How often this process checks whether one of its in-flight runs was marked Cancelled from
outside (the Stop button writes the row; see AgentRunWatchdog.Track). Bounds how long
a stop takes to reach the running agent, so it is deliberately much shorter than the heartbeat.
How often the owning process stamps a liveness heartbeat for its in-flight runs.
A Running run whose heartbeat is older than this (or NULL with an older StartedAt) is force-failed.
How often this process re-scans for orphaned runs (mid-life, without waiting for a restart).
Once this process has tracked a run, the sweep keeps running for the life of the process, even
with nothing tracked, so it also closes runs that other, crashed instances left Running.
0 turns the periodic sweep off (for a database that should be able to pause when idle): the
host's one-time startup sweep (AgentRunWatchdog.SweepOrphanedRuns) still closes runs
orphaned by a crash, on the next restart, and the tracked-set prune moves to the heartbeat.
Tuning for the agent-run watchdog. All durations are in their named units. Defaults are deliberately conservative: the stale threshold is many heartbeat intervals wide, so a momentarily-slow-but-alive process is never mistaken for a dead one.