When reset accusers some workers don't get reseted. I guess it has something to do with time the share is submitted and calculated during reset(score=score+share*value)
If this is what is happening, then all that is needed is a mutex_lock/unlock around the score update/reset functions, which is an easy fix. Another option is to trigger second reset if some of the miners has 50+% from total score, which is a workaround, but might be easier to implement depending on the pool structure.
I doubt Slush wouldn't have fixed something simple like this until now, but he is busy with other things lately, so it is possible that he have just missed it.
My personal guess is that it is probably time/clock based and some of the backends is not reset, because it's time is off - that would also explain the situation with the total score not being reset few times too.