the float variant was always shadowed by the int version as
Rect2d is implicitly convertible to Rect.
This swaps things which is fine, as the vector of boxes was always
copied and the computation was done in double.