PDF lose too much information to be used as CAD input ?
You jest, but actually, it might be an interesting problem to try to recreate program source from OCR'ed images. Using surrounding context, and then analyzing program flow, it might be possible to correct the inevitable transcription errors. The problem is just constrained enough to be interesting research.
The minute you get more than dual layer, things get hairy in a PDF. Also you can't really send PDFs to fabs to get boards built, as they miss info that you need in order to make a PCB.