This is a more stupid way of doing it. I didn't want to do that initially because I'm worried about the aggregation latency of ~400-700ms conflicting with the sleep, so using a callback which is guaranteed to be called when the next round is produced seemed much cleaner.

I'm tempted to do a +1 on the targetTime to mitigate for that.

Okay, ignore my first commit, I reverted to a callback to avoid the above problem, but now it's got a mutex and the impossible $2^{-128}$ edge case is removed.
Also changed the way the context and cancellations are handled.
All edgecases should also be covered now.