carlos / platform Public

Flaky CI test: stopevidence TestFDIdentityFindsAHardLinkHeldByAChild open

claude agent, operated by paul@keymail.dev operated by paul@keymail.dev session 5b793570 · claude-code/2.1.288 · DOCS-T1

Flaky in CI only: TestFDIdentityFindsAHardLinkHeldByAChild in internal/stopevidence failed once in CI on zdd-6-surfaces (d12a71c1, a branch that does not touch the package).

Locally it passed 30/30 on that branch and 50/50 on origin/main 30866251, and a fresh CI run passed. It looks like a timing race in the test under CI load. Filed so it can be fixed rather than re-pushed past.

claude agent, operated by paul@keymail.dev operated by paul@keymail.dev session c4a04fdc · claude-code/2.1.292 · DOCS-T1

Fixed on branch test-suite-debt (8fef823e), CI green, not merged.

Cause. holdOpen started sleep 60 and scanned /proc as soon as Start returned, while the child was still loading. A probe saw four different files in turn on fd 4 (ld.so.cache, libc.so.6, locale-archive, locale.alias). FDHolders stats through the descriptor and, on ENOENT, lstats the entry; a descriptor closed and reopened between the two gives ErrIncomplete. A scratch test with a child churning descriptors got ErrIncomplete in 19 of 2000 scans. The original CI message was not recorded, so this is the mechanism, not a replay of that run.

Fix. The holder is sh -c 'echo ready && read wait' and the test waits for ready before scanning. No sleep, no retry. The mount-namespace test had the same race and uses the same handshake.

Left alone on purpose: FDHolders still returns ErrIncomplete for a live process that reopens a descriptor number mid-scan. That is the fail-closed rule from 9cf9f18f.