[backnotprop]

Cause

Since the split, the decisions-guides shard runs M3 first, straight after the warm-up, on its own runner. That runner is slower than the old single job: M3 took 386 to 524 s there, against 338 s on 046a3a5. Main's code passes locally 3 of 3 times. The failures are races that only a slow runner hits, plus one real ordering dependency. Each failure then spread to the next class, and in three runs to a hung job.

1. Back tapped while the decision sheet was still closing. Run 38010716662 (PR 1819):

   t = 226.06s Tap Button (Element at index 0)
   t = 226.18s     Check for interrupting elements affecting "decision-cancel" Button
   Failed to tap "Decisions" Button: No matches found for Descendants matching type TabBar
   

2. The New message sheet was read as closed 3 s after Send. Run 37973911538 (main, b0f60ec): DecisionsProofTests.swift:132: XCTAssertFalse failed - the sheet closes once delivered. The text field still existed at t=275, 277 and 279 s after the Send tap at t=272 s.
3. M4 opens the shard's first web view. Run 38017849289 attempt 2: the guide cover showed only its spinner for 30 s (the guide's title not found, with the screen printed). Before the split, AttachmentProofTests paid WebKit's start-up cost on the simulator first.
4. A failed class kept its screen recording running. The next class's simctl io recordVideo exits at once with Host recording is already in progress (exit 16, reproduced locally). /video/stop then waited for an exit that had already happened, so the next class failed on NSURLErrorDomain Code=-1001 "The request timed out.". This happened to M4 in 37973911538 and 38010716662, and to RelayPushTests in pair-and-push in 37958261402 and 37987495244.
5. The job hung until the 45 min timeout after a ProofCase failure. This happened on d5d5f54 (37958261402), 38017849289 attempt 2 and 38023662529. An async XCTest runs on past a failure. ProofCase's failure tearDown terminated and reinstalled the app while the test body was still tapping. The log shows Tear Down, then Terminate, then the body's Tap Button (Element at index 0), then a second Tear Down, then nothing for 30 min.
6. Also on a slow runner: Reply instead was read as not hittable before the swipe had settled (38023662529), and the 8.1 menu did not close on the first tap outside (38027139806 attempt 2).

Fix (proof code only; every class still runs)

• DecisionsProofTests
  • Is now a ProofCase: cold install at setUp, shared helpers, duplicates removed.
  • Waits for each sheet to close (after Cancel, after Done and before Back, after both Sends).
  • Waits for Reply instead to be hittable.
  • Closes the 8.1 menu and the 8.3 words by tapping outside, up to three times.
• GuideProofTests: the first wait for the surface uses ProofWait.opening, like AttachmentProofTests' surface waits. This is the one bound that changed. It is where WebKit's first start is paid in this shard.
• Waits now throw (require): the pairing waits, the guide's waits and M3's sheet waits end the test body when they fail. The pairing waits print the screen when they fail. Two older pairing failures had nothing to read: 38013707345 (the cover did not open) and d5d5f54 (no list after the code).
• ProofCase no longer resets in a tearDown. The next class's setUp resets. RelayPushTests, the one class after a ProofCase that did not reset itself, now resets at setUp.
• proof.ts: stopVideo() returns when the recorder has already exited, and /video/start stops one that is still running.

Runs

• Local, iOS 26.5, proof.ts --prebuilt --only DecisionsProofTests,GuideProofTests:
  • main's code: 3 of 3 pass.
  • fixed code: 5 of 5 pass across the heads, one of them on 3c0a4d8.
  • pair-and-push passes once on 1f2e9a6.
• CI, decisions-guides on each head:
  • f56fd30: green, then red (Guide's first web view, the tearDown hang).
  • 03de7dc: red (Reply instead, the tearDown hang).
  • 1f2e9a6: green, then red (the 8.1 menu; no hang, and M4 passed after it).
  • 3c0a4d8: green (run 38031955782).
• Other shards on 3c0a4d8: relay-transport and attachments green. pair-and-push red on a different, older fault: ProofTests could not find "Try Again", the same failure as dependabot run 37987495244 before this PR. RelayPushTests passed after it, so the reset at setUp works. Not addressed here.