A booting forge removes only its own dead checkouts, not the checks another forge is running #1887

A booting forge no longer deletes the checkouts another forge is still using, so a roll can't pull the working directory out from under a running check (#1884). test/checks.test.ts passes (32 tests), tsc --noEmit is clean, and the new test fails without the change.

What changed

• startChecks used to run rmSync(dataDir/checks, recursive) on boot. It now calls a new sweepCheckouts in src/checks/service.ts.
• sweepCheckouts goes through checks/<runId>. It keeps a directory only if its run is running and leaseLapsed(run, STALE_AFTER_MS) is false.
• It removes everything else, including directories with no matching run.
• The per-run cleanup in execute is unchanged.

Test

• The new test is in test/checks.test.ts. It creates checks/<id> for a running run with a fresh heartbeat, for a running run with a lapsed heartbeat, and for an id with no run, then calls startChecks.
• Only the live run's directory survives.

Left out

• The optional change is not done: a check that fails because its own checkout vanished is still recorded as a failure, not requeued with REQUEUED_SUMMARY.
• Only checkout directories are swept. Nothing else was touched.

To try it, run bun test test/checks.test.ts.

🤖 Generated with Claude Code

Closes #1884

scout only (reviewers: not for you): the check on this pull is red on "edits on a mirror land upstream first > a commit made here appears on the origin"; fix it on this branch and push, or say in one line why the red is not yours.

scout only (reviewers: not for you): correction from the steward. The red on this pull is most likely load, not your change: the tests it failed (edits, overlap, snapshot-drill, two-process) pass on main alone (10/10, 4/4, 5/5), and they failed together across several pulls while the lane ran 9 checks × 6 shards. I asked for the check again; if it comes back red on the same test, it is yours.

The red isn't from this change. I made no code changes in response to the feedback, so nothing was pushed.

I re-ran test/edits.test.ts on this branch, which includes "a commit made here appears on the origin". All 10 tests passed, with 40 expectations. The change only touches how a booting forge removes dead checkouts. It doesn't touch the edits or mirror code. That fits the steward's account that the failures came from load while 9 checks × 6 shards were running. If the re-requested check fails on that same test, send me its log and I'll look again.