fix: read group liveness and the sentinel pgid from full ps listings

[Martin Bens (SpiGAndromeda)]

Closes out the BusyBox follow-up work #12 left open: the Alpine test leg is green and gates merges again.

The defect

BusyBox ships ps and pgrep without selection flags, and its usage errors exit 1 — the same status pgrep uses for "group has no members". On Alpine a cancellation therefore read the failed pgrep -g probe as an already-emptied group, ended the grace period at once and dropped the SIGKILL escalation, so a tool that ignores SIGTERM outlived its own cancellation. The lifeline sentinel's ps -o pgid= -p read failed outright, leaving it no group id to kill once the server was gone.

The fix

Both sites now read a full ps -A -o listing (pid and pgid columns) and filter it in bash — the one shape procps, BSD/macOS and BusyBox ps all share. pgrep is no longer invoked anywhere, so it drops out of the requirements; a ps that itself fails still degrades to the raw kill -0 group check, never to "empty", keeping the safe direction: a probe that cannot answer delays the kill rather than dropping it.

The test harness had the same dependence. Its liveness probes (ps -o state= -p — BusyBox has neither the state keyword nor -p) now share one _mcp_proc_state helper reading a pid=,stat= listing, the fixture process census reads stat=,args=, and the hardcoded /usr/bin/true is resolved with type -P (true lives in /bin on Alpine). The pgrep-cannot-answer regression test became a ps stub that intercepts only the liveness scan's column signature, so the unknown-liveness fallback stays pinned without degrading the harness's own ps reads.

Tolerances removed

With the suite green on Debian, Alpine and macOS alike, the Alpine allowances from #12 are gone: the CI leg no longer runs continue-on-error, scripts/test-linux.sh fails a bare run on an Alpine failure, and the README requirements drop pgrep and the procps caveat for Alpine consumers.