Planning changes. Not much of the actual implementation should change however, so it's fine to do a first review already.
Also just found a crash when reading transform orientations from old files, think it's easy to fix, JFYI.

Only did very quick review so far, think this is going in the right direction, but… You are still trying to save & reload a 'foreign' data pointer!

fd->globmap & co might be nice trick, but this will not survive linking libraries. Further more, this will make remapping etc. of IDs more complicated, since it will require a special case handling.

Once again: a given datablock shall not be allowed to store a pointer to data from another datablock. Ideally, not at all, ever. In practice, this might be needed for performances reasons - in that case, pointer should be considered as purely runtime, temp cache data. It means that:

1. It should merely be NULL'ed in readfile.c.
2. A way to retrieve pointer should be stored in a robust, failsafe way (here, probably name of custom transform).
3. A way to sanitize/validate/restore proper pointer from that 'failsafe' storage should be provided, as well as a way to tag pointer as dirty (a new flag in twflag? ).

Point 3 is crucial for any serious datablock management! It also allows for lazy loading (you only search for valid pointer when you need it).