* Two separate bugs, with very similar symptoms. * The distribution binary search didn't work correctly in cases where there were a lot of faces with 0 weights. * Maximum distribution sum should have been exactly 1, but due to the wonderful nature of floats this wasn't the case at all.