I actually had the geometry in mind when talking about lossless compression. They almost surely perform lossy compression on color and normals (assuming they do not calculate them on the fly). I automatically assumed, however, that they do not throw away any geonetry information when converting from polygonal to octree, because geometric detail is supposed to be a strong point of voxel technology.
This is true. But for storing geometry information you only need 1.5bit per voxel which is really really small (and nice). Material information (color/normal/specular/emissive) can be stored in seperate textures that are strongly compressed both on disk and in graphic memory (DXT1 format for example).
In this thread I've tried to explain how you can get such good compression:
It is actually quite simple.
EDIT: In previous posts you had a fight about color inheritance in child nodes. This kind of inheritance is completely unnecessary because if color is stored in DX format texture, it will always be very well compressed - there is no need for bother with inheritance (will produce messy code). Monochromatic surfaces will be compressed very well because of DX format inner workings.