i think with fixed function dot3 bumpmapping it isn't possible to transform the light vector to texture space at every vertex. it doesn't look right on non-planar models with different surface normals.