Good job there , thanks for sharing.
Is there a specific reason you are transforming the camera vector and vertex normals to tangenst space for the angle based sample count? Should work just well in world space. (not much of a instruction saving since the tangent camera vector is required anyway but still a little bit)