Tracks the OS processes of running checks so code-review runs can be cancelled mid-suite.

Killing the runner task (Task.shutdown/2) closes the check's port but does not kill the external process — a test suite mid-compute survives its port. The tracker closes that gap with a tombstone protocol:

• register_check/2 — the execution layer registers the check's OS pid right after spawn. If the run was already aborted it returns :cancelled and the caller must kill its own process and stop, which closes the race between an abort and the next check spawning.
• abort/1 — writes the run's tombstone, SIGTERMs the currently registered OS pid synchronously (milliseconds), and hands SIGKILL escalation to a fire-and-forget supervised task so callers (the scheduler's cancel path) never sleep.
• release/1 — purges the run's tombstone and entry; the scheduler calls it after the runner task is fully shut down.

Registrant processes are monitored, so entries are purged even when a runner dies without completing its checks.