PID limit exhaustion from zombie processes

A CI build host starts failing new builds with 'fork: retry: Resource temporarily unavailable' and SSH logins are refused, though CPU, memory, and disk all look fine. The host runs many short-lived containerized build steps. ps -eLf | wc -l shows ~30,000 tasks; cat /proc/sys/kernel/pid_max is 32768. A build job introduced last week spawns a per-test helper process but, on a code path hit by a flaky test, never reaps the children, ps shows thousands of <defunct> (zombie) processes parented to that job. Triage and fix.