This is causing really awful scheduler behavior where the main thread scheduler is
continually waking up, stealing work, discovering it can't actually run the work,
and sending it off to another scheduler.
No test cases because we don't have suitable instrumentation for it.