daily-fix: pod-side liveness probe before LAUNCH FAILED

Workflow gap

Two same-day #1739 fan-out patterns wrote false negatives into durable/driver records because the reporting layer keyed on the SSH channel instead of pod-side state:

1. Launch-wrapper false "LAUNCH FAILED". The 6-pod jobd/r2aug fan-out's launch ssh hung (~35 min for jobd-evil), was killed, returned non-zero — and the wrapper wrote "pod-1739-jobd-evil LAUNCH FAILED" while the detached pod-side run was alive (pid 1885 verified). Recurred for jobd-sycophancy at 07:48Z. The teammate shipped a session-local v2 driver with a 60 s launch-watchdog that probes the pod-side leg before reaping hung sshs — the recipe exists but lives only in that session. A sibling instance the same day: pod-1739-a1apilot "LAUNCH FAILED (ssh_rc=124, no live pod-side proc)" needed a manual re-pin + relaunch (that one was a REAL failure — the probe is what tells them apart).
2. Heartbeat false "UNREACHABLE — investigate". The R5 wall-clock heartbeat flagged terminated-after-verified-upload pods as "UNREACHABLE (ssh rc=255) — investigate before assuming healthy" twice (07:45Z, 10:01Z); the 10:02Z heartbeat corrected the copy to "TERMINATED (verified upload) … NOT a fault" by checking the termination sentinel first.

verified-at-filing: all quotes are probed tool_result/driver-log readbacks from session 21e049f7 (rows 4113–4260, 4222/4556/4569) and 2f4940f0 (rows 271–328). grep -n 'LAUNCH FAILED\|liveness' .claude/rules/pod-side-reporting.md | head run at compose time — the (re)launch contract covers pid-file/log/sentinel duties but has no "probe before declaring failure" clause for wrappers/heartbeats.

Proposed change

Add to .claude/rules/pod-side-reporting.md (the (re)launch contract section): a launch wrapper never writes LAUNCH FAILED without a pod-side pidfile/process probe (the 60 s launch-watchdog pattern: ssh channel outcome ≠ leg outcome), and a heartbeat classifies ssh rc=255 against the termination sentinel / pod.py list-ephemeral before writing UNREACHABLE. Name the v2-driver watchdog as the reference implementation so the next fan-out inherits it instead of re-deriving it.

Provenance

• fingerprint: 068a96affa5c

• workflow_fix_target: .claude/rules/pod-side-reporting.md

• origin: /daily 2026-08-05 problem sweep — miner 2 P2/P21.