Revert #137: it destabilized the daemon-gates (syncookie-v6 wedged on main) #152

Merged

vxfemboy merged 1 commit from sp-revert-137 into main

2026-07-07 15:28:32 +00:00

vxfemboy commented

2026-07-07 15:22:50 +00:00

(Migrated from github.com)

Reverts #150 (the #137 lab-teardown-residue fix). It regressed previously-green gates: on main, deception-syncookie-v6 wedged (~24min on a 6min step) — that gate had passed reliably for many runs before #137. In PR #151, deception-syncookie wedged similarly. Both are daemon-gates that were stable pre-#137.

Cause: the #137 change added an unbounded pre-gate sudo lab down before every gate plus a reap_orphan_lab_procs() pkill — one of these intermittently hangs / races the daemon-gates' engine, wedging a gate that used to pass.

Reverting restores the known-green state (original teardown, no pre-gate sweep, the 4 known-flaky gates quarantined). The #137 root-cause finding (orphans in anonymous namespaces) stands and is documented on the issue — but the fix needs a bounded, non-racy reimplementation (timeout-wrap the sweep; make the orphan pkill signature strictly scoped and incapable of matching a live gate's engine) before re-landing. #137 stays open.

Restores CI to green; no product code affected (lab-harness + CI only).

Reverts #150 (the #137 lab-teardown-residue fix). **It regressed previously-green gates:** on main, `deception-syncookie-v6` wedged (~24min on a 6min step) — that gate had passed reliably for many runs before #137. In PR #151, `deception-syncookie` wedged similarly. Both are daemon-gates that were stable pre-#137. **Cause:** the #137 change added an **unbounded pre-gate `sudo lab down`** before every gate plus a `reap_orphan_lab_procs()` `pkill` — one of these intermittently hangs / races the daemon-gates' engine, wedging a gate that used to pass. Reverting restores the known-green state (original teardown, no pre-gate sweep, the 4 known-flaky gates quarantined). The #137 root-cause finding (orphans in anonymous namespaces) stands and is documented on the issue — but the fix needs a **bounded, non-racy** reimplementation (timeout-wrap the sweep; make the orphan `pkill` signature strictly scoped and incapable of matching a live gate's engine) before re-landing. #137 stays open. Restores CI to green; no product code affected (lab-harness + CI only).