Pfew… This has been epic (nearly 8 hours today), but I think the core function is now working correctly. For now I’ve setup a very dirty way to expose it to Python (proper solution will probably need to store it in CD and expose it as a loop’s member, as pyrna does not seems to really support arrays of vectors ).
That WIP patch will allow the most motivated to test it (again, it’s not yet usable!) – others will have to champ at the bit! :p It’s too late to use this in OBJ export code and upload a build to graphicall, and I’ll likely not have any blender-time tomorrow, so I’ll try to do this on Thursday…
A very basic profiling shows it’s about twice slower as regular faces+vert normals computation, not that bad!
Btw, if any of you has a good free OBJ visualizer working under Linux, I‘d be very interested, else it’s hard for me to test complex cases! 😉