Sorry, I mean if the boundary box check is engaged, a single
straight line segment would not even be checked for distance
so the problem is still there.

ehm, not really related to int/float.
The last one was instead related to use of the bounding box
for a shortcut. We can't use the bounding box when there
are nonscaled strokes to consider.
For now I dropped the shortcut. Will be slower generally but
more correct. For an improvement, we should maintain a flag
telling us whether nonscaled strokes are there at all.

Note that nonscaled strokes also influence boundaries
for invalidated bounds. If you modify latest testcase so
that it reduces in size, you'll find invalidated bounds
bugs.

I have to add that the bug is only exposed if horizontal
scaling is different from vertical scaling.
The proprietary player also behave in a weird way when this
happens, in that it doesn't seem to consider the caps,
but still gets the horizontal bound correctly.