The problem is, that you loose precision when you split your polygons.
I think one solution would be to reuse the same vertex (on the GPU) for polygons that have been split. So if you split polygon A into B and C then B and C will share two vertices at their split-border. If you then render B and C, both should index the same vertex in the vertex-buffer array. Because when you use the exact same vertex then two adjacent polygons should not produce such missing pixels.
Jan.
PS: Well, T-vertices are of course a bigger problem. Hm, no idea right now.
Do you believe it can be faster avoiding to generate them?
We collect triangles (cloning them) within our scene from totally different objects. When we process the split for a triangle we can only know the owner, nothing else. This is the reason why I am so skeptical to find a reasonable fast approach.
BTW I still need to read carefully the docs you suggested.