Intermittent ready-age cold poll replay refusal after native restart (MariaDB restarts lane) #1096

Observation

In Shared fixtures run 42431342149, job 128581664078 mysql-execution (mariadb-10.11, restarts) failed on PR #1095 head d9c4d5d2. #1095 changes only CI install scripts and tests; no runtime or product file.

scripts/conformance/server-parity/ready-age-restart.php failed in its finish phase (line 80 → readyAgeClaim, line 28):

PHP Fatal error:  Uncaught RuntimeException: Native successful claim lacks its declared literal ready-age value.

The finish phase replays the original workflow- and activity-poll requests (same poll_request_id) against the recovered native node, after the original node was killed with SIGKILL. The check requires server_capabilities.claimed_task_ready_age === true and an integer task.ready_age_ms in the replayed claim.

Timeline from the job log:

• The range profiles finished.
• Restart, cooperative and ready-age prepare ran against native-a, then docker kill at 00:55:33Z.
• The recovered node became ready at about 00:56:05Z.
• finish ran at about 00:56:06Z.

That is roughly 30–40 s after prepare, inside the 60 s workflow lease plus the 5 s answer retention, so answer expiry is not the explanation on this evidence.

The same native source and lane passed on #1095's previous head (42429372674) and on recent main runs. The cause is unknown.

Evidence gap

The finish script throws without printing the replayed response. Only ready-age-restart-prepare-controls.txt is uploaded (artifact 13598420191, {"outcome":"prepared","claims":["workflow","activity"]}). The receipt JSON with prepared_at and the saved claims is not uploaded. So this run cannot distinguish:

• an empty or non-leased poll answer;
• a missing capability;
• a missing ready_age_ms field.

Next action

1. Diagnostics (Server). On refusal, ready-age-restart.php should print the bounded HTTP status, poll_status, the presence of server_capabilities.claimed_task_ready_age, and the task key set (no payloads), and keep the receipt JSON in the evidence upload. This is a CI-only change and does not weaken the assertion.
2. Classification. Reproduce or attribute it once those diagnostics exist; until then treat it as an intermittent product-check result, not a pass. The decision on Install profile SDK fixtures from a verified job-scoped dist cache #1095's current failed job rests with General.

Owner: Server.