fix(studio): the timeline scroll gate catches dropped frames in the virtualized view

[miguel-heygen]

The timeline viewport gate now catches a scroll that drops frames in the virtualized timeline on CI. A 45 ms stall on every scroll step used to pass it.

What changes

• The CI virtualized arm gets its own pair of limits. Pooled over all 315 measured scroll steps, the interaction p95 must be ≤58.3 ms (3.5 frames) and the frame-interval p95 ≤25 ms (1.5 frames). In plain terms, at most 5% of scroll steps may drop a frame, and at most 5% may take 4 or more frames.
• The unvirtualized arm keeps 75/75, because its normal interaction p95 is already 49-60 ms. The primary tier keeps 50/33.3, and the throttled and high-DPR tiers keep 75/75.
• Which pair applies is decided in one place, responsivenessLimits in the verdict module. The gate script reads it from there.
• One rerun per arm. When an arm's timing attempt fails (3 warm-up runs dropped, then 5 measured runs), the gate measures that arm once more with the same protocol and fails only if both attempts fail. The loop is inside the gate script, not a CI retry. Direct-scroll and memory checks run once. The step summary prints every attempt's p95s, limits and verdict.
• Per-step samples in the artifact. The timeline-viewport-gate-evidence artifact keeps every measured step's interaction and frame interval, per arm and per attempt (attempts[].runs[].interactions / frameIntervals), so a red can be recomputed later.

Why these numbers

A pooled p95 over 315 steps fails once 16 of them are over the limit.

False fails on normal builds. Every CI run today that carries per-step evidence: 68 normal virtualized attempts from 67 PR runs, judged single-shot.

Virtualized limits (interaction / frame)	Attempts that fail	Most steps over the interaction limit
75 / 75 (main)	0 of 68	3
41.7 / 25 (first pick)	16 of 68 (23.5%)	39
58.3 / 25 (this PR)	0 of 68	6

At 41.7 ms the fails were all on interactions: 17-39 steps at 3 frames, spread across all five runs (a slow runner, not a slow start), and frame intervals stayed under the limit (at most 9 steps over 25 ms). The earlier analysis predicted 0% at 41.7 from too little per-step data, so the fallback rule fired before merge and the interaction limit stepped to 58.3 ms. With the rerun a false fail needs two failed attempts in a row; at 58.3 none of the 68 had a single one.

Catching the 45 ms stall, with the rerun. A throwaway build injected a 45 ms stall into every timeline scroll step and ran the virtualized arm 10 times on CI.

• 20 of 20 attempts failed, so 10 of 10 gate runs failed.
• Every attempt had 270 of 315 frame intervals over 25 ms (every step that scrolls stalls), against 16 needed.
• Interactions over the limit: 142-193 of 315 over 41.7 ms, and 45-47 over 58.3 ms. Either half alone fails the stall.
• Single-attempt detection d: 20 of 20 observed, with a margin of 270 against 16. With the rerun the gate fails the stall at a rate of d², which is also about 1, above the 95% bar. Counting attempts alone, 20 of 20 bounds d at ≥0.86 with 95% confidence; the step margin is what carries it.

CI time per arm. A normal pass runs one attempt, unchanged: about 40 s for the virtualized arm and 35 s for the unvirtualized one. A failed attempt adds one more, about 17 s on a normal build and 30 s on the stall build (a stall arm takes 67-68 s end to end). The steps per run are unchanged.

False fails are counted per arm, each against that arm's own limit: the unvirtualized arm's against its 75 ms, never toward the virtualized limit. For example, #4836's unvirtualized arm read 82, 82, 39, 49 and 47 ms interaction p95 per run on a busy runner; that counts against 75 ms.

This PR's own CI. The virtualized arm applied 58.3 / 25 and passed on its first attempt: interaction p95 33.2 ms with 0 of 315 steps over 58.3, and frame p95 16.8 ms with 1 of 315 over 25. The unvirtualized arm passed at 75 / 75 (61.8 / 33.4 ms).

Tests

timeline-viewport-verdict.test.mjs:

• The CI virtualized arm: 16 of 315 steps two frames late fail and 15 pass; the same for frame intervals that drop a frame. Both fail with the virtualized limits set back to 75.
• The rerun: a failed attempt followed by a passing one passes; two failed attempts fail; a third attempt is never counted.
• The other arms and tiers stay at their old limits.

Before

Main's limits on recorded CI samples: the 45 ms stall run passes.

After

This branch's limits and rerun on the same samples: both stall attempts fail, and a normal run still passes.

[github-actions]

Edit accuracy: 835 passing here, 835 on the base branch

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (9)

• sequndo-none-center-r0-nested-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): fail, fails
• sequndo-none-center-r0-root-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): pass, passes
• sequndo-none-pct-r0-nested-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): pass / pass / fail, passes
• sequndo-none-pct-r0-root-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): pass / pass / fail, passes
• sequndo-none-px-r0-nested-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): fail / fail / fail, fails
• sequndo-none-px-r0-root-z100 (fixed by fix(studio): an undo that lands mid-drag keeps the box under the pointer #4853): pass / pass / pass, passes
• seqnudge-none-pct-r0-nested-z100 (fixed by fix: a save's temp file no longer reaches Studio as an outside change #4857): pass, passes
• seqnudge-none-pct-r0-root-z100 (fixed by fix: a save's temp file no longer reaches Studio as an outside change #4857): pass, passes
• seqrepeat-none-px-r0-nested-z100 (fixed by part C (fix(studio): undo right after an edit undoes that edit, even before its save lands #4807 stack)): pass / pass / pass, passes