Specialize replica selection algorithm in Policy itself

It seems to be a pretty lower hanging fruit that if we apply the algorithm added by #2694 into the Policy's replica selection code itself, it probably can speed up the code that is current run on every mutation.