does it deal with shapes that have holes in them?

I think it should, though it may need some modification to isolate the two (or more) polygons. I'll need to construct a known testcase and see what happens.

As for the source of the errors. The presence of 'holes' could explain the anomolies I'm seeing. But either way, the problem is exacerbated by the presence of duplicate points that are being returned in individual polys. I've tried printing the values from the simple 'squares' testcase I posted earlier, to the fullest precision Perl can give me, and this isn't a case of points that differ by infinitesimal amounts being mapped to the same pixel. The values as returned bu M::G::V are all exact integer values.

I'm currently filtering each of the individual polys for adjacent duplicate vertices, but that won't handle the case of there being 'extra points' in the solution. If you run the code above with -PAT=hex2 and look at the grey (unfiltered) polys on the left-hand edge, you'll see that there are various edges showing up that do not obviously result from the normal dividors from the points.

In particular, if you look at the poly in the lower left corner, the extreme left-hand edge appears to be parallel to the left edge of the coordinate space (white box). There is no way you should be able to generate a 'parallel normal' from points wholy contained within the white coordinate space--but there it is.

I'm not accusing the underlying libraries of having errors, I just don't know any way to explain the presence of that edge (and others) based on what I know of the algorithms involved. Now I'm wondering what hoops would be involved in compiling the underlying library on Win32? (Did you post a pointer to teh source code anywhere?)

I had to make a couple of minor edits to memory.c (in myalloc() to get it to compile with MSC. Apparently cl doesn't know how to calculate sizeof( void* )? I just switched the two occurances to char* until I got a chance to investigate that, work out the correct solution and produce a patch for you. But I do not see that it could have any affect on the math.

I'll play and try and come up with a simple dataset that produces a 'hole' to see what happens. If you have any ideas on how to produce such a dataset, please let me know. I'm having trouble conceiving of how that can arise right now?