In your post title you mention 0.2.1/0.3.1 but in the text you only mention 0.2.1. Do you know if this problems still exists on develop? I know it might not be possible to test.
There was a bug fix a few months ago which might also be applicable to PolyVox 0.2.1:
https://bitbucket.org/volumesoffun/poly ... eff742bd12Also, it is in theory possible for zero-length normals to be generated. Imagine a 3D checker board for example. In this case the normal at each voxel will be zero, and so the interpolated normal at vertices will also be zero. It's hard to imagine how this would be happening in your case though.
Can you see NaN/Inf normals in the debugger? Can you trace it back at all and see what voxel values generate them?
As for working around it, you also have the option of ignoring the PolyVox normals and computing some yourself via face averaging. Actually, does ignoring the normals fix the black face problem? Cam you turn off lighting completely and/or see anything happening in wireframe mode?