I've attached a new patch file with some more tweaks, including the additional test case from msg414665.

I've pretty sure that the changes to getpath.py are slightly worse in the new version, but aren't fully correct in the initial version as well (see my previous message about searching outside of the framework). 

I may return to this later when I have time to try to really understand what's going on and compare the logic with the previous C code (especially for framework builds).  But for now I'm giving up.