I suspect that the issue is with the constraint edge recovery process. Currently this uses a very simple process of swapping edge diagonals which works much of the time. However, I am pretty sure that in certain situations this won’t work without inserting additional points. It would be interesting to triangulate without constraint edges and see what you start with, and then see if swapping can produce a correct final result or whether new points need inserting. Maybe we can learn something and improve the algorithm.