One of the key abstractions we use is that we consider three coordinate systems: PDF, projection, and geodetic. From the publishing perspective, we assume that the georegistrant (georegistor?) knows the projection, parameters and datum used to create the map. Given that, we can do inverse projections, datum shifts, whatever. However, we don't worry about any of that during georegistration. All we're doing is registering two two-dimensional Cartesian coordinate systems*. For a wide variety of maps, and nearly all working maps produced today, this works well. However, you start to run into trouble with scanned images, sketchy underlying landbases, etc. One thing we don't do is munge the image, or the graphical representation of the map. So, what we're going to provide is a way to munge the linear mapping between the PDF and projection coordinate systems, via a multi-point georegistration process. We've played around with this internally, and I hope we can get it to you, soon.
* As I hinted in a previous post, georegistration is slightly preferable to georeference in this context, IMHO. Let me know what you think.