Process-group watchdog: the recorded pid may be /usr/bin/time, which dies on TERM before the group does

Pattern: ( exec setsid /usr/bin/time -o t.txt timeout --foreground N cmd ) & p=$!, then a memory guard kills -p with TERM, sleeps, and sends KILL only if ps -p $p still shows pgid p. Bug: GNU time ignores only SIGINT/SIGQUIT, so the group TERM kills time at once. If $p was started inside $(...), its parent has already exited and init reaps it immediately, so the KILL check finds no process and skips. timeout and cmd stay in the group; if cmd ignores TERM (or handles it slowly), it keeps running after the guard reports success. timeout without -k also waits forever after forwarding TERM. Fix: pick liveness and kill targets by group membership, not by the leader pid: live processes with pgid == sid == g and your user (ps -eo pid=,pgid=,sid=,user=). A pid can't be reused while it is still the id of a live group or session, so kill -- -g is safe while members exist. TERM, poll until no members are left, KILL, poll again, and fail loudly if any remain. Add timeout -k 10. Test it with a fake workload that runs trap '' TERM before allocating memory (children inherit the ignore).