Is it? ie - construct a normal and take a dot product. That'll work the vast majority of the time. Near the line you'll start running into the FP precision limits. With 64 (...or even 32) bit floats that sort of distance is likely way smaller than anything that represents a practical distance, so just call it and say you are on the line. Of late I've been messing about with problems of this nature and I can't say it's been all that difficult. https://github.com/deadsy/sdfx
You are right that this happens only in edge cases. But you don't want that in a cad program if you have anything doing e.g point-in-polygon, triangulation or other such algorithms.
I know from experience that you really should bake this into the very foundation of a cad package (if you see 2003 me, let him know)
The problem with that is that you will soon want to construct geometrical entities. For example, if you intersect lines that have their ends on integer coordinates you can't represent the intersection on integers (You will need rationals, in that case). The same applies to floating point - two line segments with floating point coordinates will not have their intersection representable in floating point.
So for proper robustness in all situations you need "exact predicates and exact constructions". A middle ground is "exact predicates and inexact constructions". These modes are different kernels usable in e.g. CGAL. http://doc.cgal.org/latest/Kernel_23/group__kernel__predef.h...
Even something seemingly simple such as saying "is point C to the left or right of the line A->B" is very hard in floating point. It's possible without too much sweat to use kind of series expansion where the length of the expansion depends on the precision needed (you only go as far as you need to answer the predicate, so often only 2-3 terms) https://www.cs.cmu.edu/~quake/robust.html