testmgr flake: aarch64 c-conformance shard3 (00040.c) exit-124 under full load

• Type: bug (test infra — flaky timeout). Track T (testmgr tiers / timeout margins).
• Status: done
• Opened: 2026-07-14 (night session).

Evidence

Three occurrences on the same box (full tier, ~16-way parallel):

• b310's gate (2026-07-13, noted in 603cf2bd's commit message): shard3 timed out twice under load, 8.2s standalone, pinned pre-change compiler identical.
• 2026-07-14 night, full run for fec98091: 00040.c — exit code 124 (want 0), PASS standalone in 14.7s.
• 2026-07-14 night, full run for 30eb98f0: same signature, PASS standalone.

Always the same shard, same test, only under full parallel load, never a wrong RESULT — a qemu-user process starved past the per-test timeout.

Suggested fixes (T's call)

• Scale the per-test timeout for qemu conformance jobs with the load factor testmgr already computes (it scales scheduling; the per-test kill timer apparently not), or
• give cross-conformance shards a lower parallelism class, or
• one automatic retry for exit-124 within a conformance shard (a timeout is load-shaped, not result-shaped, so a retry cannot mask a real failure that reproduces).

Impact

Every full-tier run on a loaded box has ~⅓ chance of a spurious RED, which costs a manual standalone rerun to dismiss — tonight it did so twice.

Log

• 2026-07-15 — resolved, commit ab3a5b2a.