[thymikee]

Summary

Follow-up to #3901, applying the review's non-blocking notes. 2 files changed.

• AccessibilityNodeReads.child now reads through nullIfGone, a small generic seam that treats a throwing read as a gone node. The JVM suite can reach it, and a new AccessibilityReadBatchTest case fails if the catch is removed. A non-vanishing failure (IllegalStateException) still propagates.
• A doc comment says a null node is a programmer error that the catch would also hide.
• child.getChildCount and the bounds reads in appendNode need no guard. In the framework sources (API 35), getChildCount() returns mChildNodeIds.size() and getBoundsInScreen(Rect) copies mBoundsInScreen. Both are field reads on a sealed node with no interaction-client call. Only getChild goes through AccessibilityInteractionClient, and fix(android): read a child that vanished mid-read as gone #3901 already covers it.

Validation

• Head a09cc7d: pnpm check:affected --run --head a09cc7df06 passed (exactCommit). The snapshot helper build runs SnapshotHelperTestSuite, which includes the new case.
• With the catch changed to a different exception type, the same build fails with NullPointerException: window closed from the new case.
• GitHub owns the Android lanes. No device run: the change is a refactor of the already-verified catch.

🤖 Generated with Claude Code

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.54 MB	5.53 MB	-6.9 kB
Package (unpacked)	5.54 MB	5.53 MB	-6.9 kB
Package (download)	1.68 MB	1.68 MB	-1.5 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	31.4 ms	32.9 ms	+1.5 ms
CLI --help	64.8 ms	65.9 ms	+1.1 ms

[thymikee]

The code at a09cc7d is ready, and the new test correctly pins that a child vanishing mid-read reads as gone. I read the FutureTask and await path to confirm the test would fail without the fix, but I did not run the JVM suite or mutate the catch. I could not check the API 35 framework sources, so the note that getChildCount and getBoundsInScreen are plain field reads is not verified here. There are no conflicts. Smoke Tests is still running, and it covers this change through the Android snapshot helper build and SnapshotHelperTestSuite, which now includes the new case. The other 15 checks are green, so this PR can merge once Smoke Tests finishes green.

Not blocking, and you can take or leave these: readsAChildThatVanishedMidReadAsGone calls nullIfGone directly, so the JVM suite would stay green if child went back to a bare node.getChild(...); a package-private child overload that takes the framework read as a lambda would pin that wiring, or the emulator lanes can cover it. A null node also throws inside the lambda and reads as a gone child, which Objects.requireNonNull(node) in child would avoid and let you delete the doc paragraph at line 42. The nullIfGone javadoc could also name the result-integrity NPE it handles instead of the general "vanished node" wording.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-22 11:14 UTC