Dec 032012
Putting another sphere into the scene revealed that sometimes the reflections were wrong because a negative cos was calculated. The issue was that the change of basis was done incorrectly for transforming the samples. Now it looks sane:
But you can still see a strange highlight at the top of the bottom sphere. My reflectance just based on the cos term, using a real material property, like stainless steel, results in a normal looking bottom sphere (1000 samples):