Numerical stability is a nightmare for the winding number algorithm. I've spent about 2 years on it already, but I believe I've managed to make it work. I used the paper from Eric Lengyel which worked it out with quadratic and extended it to cubics. I kept the topology classification strategy, but had to work out a different solution for the analytical resolution stitching. I'm planning to write about it; this serie of math articles are actually building up the foundations I need to explain it.

I'd be happy to discuss it though.