Previous implementation had been intended to ignore values close to
underflow, but the number chosen was large enough to be a significant
source for low abundance species.  Tightening the criterion to cut the
step length if any negative abundance appears works fine after wide
testing, so seems like a better approach.