[arun279]

Why

The iOS perf probe failed its own runs with "no responseMs, doneMs": in run 39878809507 (pass 1, the change app's show detail push), in 39867496265 (pass 2, show detail push), and twice on the first tab switch (39857698481, 39737186653).

I went through the raw videos and the fake Trakt log. In each case the app had acknowledged the command, but the screen did not change within the 2.5 s window.

• Show detail push, 39878809507:
  • the recording has a single frame, and simctl recordVideo writes a frame only when the screen changes;
  • show detail's requests (seasons, related, fanart) reached the fake Trakt server before the next segment's, so the screen had mounted;
  • the next segment's recording is still on Up Next for 1.5 s, then Up Next's header snaps into place, show detail slides in, and the episode sheet opens on top;
  • that late push landed in the episode sheet's window. It is that segment's 1210 ms response and 2276 ms finish, the maximums in that run.
• Tab switch, 39857698481: the tab bar moved to Library at once, but the content stayed on Up Next for the whole window.

The probe decided whether anything had happened by looking for a scene change anywhere on screen. A frozen screen has none, and the tab bar's change is below the threshold, so both runs failed as "the probe missed it" instead of being recorded as a late change.

What changes

• The probe takes a screenshot just before each command.
• If the content region has not changed by the end of the window, the probe polls screenshots until the content differs from that one, for up to 15 s, before it starts the next interaction.
  • Content that changed late is recorded as censored (>2500), as a late change already was.
  • The late change can no longer land in the next interaction's window.
  • Content that never changes still fails the run.
• The whole-screen scene pass is gone. Each segment is analysed before the next starts, because the decision to wait depends on the result; that costs about 0.3 s of ffmpeg per segment.
• A launch that never reports ready now leaves launch-timeout.png, so a stopped run shows what was on screen.

The screenshot digest covers only the content region's pixels, so the status bar clock does not count as a change.

Tests

• afterWindow censors content that changed only after the window, keeps it empty when it never changed, and leaves measured times alone without waiting.
• waitForChange resolves once the content differs, and gives up at the limit.

Four mutations each fail a test: always censoring, ignoring measured times, an inverted comparison, and reporting a change on timeout. The screenshots, recordings and the wiring in measure() run only against a device, so they are checked by probe runs rather than unit tests.

Verification

• Probe run 39886057572 (5 passes per app, both platforms): iOS completed, every segment measured and none censored. The stall did not recur in that run.
• Android stopped when the change app's fifth launch never reported ready. That app is Read the runtime's three stores in parallel at startup #201's code, and Read the runtime's three stores in parallel at startup #201's own Android lane failed the same way, on "Couldn't start Cue" after a relaunch. Details are on Read the runtime's three stores in parallel at startup #201.
• Rerun 39889357440, Android only, with the launch screenshot in place: every pass of both apps completed, no value was censored, and every launch reported ready. The launch failure did not recur, which fits the intermittent startup failure described on Read the runtime's three stores in parallel at startup #201.

[github-actions]

Pull request footprint

measurement	base to head
Product code lines in core/src, native/src, native/app, and native/modules	+0
Test lines in test, packages/*/{test,__tests__,e2e}, and .maestro	+40
CI and tooling lines in .github and scripts	+48
Product comment lines identified by a comment prefix	+0

User-delivered artifact sizes

measurement	base	head	delta
Expo iOS JavaScript bundle, raw file	5.45 MB	5.45 MB	0 B
Expo Android JavaScript bundle, raw file	5.48 MB	5.48 MB	0 B
Firebase tester APK, arm64-v8a and all densities	41.17 MB	41.17 MB	0 B
Play download estimate, XXXHDPI arm64-v8a English Android 15	19.83 MB	19.83 MB	0 B

Most complex product function

measurement	base	head
Cyclomatic complexity, oxlint complexity	10	10
Cognitive complexity, Biome noExcessiveCognitiveComplexity	15	15

[arun279]

Superseded by #212, which removes the perf probe from CI along with the rest of the custom device-lane machinery.