test/apps-adversarial-teardown.test.ts runs under 10 s #3571

test/apps-adversarial-teardown.test.ts now runs in about 3 s here, down from 9.9 s on this machine (the 36.4 s in scripts/test-times.json was recorded on the vxrail). Both tests are still there and cover the same behaviours. I did not run it on the vxrail, so the under-10 s bar there isn't confirmed.

What changed

• Resurrection test: it no longer sleeps 3.5 s to see whether the preview comes back. It watches setTimeout for the 2 s restart backoff (RESTART_BACKOFF_MS[0]), lets the exit handler run once the stop has settled, and asserts no restart timer was armed. The directory, host and "restart" run assertions stay.
• Mid-deploy close test: it no longer sleeps. It waits for the in-flight deploy run to reach a final state, then asserts the directory is gone and the preview host answers 404.
• Waits: the directory-gone polling is now one gone() helper, and the polling sleeps are 10 ms instead of 50 ms.
• Timing record: scripts/test-times.json now records 3 s for this file. That figure is from this machine, not the vxrail.
• Production code: untouched. The file already builds one forge and one app runtime in beforeAll.

Checks

• Repeat runs: the file passed 10 runs in a row, each 3.1–3.9 s.
• Test would catch the bug: I temporarily broke the production fix (kept state.live set in teardown) and the resurrection test failed. I then restored the fix.
• Type check: bun x tsc --noEmit failed here with a read-only temp dir error, so I did not get a type-check result.

Not done: no test was deleted. The restart check now watches for the timer being armed rather than waiting out the 2 s window, so it would miss a restart scheduled with a different delay.

Closes #3561

test/apps-adversarial-teardown.test.ts now runs in about 3 s here, down from 9.9 s on this machine (the 36.4 s in scripts/test-times.json was recorded on the vxrail). Both tests are still there and cover the same behaviours. I did not run it on the vxrail, so the under-10 s bar there isn't confirmed.

What changed

• Resurrection test: it no longer sleeps 3.5 s to see whether the preview comes back. It watches setTimeout for the 2 s restart backoff (RESTART_BACKOFF_MS[0]), lets the exit handler run once the stop has settled, and asserts no restart timer was armed. The directory, host and "restart" run assertions stay.
• Mid-deploy close test: it no longer sleeps. It waits for the in-flight deploy run to reach a final state, then asserts the directory is gone and the preview host answers 404.
• Waits: the directory-gone polling is now one gone() helper, and the polling sleeps are 10 ms instead of 50 ms.
• Timing record: scripts/test-times.json now records 3 s for this file. That figure is from this machine, not the vxrail.
• Production code: untouched. The file already builds one forge and one app runtime in beforeAll.

Checks

• Repeat runs: the file passed 10 runs in a row, each 3.1–3.9 s.
• Test would catch the bug: I temporarily broke the production fix (kept state.live set in teardown) and the resurrection test failed. I then restored the fix.
• Type check: bun x tsc --noEmit failed here with a read-only temp dir error, so I did not get a type-check result.

Not done: no test was deleted. The restart check now watches for the timer being armed rather than waiting out the 2 s window, so it would miss a restart scheduled with a different delay.

I addressed the review in 1e59b72d and pushed it. test/apps-adversarial-teardown.test.ts now runs in about 3 s here, down from 9.9 s on this machine before the change. The file passed 8 runs in a row (2.7–3.4 s). I did not run the type check or the rest of the suite; the forge's check does that on push.

Changes since the first push

• Spy restore: the setTimeout spy is now restored in a finally, so a failed expect in the armed section can't leave it mocked.
• Mid-deploy test: it now calls a new quiet() helper, which waits until the pull's own deploy runs exist and every one is cancelled, failed or ok. It then asserts the directory is still absent and the host answers 404. It no longer uses settled() for this.
• Restart watch: it keys off RESTART_BACKOFF_MS[0], which is now exported from src/apps/runtime.ts. That export is the only production change; it adds no behaviour. The no-new-"restart"-run assertion is kept.
• Disagreement on the watch: I did not match all three backoff values. The first attempt did, and it failed on every run because unrelated 10 s and 30 s timers fire. A fresh environment's first crash always uses index 0, so one value is the right one to watch.
• Replacing the setImmediate loop: it is now one tick, not five, and I could not wait on the stop promise itself because it is internal to the runtime. Instead the test waits for the directory to go. teardown calls removeEnv only after await live.handle.stop(), so the directory being gone means the stop has settled. One loop turn then lets the exit handler run.
• Mutation check: I again broke the teardown fix (kept state.live set), and the resurrection test failed. I then restored it.
• scripts/test-times.json: it still records 3 s for this file, measured on this machine. The rest of that file's figures come from the vxrail, so the new one is not comparable until someone re-measures there.