hm... i did some tests with messing up the second model (moved some vertices around so that the correct triangles won't be found on the second model) and the algorithm seems to be quite forgiving. i didn't get any crashes. just the light map was wrong then on those triangles.

jetpack_monkey's test model worked for me and william's model seemed to work too but he forgot to send me the light map so i couldn't check if it displays correctly. he had overlapping areas in the light map uvs though which probably won't give good results.

jetpack_monkey's test model: