Regression in RandomScheduler / fuzz testing

[colin-scott]

Something seems to be wrong with RandomScheduler -- it's no longer able to trigger bugs during fuzz testing after several hours of execution, in cases where it was previously able to find bugs within <10 minute.

I'm guessing the root cause is related to the recent changes I made for InteractiveScheduler. (Although, strangely, ReplayScheduler and InteractiveScheduler are both able to reproduce existing recordings correctly, so this regression only seems to be affecting RandomScheduler?)