgets.md

File metadata and controls

50 lines (41 loc) · 2.72 KB
Raw

Performance and budgets

The page manager accounts for decoders and tracked bytes across players. A host may configure an explicit lower resource policy; there is no implicit small-asset byte ceiling. A player reserves before allocation, publishes exact category ownership, and rejects when an actual or configured resource boundary is reached. Resource rejection is terminal and reported to the consumer; hidden or reduced-motion players release animation resources as a nonfatal policy condition. None of these paths manipulates application DOM.

Authors own the cost of unit length, canvas/rendition dimensions, and bitrate. Use authored restart runways and measure the real target devices. The runtime never downscales or shortens a valid asset to fit. getDiagnostics() is bounded and suitable for support logs, but verbose tracing, screenshots, synchronous readback, and devtools perturb timing benchmarks.

Normal playback configures the selected decoder and fills the six-frame live presentation ring before revealing frame zero. It does not run the exhaustive all-routes decoder rehearsal during page startup. runAllRoutesReadiness() and the certification application retain that full proof for an explicit, dedicated development or CI candidate; it must not run in the background on a live candidate because it owns decoder generations and probe presentation.

CI performance comparisons are advisory. A named scheduling certificate requires at least 300 post-warm-up outputs at 1.5× authored real time plus exact deadline and boundary rules. Heap, RSS, GPU-process memory, and energy are observational; explicit ownership counters are the leak gate.

JavaScript delivery gates

The release authority measures the exact production output created for a consumer that imports @pixel-point/aval-element/auto. The complete working player—entry chunk, one lazy runtime chunk, and the external decoder worker—must be at most 60,000 bytes with Brotli quality 11. This aggregate is the bytes the consumer must actually ship; source-module totals and per-boundary gzip figures are diagnostic only.

scripts/performance/measure-m8-bundles.mjs builds that consumer with pinned Vite 8.1.4/Oxc, verifies that the worker is self-contained, requires exactly one lazy runtime boundary, rejects duplicated module ownership, and enforces the aggregate cap. The verified 2026-07-17 build measures:

  • entry chunk: 13,376 Brotli bytes;
  • lazy runtime chunk: 40,554 Brotli bytes;
  • decoder worker: 992 Brotli bytes;
  • complete working player: 54,922 Brotli bytes.

That leaves 5,078 bytes of headroom under the release cap. The exact release candidate must be measured again; these figures describe the current checked workspace, not a substitute for candidate evidence.