This may have been related to the delaunay optimality condition for triangles near the don't-care condition. There was a negative tolerance allowing sub-optimal elements to appear in the mesh which violated some assumptions and led to mesh divergence issues