Luke Parker
1f61eb5ca9
Timeline Layout Continuity
Run from packages/app:
bun run test:stability
The suite runs a production build in one Chromium worker. Selected scenarios use deterministic 4x CPU stress after application readiness. This is a stress profile, not emulation of a specific device.
What It Proves
The continuity probe samples DOM-derived layout and visibility state across browser render opportunities. Tests declare explicit contracts such as:
- Preserve a visible semantic anchor while the user is away from the bottom.
- Preserve end anchoring while active content grows or new content appears.
- Keep adjacent visible rows ordered without material overlap.
- Keep user-selected disclosure state through updates and virtualization.
- Avoid a sampled blank interval while one visible surface replaces another.
- Preserve logical row and control identity where local state or focus depends on it.
- Keep keyboard, wheel, and nested-scroll ownership consistent during remeasurement.
The suite exercises real browser reducer, projection, component, virtualizer, layout, focus, and interaction code. The backend and event producer are controlled fixtures.
What It Does Not Prove
The pass/fail oracle does not inspect every compositor-presented pixel. A sample taken after requestAnimationFrame is a DOM/layout observation, not proof that every sampled state was displayed or that every displayed frame was sampled.
The suite does not provide complete coverage for:
- Compositor-only or raster-only glitches.
- Color, contrast, canvas, WebGL, masks, irregular clips, or arbitrary occlusion.
- Physical display refresh rates, native OS scaling, or a named low-end device.
- TCP packetization, proxy buffering, or the complete real server/provider pipeline.
Playwright video, trace, screenshots, and observation JSON are diagnostic evidence. They are not pixel baselines and do not participate in normal pass/fail decisions.
For optional before/violation/after screenshots, set OPENCODE_STABILITY_CAPTURE=1. Capture is opt-in because compositor readback can perturb timing.
Test Layers
- Projection: admitted rows, grouping, labels, and final visible states.
- Local state: disclosure state, identity, duplicate delivery, and virtualization restoration.
- Interaction: wheel, keyboard, nested scrolling, actionability, and focus behavior.
- Layout continuity: anchoring, adjacency, responsive reflow, and visible surface handoffs.
- Reducer hardening: validly shaped but intentionally reordered, duplicated, removed, or replaced events.
- Oracle contract: pure analyzer and browser sampler calibration tests.
Production-lifecycle fixtures should model states emitted by the current producer. Impossible or reordered sequences belong in reducer-hardening tests and must not be described as normal provider behavior.
Diagnostics
Failures retain:
video.webmtrace.zip- failure screenshot
- sampled DOM/layout trace JSON
- event markers and summarized violations
The analyzer records both unclipped layout bounds and ancestor-clipped visible intersections. Scrollbar and raw scrollTop changes alone do not fail continuity checks; user-visible semantic anchor movement does.