If you use the tools/MeshToVolume and the input mesh has NANs (or likely INF as well..) it will end up chewing through infinite memory before being killed by the OOM.

My thought is to add to VoxelizePolygons() a just-in-time test for each tri to verify all points are finite, if any aren't, to ignore it.  This will re-check points 6x more frequently than a pre-pass, but I'm hopeful the cost of the check will be shadowed by the other operations.  I'll run some timings to verify this.

I don't like the idea of removing the NANs a the input stage as that will likely be a similar amount of work anyways, and leaves future users of the API vulnerable to NAN injection.