test/explore-rank.test.ts throws no unhandled error: every run of the file ends with each test named, pass or fail #5670

I couldn't reproduce the (unnamed) failure, so this change is a likely cause fixed on inference, not a confirmed one.

bun test --rerun-each 30 test/explore-rank.test.ts passed: 180 tests, 0 failures, no (unnamed) entry. After my change, --rerun-each 5 passed (30 tests), and the type check is clean.

Likely cause. beforeAll seeds 500 repositories in a loop of createRepo calls, each running git init. If the hook gives up first on a loaded check machine, afterAll closes the app and deletes the data directory while that loop is still running. The loop then writes into a closed database. The error has no test to belong to, so it shows up as (unnamed). I didn't see it happen here, so this is inferred from the code.

What changed (test/explore-rank.test.ts only).

• The seeding now runs as one promise, setup, which beforeAll returns.
• afterAll waits for setup to settle before it closes the app and removes the directory.
• It also skips app.close() if the app was never created.

Not touched. No product code changed, and no test was weakened. There are no retries or skipped cases, and the 300 s hook timeout is unchanged.

Open risk. If the (unnamed) failure comes back, the cause is something else. The next place to look is the bus.append listeners that run on repo.created.

To try it. Run bun test --rerun-each 30 test/explore-rank.test.ts (about 2 minutes).

Closes #5667

I couldn't reproduce the (unnamed) failure, so this change is a likely cause fixed on inference, not a confirmed one.

bun test --rerun-each 30 test/explore-rank.test.ts passed: 180 tests, 0 failures, no (unnamed) entry. After my change, --rerun-each 5 passed (30 tests), and the type check is clean.

Likely cause. beforeAll seeds 500 repositories in a loop of createRepo calls, each running git init. If the hook gives up first on a loaded check machine, afterAll closes the app and deletes the data directory while that loop is still running. The loop then writes into a closed database. The error has no test to belong to, so it shows up as (unnamed). I didn't see it happen here, so this is inferred from the code.

What changed (test/explore-rank.test.ts only).

• The seeding now runs as one promise, setup, which beforeAll returns.
• afterAll waits for setup to settle before it closes the app and removes the directory.
• It also skips app.close() if the app was never created.

Not touched. No product code changed, and no test was weakened. There are no retries or skipped cases, and the 300 s hook timeout is unchanged.

Open risk. If the (unnamed) failure comes back, the cause is something else. The next place to look is the bus.append listeners that run on repo.created.

To try it. Run bun test --rerun-each 30 test/explore-rank.test.ts (about 2 minutes).

test/explore-rank.test.ts now stops its seeding loop before teardown, so afterAll no longer waits out the remaining git inits. I never reproduced the (unnamed) failure, so this fix rests on the code, not on a failing run I watched go green.

What changed. afterAll sets a stopped flag and then waits for the setup promise. seed() checks the flag on every iteration, so the loop ends within one createRepo. The wait finishes well inside bun's default 5 s hook timeout, so app.close() and the rm do run. I took juniper's second option and left afterAll at the default timeout. The 300 s beforeAll timeout is unchanged, and no case was weakened.

Evidence.

• Original run: 30 reruns of the file passed, 180 tests, with no (unnamed) entry. That run was before any fix.
• Lowered-timeout run: I set the beforeAll timeout to 1.5 s once, locally, and did not commit it. The run reported a single hook-timeout failure and finished in 2.0 s, with no stray error after it. I did not run the same lowered timeout without the fix, so I haven't seen it fail the other way.
• Normal runs: the file passes (6 tests) and the type check is clean.

What would confirm it. The (unnamed) label should stop appearing on the forge's checks. A hook timeout on a loaded machine would now show as one named failure, not as an error with no owner.

Not touched. No product code and no new dependency.