Transcript geometry gate fails on geometry-mixed-24-turns (two cold reader slips) #5722

What happened

The test job step "Transcript geometry invariants" fails intermittently on main:

{"scene":"geometry-mixed-24-turns","trial":0,"coldHeightDrift":5945,"warmHeightDrift":0,"coldReaderDisplacement":365.375,"coldReaderSlips":2,"warmReaderSlips":0,...}
ExpectError: geometry-mixed-24-turns: unmeasured rows moved the reader, worst 365.375px

The gate allows at most one reader slip on the first cold upward sweep over unmeasured rows (coldReaderSlips <= 1, scripts/perf/geometry-ablation.mjs). Sometimes the sweep produces two.

How to reproduce

The step runs xvfb-run -a node scripts/perf/geometry-ablation.mjs --assert-stable against apps/desktop/storybook-static. The scene is the geometry-mixed-24-turns story in apps/desktop/stories/app-shell.stories.tsx.

It is intermittent. On recent main pushes:

The same failure also hit an unrelated PR (#5721, run).

Every CI failure reports exactly the same numbers (coldReaderDisplacement 365.375px, 2 slips, coldHeightDrift 5945). Every pass stays at one slip or fewer. On the #5721 branch, a rerun of the failed job failed again with the same numbers.

Locally (macOS, npm run build, npm --workspace @maka/desktop run build-storybook, then GEOMETRY_REPETITIONS=1 node scripts/perf/geometry-ablation.mjs --assert-stable), a clean main at f5aa3f0 fails 3 of 3 runs: coldReaderDisplacement 365.0625px, 2 slips. So this is at least partly a real regression, not only CI timing. The single passes in CI suggest that timing decides whether a second row gets measured too late, for example font or Markdown readiness, or the virtualizer's measure-ahead margin.

The fix should remove the race. Raising the slip threshold would only hide it.

Environment

GitHub Actions test job on main (Linux, xvfb, Electron via Storybook static build).

Logs, screenshots, or additional context

The assertion and its rationale are at the end of the per-scene loop in scripts/perf/geometry-ablation.mjs.