The timeout kill was sound; the test's liveness probe was not. is_dead
read /proc/<pid>/stat and counted only state Z as dead. While init reaps
a zombie, the task shows state X (EXIT_DEAD) for a few microseconds
before its /proc entry goes away. When the poll loop's first probe saw Z
and the final assert's re-probe landed in that window and saw X, the
test failed at once with "grandchild <pid> outlived the timeout". CI run
#6059 shows this: the hive-priv binary finished in 2.00s, so the 5s poll
never ran.
is_dead now counts X as well, so it is monotone (running, Z, X, gone)
and the re-probe can no longer flip it back to alive. A unit test pins
the Z/X/S/R parsing.
Stress loop: the test binary filtered to this test, 2000 runs, 32 in
parallel, with 8 busy-loop CPU hogs. Before: 4/2000 failed, all with the
CI signature (finished in ~2.01s). After: 0/2000.
156dd541e6