[earakely-scale]

What changes

Checks that failed on main and on unrelated branches over the last day and a half of local-backends runs, and on CodeBuild, each on a busy runner. Each either gets room it needed, or loses the part only timing decided:

Check	Failed with	Change
The gate-3 probe's recording check (conformance/checks.py, _segments_match)	"8 frames over 1.40 s of a 2 s recording at 5 fps" (main e5c1ab9), "9 frames over 1.59 s"	The segments must span half the recording, as the frames must already number half. A recorder that takes a moment to start no longer fails gate 3; one that stalls after its first frame still does (test_a_stream_that_stalls_after_its_first_frame_fails). A probe check that fails fails the compose, and so every test that needed that computer: computer_app_test's WebSocket test failed this way on main.
page_driver_test test_a_call_is_bounded_from_when_it_asks_for_its_session_and_leaves_no_session_half_open	a gateway session owner (ChildSessions._own) still running 0.1 s after the bounded call (main e5c1ab9 and 3 other branches)	Waits up to 5 s for the owners to end (OWNERS_END_WITHIN_S, polling every OWNERS_POLL_S), so one that never ends still fails.
test_computer_app_server test_the_drivers_readiness_asks_every_apps_server_at_once	3.15 s with 2.5 s allowed, on CodeBuild (#666)	Times the readiness request alone, not the fake driver's setup, and each app's check takes 2 s: asked at once the status takes about 2 s, one after another 6 s, and the bound is 4 s. The driver does ask at once (runtime.ready_checks gathers them).
desktop_example_test test_a_scripted_drive_is_recorded_on_the_clock_its_screenshots_carry	the clicked cell's first changed frame 0.137 s after the screenshot that showed it, 0.1 s allowed (main 8668b2b); and on #666, no recorded frame showing the change at all (StopIteration)	The click-repaint timing is gone: it failed on 5 commits, both ways down to when the recorder writes a frame under load. The test still checks each screenshot against the recording's nearest frame, in time and in pixels, which is what puts the two on one clock.
desktop_example_test test_the_desktop_examples_computer_passes_the_probe_at_full	"scored below 1 (probe: 0)" on #666, with no word of which check	Its failure now names each probe check that failed and why, from the run's verification.

Testing

• page_driver_test.py, test_computer_conformance_recording.py and the readiness test pass locally.
• desktop_example_test.py runs in integration-local-slow on this PR.

Not in this PR

• launcher_test.py ×3 (KeyError: 'image'): two composes of one disk racing; folded into fix(computers): a disk form's console shows the session's programs again, so releases publish #666.
• Gate 3's P7 (a drag through its waypoints, the reference driver's Tk desk) and P6 (a chord selects all, cua-driver) failing now and then: fix(computers): the probe tries a drag or a chord once more when a busy desk may have missed it #671.
• run_bundle_test's killed-run test: ECR Public rate-limiting a pull of nginx:stable-alpine.

🤖 Generated with Claude Code

Confidence Score: 5/5

This PR appears safe to merge.

Summary

This PR gives busy runners more room without removing the remaining recording and cleanup checks.

• The recording probe allows startup delay without skipping frame checks.
• Failed desktop probes show which checks did not pass.
• Desktop recording checks no longer time click repaint against video frames.
• The readiness test distinguishes parallel checks from slow setup.
• The page-driver test gives session owners more time to finish.

The previous unnumbered polling-delay finding is fixed.

_{Reviews (6) · Last reviewed commit: "test: the desktop example's scripted dri..." · Reviewed by Greptile}