Probably a regression from the prevalidator rework - was enforcing non-negative int when precision=0 (i.e. non-float), should have been enforcing only full int range.