Seen 2026-08-25 on cranky-toaster-86 (node-5e0ea2c8, Omarchy). An exec ran 'adb shell am start -W ...' against an asleep Quest 2. adb blocks forever; the exec had timeout 90 but was never killed. All 4 exec workers ended up wedged on hung adb children (in_flight stuck at 4), the queue backed up, and the node's exec channel was dead until the child processes were killed locally. DELETE /exec/queue clears only queued items and DELETE /exec/queue/{id} refuses in-flight items, so there is no remote recovery path.
Fixes needed: (1) enforce the exec timeout with process-group kill (setpgid + kill(-pgid, SIGKILL)) so children like adb die with the shell; (2) make the cancel endpoint work on in-flight execs with the same group kill; (3) consider a watchdog that reaps any exec exceeding its timeout by 2x.
Recovery used: local pkill of the adb children on the node. Lesson for runbooks: adb commands over cluster exec must always wrap the adb call in 'timeout N' because adb itself never times out.