How fat vertices has been used in the testing? Storing a full tangent space matrix to the input vertices, replicating it to all tessellated vertices and transforming/interpolating it in domain shader for each generated vertex could become a perf hit easily. As tangent frame calculation can be moved to the pixel shader (by using gradient instructions) and some games already are using this trick, it should be benefical to test the tessellation speed with as light input vertices as possible (pixel shader tangent frame calculation will be popular with tessellation).
Yes, it's kind of inexact for culling backface triangles, but very efficient and highly usable for culling offscreen triangles before the tessellation. As you have full control over DS and HS, you know exactly the maximum value that a single vertex can move by your displacement mapping. In our engine for example all displacement maps are 8 bit snorm maps (artists use full range from black to white), and there is a per object (float) scale factor for the displacement. In this case you can just add the displacement scale factor to your viewport culling test, and you get a perfect culling result (and a major performance boost).
It's a specific term in the shader, but a relatively simple one based on the light and viewer directions and wave orientation.
