test(subagent): a git failure during worktree removal is not a test failure

[Lstarsky0]

What and why

failed_spawn_removes_its_worktree_despite_post_checkout_hook_output failed once on ubuntu-latest, in the CI run for #6952 (which doesn't touch subagent code): https://github.com/codewhale-hq/Codewhale/actions/runs/38057245676/job/114383431465

git ["branch", "--list", "codex/agent-hook-probe"] failed: fatal: failed to read .git/worktrees/codex-agent-hook-probe-path/commondir: No such file or directory

The test waits for the refused spawn's cleanup by polling git worktree list and git branch --list every 50 ms while the cleanup runs git worktree remove on the blocking pool, and each poll asserts that git exited 0. git checks that a linked worktree's commondir exists and then reads it (get_common_dir_noenv in setup.c). If the removal deletes the file between those two steps, the polling command dies with the message above. The cleanup itself goes on and finishes, so the product did the right thing and the test still failed.

The loop now reads a failed git command as "not settled yet" and polls again until the 30 s deadline. The assertions after the loop are unchanged and still use the strict helper. failed_spawn_removes_the_worktree_it_created shares the loop, so it gets the same fix.

Not in this change: remove_worktree_if_expired in the lane crate also starts with git worktree list --porcelain, and returns without removing anything when that command fails. Two cleanups overlapping in one repository could hit the same window and leave the second worktree behind. I haven't seen that happen, so I left it alone.

Issue

No-Issue: flaky test seen on CI, no issue filed for it

How I tested it

• The two refused-spawn tests on their own (cargo nextest run -p codewhale-tui --lib with a filter): 2 passed.
• To exercise the failure without waiting for the race, I put a wrapper named git first on PATH that makes the first two git branch --list calls exit 128 with git's message and passes everything else through. On main both tests fail at the first poll with the same panic as the CI run. With this change both pass.
• I could not hit the real race on my Mac: 400 git worktree add / remove cycles against four concurrent git branch --list loops, about 7300 polls, no failure. A commondir that exists for lstat but not for open (a dangling symlink) gives the exact message from the CI log.
• cargo fmt --all -- --check, cargo clippy -p codewhale-tui --all-targets --all-features with CI's flags, and cargo check -p codewhale-tui --all-targets: clean. Clippy also on 1.99, plus cargo +1.89.0 check -p codewhale-tui.
• cargo nextest run -p codewhale-tui --lib: 15248 passed, 44 failed. The same 44 fail on main (0ea319a) on my machine; most need an extension-host Node version I don't have.

No new tests; this only changes how two existing ones wait.

Checklist

• One focused change, rebased on current main
• cargo fmt --all passes, and tests cover new or changed behavior
• If this adds or changes a user-facing feature, I updated its docs and its row in docs/features.toml (cargo test -p codewhale-tui --test feature_registry checks it) — n/a

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-11T11:29:57.020465Z	e929899	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Hmbown]

Thank you @Lstarsky0 for narrowing the worktree-removal observation race: a transient Git read failure while commondir is being removed should wait for settlement, while the final worktree and branch checks must stay strict. Your fix has merged as its own PR in bbae5043, with your original authored commit and branch retained.

Exact own-head hosted results: macOS19,035passed/0failed/33skipped, Ubuntu19,032/0/34 and Windows18,301passed(oneleaky)/0/27; owner recovery passed all three. Both cleanup cases passed macOS/Ubuntu. The hook-output case is Unix-only; the portable cleanup case also passed Windows. The existing browser-launcher leak remains recorded. This carries the fix on main for the next0.10.2 integration; 0.10.2 has not shipped, and final assembled/package acceptance remains separate.