this idea of getting the eight adjacent voxels is the direction I was going in.
now, 'outside' voxels can be autonormalised based upon general adjacent outside emptiness. That all seems to work really well- as well as we can expect, anyway
the thing is these voxels that fill the gaps in slopes.
there aren't many of them (e.g. 236 in the TS APC). And we can easily identify them.
I have thought of three approaches:
1) the simplest thing to do is to just apply the most common adjacent auto-normal to them.
2) or we could try slope-detection. Problem with that is that I guess there could be loads of times when these slopes are ambiguous etc
3) actually, surely it is only the diagonals that count? Therefore, there are only eight possible voids.
will elaborate (3) later.. got to go now :(