Sorry for necroing this thread, but I’ve spent a lot of effort trying to solve the issue in-material, to limited success. I’ve posted my analysis on this forum thread.
Moreover, Adam Davis seems to have acknowledged the same problem as an issue in this AnswerHub thread.
Could you clarify this further? If this is not a bug, I think there’s a lack of functionality here. There should either be an option to recompute normals properly, or a way to avoid any normal distortion at all.