the #387 bounded-diff test exceeds its 5s budget on windows-latest, then EBUSY on cleanup

[SUaDtL]

The failure

farm.test.ts > farm artifact publication (#397 / #387) > #387: the unavailable-diff list is bounded and still reports the true total fails on windows-latest with two errors:

Error: Test timed out in 5000ms.
 ❯ farm.test.ts:1730:3

Error: EBUSY: resource busy or locked, rmdir
'C:\Users\RUNNER~1\AppData\Local\Temp\farm-artifacts-…'
 ❯ farm.test.ts:1488:5   (afterEach)

Observed at 5331ms against a 5000ms budget — it is over by ~7%, not by an order of magnitude. Passes on ubuntu-latest; passes locally on Windows.

Two distinct problems, in order

1. The budget. This case plans 12 tasks at FARM_CONCURRENCY: 6, each spawning a gate process. A GitHub Windows runner is materially slower at process creation than a developer box, and it is the heaviest case in the file.
2. The cleanup. The afterEach rmSync then hits EBUSY, because the timeout abandoned the run mid-flight and something still holds the temp directory. This is the more interesting half: a timed-out run leaks a live handle, and the teardown reports it as a second failure that obscures the first.

Why it is NOT #515

#515 was a genuine race — concurrent git worktree add reading a sibling's half-written commondir, fixed under a worktree lock in #534, with the failure signature worktree add failed: … failed to read .git/worktrees/<other>/commondir.

That signature does not appear here. This is a wall-clock budget on a slow host, and the escalation #515 described is absent. Same test, different cause — worth stating plainly, because the obvious reading is that #515 regressed and it did not.

Note the lock added in #534 serializes the (short) registry calls; on a host this slow that is a plausible contributor to the margin and should be measured rather than assumed.

How it surfaced

plugins/ca/tools had never run on Windows in CI — the tools job is ubuntu-latest only. #521's coverage-union job added a windows-latest cell, and #539 cleared the containment failure that was masking everything behind it.

Acceptance criteria

• AC-1: the margin is measured on windows-latest rather than guessed — how long this case actually takes there, and how much of it is process creation versus the serialized registry calls.
• AC-2: the budget fits the slowest supported host with headroom, or the case is made cheaper without weakening what it asserts. It must still prove the unavailable-diff list is bounded at 11 while reporting a true total of 12, and FARM_CONCURRENCY must stay at 6 — flaky: a task escalates under full-suite load, flipping exit 0 to 2 in the farm artifact-publication tests #515 AC-4 applies unchanged.
• AC-3: a timed-out run releases its handles, so afterEach cannot fail with EBUSY and bury the real error. A teardown failure must never be the loudest thing in the log.
• AC-4: the test passes across repeated windows-latest runs.

Not in scope

Found while closing #539. Filed separately from #515 (fixed, different cause) and #541 (a worktree-fs refusal in the same CI run).

[SUaDtL]

Reopening: my first fix was incomplete, and #545's CI caught it.

I sized the budget from the one case this issue named and left the rest of the block on the 5000 ms default. Seven cases then timed out together on windows-latest, turning the ca coverage-union cell red again.

The cause was never that case's concurrency — every case in farm artifact publication launches real farm subprocesses, and process creation on a hosted Windows runner is far slower than on a developer box. The heaviest case simply crossed the line first, so sizing from it was always going to find the next-slowest case next.

Fixed properly in #546: the budget moves to the describe block, and the redundant per-case timeout is removed. AC-3 (reaping abandoned children so the teardown cannot report EBUSY over the real error) shipped in #543 and is unaffected.

[SUaDtL]

Completed in #546. The budget now belongs to the farm artifact publication block rather than to one case in it, and both ca coverage-union cells passed on that PR — including windows-latest.

AC-1 measured (3553 ms isolated / 5331 ms under CI load for the heaviest case), AC-2 satisfied without weakening the case (FARM_CONCURRENCY stays 6, assertions unchanged), AC-3 shipped in #543 (abandoned children are reaped before teardown so EBUSY can no longer report over the real error), AC-4 green.