fix(ios-e2e): avoid alert probe after stalled WebView read

[thymikee]

Summary

Give the WebView route's first native landmark up to 45 seconds before probing for a deep-link system alert. The linked failed run restarted XCTest during bridge fallback; its 15-second landmark wait expired, and alert get queried an already visible WKWebView, leaving the runner busy. A healthy WebView route still finishes this wait as soon as the landmark appears.

The shared 15-second default and failed-wait → alert get/accept path remain in place for real “Open in” confirmations. Two files changed.

Validation

Commit e358595a5905807d97ec3607f1eceb94edbd5f26: pnpm format and pnpm check:affected --run passed (format, lint, typecheck, Fallow; Vitest found no related tests). The failed WebView run and a successful run with a real confirmation informed the budget and retained alert path. The iOS simulator replay on this PR head is pending; it is the live behavior check. A restart longer than 45 seconds can still reach the alert probe.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.80 MB	4.80 MB	0 B
Package (unpacked)	4.80 MB	4.80 MB	0 B
Package (download)	1.44 MB	1.44 MB	-3 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.2 ms	27.6 ms	+0.4 ms
CLI --help	79.3 ms	80.0 ms	+0.7 ms

[thymikee]

Reviewed at e358595. The code looks good: the longer 45 s landmark wait applies only to the WebView route, and the shared 15 s default and the alert path stay the same.

Not blocking: no run has yet shown the XCTest restart path finishing inside the new 45 s budget, so the fix is supported by the failure pattern rather than by a direct run. Checks were still queued at review time.

[thymikee]

Recommend closing this in favor of #2901 and #2902, both review-clean:

• fix(ios): find a popover dismiss region in one query instead of reading every element #2901 fixes the root cause. alert get with no alert walked every element of every window one XCTest read at a time, and on a loading WebView that overran the watchdog and kept the runner busy for about 38 s. It is now a single predicate query: 0.2 s instead of 14.7 s.
• test(ios): wait out the launch an accepted deep-link confirmation releases #2902 moves the prompt helper to live-deep-link-confirmation.ts and only re-checks the prompt when a wait fails with APP_NOT_RUNNING. A stalled WebView read no longer leads to alert get at all. This PR's docblock ("alert get against a live WKWebView … exceeds the watchdog") is out of date once fix(ios): find a popover dismiss region in one query instead of reading every element #2901 lands.

One gap: this PR allowed up to 45 s for the landmark before the 20 s page wait, while #2902 allows 15 s plus the caller's 20 s. A runner restart longer than about 35 s would pass here and fail there. No such restart has been observed. Both PRs edit the same helper, so they will conflict.

[thymikee]

The exact-head iOS smoke job for this PR passed, but I’m keeping it open as a draft for now. I agree #2901 addresses the expensive no-alert query. I found a real prompt path missing from #2902’s APP_NOT_RUNNING classifier and documented it here: a previous passing run had readable XCTest captures until the destination wait expired, then alert get/accept handled “Open in…”. The 35 s combined wait budget also falls short of the linked failed run’s 51.9 s runner recovery. Once the replacement covers both shapes, this overlapping PR can close.

[thymikee]

#2901 and #2902 have now merged into main. The replacement keeps the native landmark wait at 15 seconds per attempt, but allows up to five bounded attempts for stalled/restarting captures; it also probes readable timeouts for a real “Open in…” confirmation. That addresses the two paths raised above without this PR's 45-second WebView-only override.

#2897's exact-head iOS smoke passed, but it did not reproduce the 51.9-second restart. The post-merge iOS smoke for #2902 is still queued (job 107700023853). I am keeping this PR in draft while that replacement run completes, then expect to close it as superseded if the run passes. No rebase is needed for a superseded patch.

[thymikee]

The replacement now has the live evidence I was waiting for. #2901's exact-head iOS smoke passed, and #2902's exact-head iOS job completed both the targeted XCTest step and the fixture-backed simulator E2E smoke step successfully. The fixture smoke includes smoke:webview-remote-content, the route involved in the original failure; the job is finishing a separate gesture replay at the time of this comment.

#2901 removes the expensive absent-alert query, and #2902 retries typed readable/stalled waits with a bounded budget while retaining the real confirmation path. This supersedes this PR's single 45-second landmark wait. I am closing this draft instead of rebasing an overlapping fallback.