test(picker): de-flake both Windows switch-picker PTY tests

[worktrunk-bot]

What

De-flakes both Windows switch_picker.rs PTY tests that intermittently blocked #3241's test (windows) job. Attempt 1 of that run failed on test_switch_picker_alt_x_keeps_cursor_sticky; attempt 2 failed on test_switch_picker_worktree_row_comments_tab_shows_thread. Both are addressed here.

Flake 1 — alt_x_keeps_cursor_sticky: racy process exit code

The test drives the interactive skim picker in a PTY, removes a row with alt-x, presses Enter, and asserted the wt process exited 0. On Windows it intermittently failed with exit code 1 even though the switch verifiably succeeded — the captured screen shows both the structured result and the success message:

{"action":"existing","branch":"wt-keep","path":"D:\\tmp\\.tmpgrsMhQ\\repo.wt-keep"}
○ Switched to worktree for wt-keep @ 'D:/tmp/.tmpgrsMhQ/repo.wt-keep'

The cursor-stickiness belief is fully verified by the screen assertions: the --format=json result reaches stdout only after the switch pipeline ran, and it names the sticky landing row. The exit_code == 0 pin coupled that belief to the interactive session's process exit code, a separate racy signal on Windows (this file's assert_valid_abort_exit_code already treats skim PTY exit codes as non-deterministic: 0/1/130). The fix keeps the screen-based success verification and accepts skim's selection/abort codes (0 or 1) instead of pinning to 0.

Flake 2 — comments_tab_shows_thread: cursor reset by async row decoration

The test moved the cursor onto feature with a single blind Down, then opened the HEAD± tab and waited for feature's diff --git. On Windows it timed out after 30s. The captured failing screen shows the > pointer stranded on main, with the preview reading "○ main has no uncommitted changes" — so the awaited diff --git never appeared:

> @ main       ? ^|                                      │○ main has no uncommitted changes
  + feature   !  ↑       +1        ↑1        +1   -1      │

Root cause: feature's row decorates asynchronously — the primed CI status resolves to "has PR" in the background. When that resolution lands it refreshes skim's item list, which resets the cursor to the top, undoing the Down. The preview re-issue nudge from #3238 then re-ran Alt-1 against main forever.

Fix: make Down cursor-confirmed. When a Up/Down arrow step carries expected content, it now names the target row, and the wait re-issues the (idempotent) arrow until the > pointer lands on that row and the screen settles — by which point the list has stopped refreshing. The step becomes ("\x1b[B", Some("feature")). This mirrors the alt-digit preview re-issue and reuses wait_for_cursor_on_row's pointer predicate (factored out as cursor_points_at). Arrow steps that pass None (every other test) are unaffected.

Verification

Both flakes are Windows-specific PTY races I could not reproduce on the Linux CI runner, so the fixes mirror this file's established de-flake patterns rather than root-causing platform exit-code / scheduling behavior. The full switch_picker suite (35 tests) passes locally on Linux, including both changed tests and every consumer of the shared helpers (wait_for_cursor_on_row, send_input_awaiting_content). The Windows CI on this PR exercises both on the affected platform.

Refs #3241

[max-sixty]

fix all the flakes please! (sorry for having introduced them)

[worktrunk-bot]

Done — pushed a fix for the second flake so this PR now covers both Windows picker PTY tests that blocked #3241.

I pulled the failing run's logs to confirm the root causes from the captured screens rather than guess:

• alt_x_keeps_cursor_sticky (attempt 1) — exited 1 despite a verifiably successful switch; the exit_code == 0 pin coupled the cursor-stickiness belief to skim's racy process exit code. Now accepts 0 or 1, with the screen assertions still proving success.
• comments_tab_shows_thread (attempt 2) — the > pointer was stranded on main (preview read "○ main has no uncommitted changes"), so feature's diff --git never appeared. The single blind Down was undone when the row's async CI/PR decoration landed and reset skim's cursor to the top. The Down is now cursor-confirmed: it re-issues until the pointer lands on feature and the list settles.

Those were the only two failures across the run's attempts. The full switch_picker suite passes locally on Linux; the Windows leg here will exercise both on the affected platform. Monitoring CI now.

[worktrunk-bot]

✅ CI is green on 7fcb90b7 — both Windows picker PTY tests (alt_x_keeps_cursor_sticky and comments_tab_shows_thread) passed, along with the full test (windows/linux/macos) matrix, codecov/patch, and pre-commit.ci. The flakes that blocked #3241 are both addressed.

[worktrunk-bot]

This flake just recurred on main — the collect affected coverage (windows) job of run 28186340000 (on 438d33e, the merge commit of #3241) failed on test_switch_picker_worktree_row_comments_tab_shows_thread with the identical signature this PR diagnoses: timed out after 30s waiting for diff --git, with the > pointer stranded on main and the preview reading "○ main has no uncommitted changes" while feature's row decorated asynchronously.

>  @ main       ? ^|                                      │○ main has no uncommitted changes
   + feature   !  ↑       +1        ↑1        +1   -1      │

So this is reproducing on the default branch too, not just on #3241's PR CI — adding a data point for the cursor-reset root cause this PR fixes. No separate ci-fix PR; #3242 already covers it.

[max-sixty]

@worktrunk-bot resolve conflicts

[worktrunk-bot]

Resolved the conflict and pushed the merge (6d663db). The branch is now MERGEABLE.

The only conflict was in tests/integration_tests/switch_picker.rs, in send_input_awaiting_content. main's #3247 added PreviewNotifier (auto-repaints a late preview when its background compute lands) and removed the test-side Alt-<digit> keystroke re-issue that #3238 used to un-strand "Loading…" placeholders. The two de-flakes here are orthogonal to that, so I kept both:

• Flake Add CI workflows and fix cross-platform test failures #1 (alt-x sticky exit code) — a separate region; auto-merged cleanly, no decision needed.
• Flake Test CI fixes and improvements #2 (comments-tab cursor reset) — kept. PreviewNotifier re-runs only the preview for the selected row (notify_row_changed); it doesn't stop on_update's request_render from resetting skim's cursor to the top when an async row decoration lands. on_update is byte-identical between the merge base and main on that path, so the cursor reset the comments-tab test hits still happens — the cursor-confirmed Down is still needed.

So the resolution keeps the cursor-arrow re-issue (restored wait_for_stable_until's nudge param, now scoped to cursor recovery and renamed PREVIEW_REISSUE_INTERVAL → CURSOR_REISSUE_INTERVAL) and drops the now-obsolete Alt-<digit> re-issue branch (is_alt_digit_tab deleted), deferring to main's PreviewNotifier for late preview content.

The full switch_picker suite (29 tests under shell-integration-tests) passes locally on Linux, including both changed tests and every consumer of the shared helpers. Monitoring CI on the merge commit now — the Windows leg is where both flakes live.

[worktrunk-bot]

✅ CI is green on 6d663db43 (the merge commit) — test (windows) passed (11m56s) with both picker PTY flakes, alongside test (linux) / test (macos), codecov/patch, lint, and the advisory matrix. Branch is MERGEABLE; the conflict is resolved.