Concurrent check installs race on the shared bun cache ("Fail extracting tarball"): good pulls look red and stall the merge queue #5093

Probe, 2026-10-22. Checks fail at install with error: Fail extracting tarball for "@typescript/typescript-linux-x64":

• pull-4781 and pull-4911 at 02:40.
• pull-4814 at 04:25, while it was the merge queue's head, so a good pull looked red and the queue paid another ~8 minute gate.

What it isn't: the per-repo bun cache (/data/crew/.cache/<repo>/bun) is intact; [email protected] has 114 files and 28 MB.

Likely cause: several checks (and gate runs) of one repo install concurrently against the same bun cache, and racing extractions of the same tarball collide.

Done when:

• No racing on the shared cache: installs into a shared cache are serialised per repo (a lock file around bun install), or each run installs into a copy-on-write or hard-link view, so concurrent runs never extract the same package into one directory at once.
• An install failure is not a test failure: one matching Fail extracting tarball, EINTEGRITY or ECONNRESET is retried once after 2 s. If it still fails, the check is failed: infrastructure; it is re-run, never counted as red for the pull or the queue.
• Test: 8 parallel installs on one cache, 20 rounds, zero failures.

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