perf(ios): reduce capture probes and text entry waits

[thymikee]

Summary

Reduce iOS latency by overlapping fresh system-host scans, deferring irrelevant keyboard reads, and shortening unreadable-value warmup. Retain first-character target checks and final verification.

On the upstream test app, median warm captures improved from 91.5 to 75.8 ms and password replacements from 3,317 to 3,044 ms. Removing the split was rejected after it typed into a successor input.

The further two-second fill investigation rejected Select All clearing: a provisional 2.55-second sample lacked login verification, and a masked-input regression failed while unchanged clearing passed. No further runtime shortcut is retained.

Scope: six runtime/test files and four evidence files. Context: tester-army/e2e#1236. Measurements, rejected experiments, and reproduction.

Validation

Tested commit: 8a3931a9c4fd8188636b7f5a36d5e56b6b97b038; comparison base: 8f3eb3d2efde3654a560fddb39127d556e068352.

• pnpm check:affected --run --head 8a3931a9c4fd8188636b7f5a36d5e56b6b97b038: passed. Receipt .tmp/validation/run-wL2vPV/receipt.json: exactCommit=true, checkoutChanged=false; nine checks passed, two deduplicated, six skipped. 1,046 related tests passed.
• Runtime verification: 17 native regressions and 43 focused capture/routing tests passed; iOS/macOS builds passed. Follow-up: control 3 passed; rejected candidate 2 passed/1 failed.
• Independent review: no remaining findings.

GitHub owns coverage, native lanes, and macOS replay; manual iOS replay lanes are parked. Physical-device verification and the full upstream suite were not rerun.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.42 MB	5.41 MB	-4.9 kB
Package (unpacked)	5.42 MB	5.41 MB	-4.9 kB
Package (download)	1.64 MB	1.63 MB	-2.0 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	43.0 ms	42.9 ms	-0.1 ms
CLI --help	78.8 ms	81.4 ms	+2.6 ms

[thymikee]

This PR is ready for human review. I reviewed commit 6a47e61 and found nothing that needs a change. All 17 checks pass at the head commit, and there are no conflicts.

Not blocking: no native test covers a masked or secure input that is removed or replaced right after its first character, so a secure-field variant of testFillFailsWhenAppReplacesInputBeforeTextIsDelivered would close that gap (the unreadable early return in readBackWarmupCharacter is covered only through the pure predicate). The investigation report in docs/evidence/ios-e2e-engine-pr1236-2026-10-19.md could also shrink to the latency evidence. Take or leave both.

The simulator runs in the PR body were recorded against 8f3eb3d plus a worktree patch, not the head commit. I could not confirm the content is byte-identical, but the PR is a single commit on that base. I also could not inspect the local evidence zip or the mutation runs. From the diff, the new tests would fail without the fix. No physical-device run was done, as you note.

The branch moved to 8a3931a while I reviewed. That commit only adds evidence files, so this verdict still covers the code.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-19 13:08 UTC

[thymikee]

Followed up on the masked-input coverage suggestion in #3820, commit 739ccd617acfd3c4bf2d0306d4e27c260015b92c.

Added native cases for a masked field removed or replaced after its first character. Both require TEXT_INPUT_NOT_FOCUSED, no recorded XCTest failure, and no text in the replacement field. The five-test selection (new cases, existing readable variants, stable masked positive) passed on iPhone 17 Pro / iOS 27.0 with zero skips before rebase. Native sources are byte-identical afterward.

In a disposable checkout, replacing the warmup read's identity-bound resolution with textValueRead(for: activeTarget.element) changed the same selection to 2 failed / 3 passed. The new masked-removal test caught the stale-handle XCTest failure. The replacement cases continued to pass. The mutant was not shipped; the disposable checkout was removed and the simulator shut down. An initial run was killed before XCTest bootstrap; after a simulator reset, the control completed successfully. The initial launch failure's cause is unproven.

I retained the original investigation report: the requested work included understanding the engine protocol and safety differences, so those details remain useful evidence. The follow-up report provides the concise latency results separately. The follow-up also has uninstrumented native verification; historical measurements remain attributed to their original source revision.