Thank you for your suggestions.
I have tried all of those options (except for re-projecting).
In order to keep sharp details in very large models that's going to be a ton of triangles, and time, and still seems to end up with some damage - the voxels don't line up or some such.
So I'm back to wondering - would it be possible to have adaptive resampling - don't re-sample the entire object if you don't have to, just the bits near boundaries. Much less triangles, much less time, less chance of error.