My gut feeling is that the number of cases where both operands are
written as literals is exceedingly small when compared to the number of
cases where one or both is an identifier. Since 2to3 doesn't support
type inference, we'd miss the relevant, interesting cases. However, if
you're interested in implementing a fix_literal_division fixer module
that would catch these cases, I'd be happy to advise on such a thing :)