A gate's checkout is not removed under it: three gates on #7608 went red on files missing from their own working tree #7634

Gates on #7608 (head squashed onto main@accaa65) went red three times in a row in setup, before or just as tests started, each time with a file missing from the gate's own checkout:

• gate 1793498673681-628892395cba: test/crew-early.test.ts failed with ENOENT reading "src/core/migrations/20261027090000-home-visit.ts", a file that is on main and in the pull.
• gate 1793499189025-df28692f87fb: mkdir -p .forge-cache && tsc ... --tsBuildInfoFile .forge-cache/tsc.tsbuildinfo failed with TS5033: mkdir .forge-cache: no such file or directory, a directory the step before had just made.
• gate 1793499279023-9a71c1629451: the first bun install succeeded, then bun install --backend=copyfile said could not find a package.json file.

Each push to the pull started a new gate within minutes of the previous one. The likeliest cause is a later gate, or a cleanup, removing or reusing a checkout directory that a running gate still uses.

Done when:

• [ ] Each gate runs in a checkout directory no other gate or cleanup removes or reuses while it runs.
• [ ] A test starts two gates for the same pull close together and proves the first still finds its package.json and finishes with its own verdict, or is cancelled with a verdict that names the newer gate. It is not left red on a missing file.
• [ ] A gate that loses its checkout is reported as an infrastructure failure, not as a red against the pull.

A fourth gate on #7608, 1793499399099-82a21fdd2bca (02:16:38Z), went red 0.6 s in on the same cause. test/crew-early.test.ts hit ENOENT reading "src/core/migrations/0069-an-event-s-place-in.ts" and test/model-grant.test.ts hit ENOENT reading "src/crew/flow_label.ts". Both files are on main@accaa65. Peak memory was 18.5 GiB of 96, so this is not memory pressure.

A fifth gate, 1793499497216-1e898d0466b7, names the cause outright. In test/model-grant.test.ts, git failed with fatal: Unable to read current working directory: No such file or directory: the gate's working directory was deleted while it ran. The rehearsal ("old forge did not come up") and release-handover ("pre-receive hook declined") reds in the same gate fit that too.

Correction to my last comment: the model-grant Unable to read current working directory red, and the release-handover push declined in the same gate, are fail-fast teardown. They follow the rehearsal red that #7632 describes, so they are not evidence for this issue. This issue rests on the setup failures that came before any test: the missing .forge-cache right after mkdir -p, the missing package.json between two bun installs, and ENOENT on files that are on main in the first shard to run.

Another setup loss on #7608: gate 1793499943106-09ad8e44e8d1 ran bun install twice and then failed on its first step with error: Module not found "scripts/gate.ts", a file on main@accaa65. No test had run, so this is not fail-fast teardown.

Live triage (automatic, not written by a person): no pillar named, so it waits behind the Later gate; name a pillar to move it.