I do think the simplification of smooth meshes is useful. As Matt says, there are less opportunities for 'perfect' simplification, but if you accept a minor quality loss then you can remove quite a few triangles. It would probably be most useful for collision meshes, but rendering would benefit too.
The main issue is really that I don't know much about mesh processing, whereas I do know a lot about voxel graphics. I was able to write the old MeshDecimator of course, but I never felt is was optimal or very robust. The algorithm itself is not so complicated, but the underlying data structure should support efficient adding/removing of vertices and a way of finding the neighbours of a given vertex. In PolyVox we just output index/vertex buffers which are ideal for rendering but don't have the above properties.
I then found that
downsampling the volume data worked quite well, and that you could then generate a low resolution mesh from that. Actually there is an example of this in PolyVox though I admit that the example is rather unclear. The approach seemed promising and so I lost some interest in mesh decimation.
Really the whole thing needs more research. In Cubiquity we expect to downsample the volume to create lower LODs (we already protoyped this), but actually it should be possible to do MeshDecimation as well. But for this I would at least like to try external mesh libraries to see how they work out, because mesh simplification feels like it should be a separate task and doesn't need to be part of PolyVox.
Note that all the above only applies to smooth (marching cubes) meshes, as we already have a solution for cubic ones.