Move to unsigend arithmetic in LRF box accumulations, and simply guard
the one subtraction that can underflow (that was the reason we were
using signed ints to begin with).  Fixes an overflow in 12-bit encodes.