When modular arithmetic transforms have already fired for a
subexpression, and that subexpression's width is at most as wide
as the bitwidth we're cutting to, there is no need to re-traverse
There was already some code to detect that case. Make it more general,
and, more importantly, sound.