Function spawn_oom_watchdog 

pub fn spawn_oom_watchdog(
    threshold_mb: usize,
    interval: Duration,
) -> Option<JoinHandle<()>>

2026-09-25: Spawn the OOM watchdog on the tokio runtime, or return None if one was already spawned.

Every interval it reads cuda_free_memory_bytes. Below threshold_mb MiB it logs an error; at the third consecutive such reading it calls std::process::exit(1). A reading at or above the threshold resets the count, and a None reading leaves it unchanged. Dropping the returned handle does not stop the task.