mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-11-17 23:47:33 +00:00
`sweep_static` is hardcoded `error`, exempt from persistence tiering, and gates the run. It fires whenever the geometry+opacity fingerprint is identical across every sample — so a held title card or logo lockup had no passing spelling at all. Measured: a fully lint-clean static card, GSAP timeline registered, paused and seekable, still failed `check`. `data-layout-allow-static` on the composition root now demotes the finding to `info`. It is not suppressed: the claim is the author's and can be wrong, and dropping the finding produced a confidently green report in which every layout, occlusion and contrast verdict came from one repeated frame with no trace left. `info` does not gate even under `--strict` and survives `--json`. Root-scoped, resolved by the same selector chain the fingerprint itself uses, so the opt-out and the measurement agree on what "the composition" is. A first attempt reused `data-no-timeline` and read it document-wide; review showed both halves were wrong. Its meaning at the engine and lint is per-host "does not register window.__timelines", not "motionless" — and this repo ships `css-spinner-render-compat` and `raf-ball-render-compat`, which carry it and DO animate via CSS keyframes and rAF. Those are exactly the unseekable cases the guard exists to catch, so that version silenced it where it matters most; a stray attribute on any child disabled the guard for the whole run as well. The predicate lives in layout-audit.browser.js beside the fingerprint it must agree with, so the root-only rule is unit-testable — reverting it to a document-wide query now fails two tests instead of passing all of them. Docs updated: the escape-hatch list and the "no way out" paragraph in lint-validate-inspect.md, a row in data-attributes.md, and cli.mdx. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>