Main's CI has been failing intermittently on test_blocked_run_does_not_stall_other_tasks with assert 2 == 1 — e.g. run 30587970128 (the #356 merge) and 30570317824. It's not a flaky test; the scheduler can actually run a task twice.

The sequence:

1. A run parks on an approval. Its task stays due in the store, since next_run only advances after completion.
2. The next tick snapshots due() and spawns a duplicate run_task. The overlap guard is checked inside the spawned coroutine, not at spawn time.
3. If the approval lands just before that tick, the parked run resumes first, finishes, and clears the guard. The duplicate then takes its first step, finds the guard empty, and runs the task again — two runs of an automation that was waiting on a human answer.

The fix claims the guard synchronously in _tick before spawning (no await between check and add) and skips the spawn when a run is already in flight. run_task keeps the same claim for direct callers, so test_scheduler_skips_overlapping_run behaves as before.

The new test forces the exact interleaving deterministically (tick → park → approve → tick → drain). It fails on current main every time and passes with the fix. The existing test passes unchanged since the duplicate spawn no longer exists. Full suite: 1016 passed, 1 skipped.