[jansan]

I have used fuzz testing to make my Bezier intersection library more robust against edge cases. The test would try to find all intersections between a random pair of curves that lie within certain bounds (you probably know that there can be up to nine intersections between two cubic Bezier curves). At the beginning it failed at approx. 1 in 100,000 randomly generated curve pairs, now I am at a point where there is not a single failure in a billion test.

My problem was how to decide if a test failed, because this would not be a crash, but failing to find an intersection between the curves. So I compared against an existing library which uses a completely different algorithm, which means the other library fails at other test cases than my own library. If the results in a test case were different, one must have failed and by testing against the found intersections I could easily decide which one.