Thanks for the files. I have the issue sorted out locally. This comes from a sort return on lin 2920 in ORMModel\Framework\Shell\SerializationEngine.cs, which tries to get a consistent sort order when two named elements have the same name. The comment says that deferring to a guid compare wreaks havoc on baselined tests (this is true, it makes the files effectively uncomparable). However, to account for this situation the code simply picked an arbitrary value. It looks like VS2012 is using a slightly different sort algorithm that calls back a second time with (y,x) instead of (x,y). Favoring one over the other based on position has worked through three VS versions, but blows up badly here. The fix is easy: just return 0.
Your model does have a ton of duplicate names in it. These will be noted in the next VS release, which flags duplicate readings in the model. I'm not 100% that this is what is causing you to hit this, but there is a pretty good change. I'd suggest scanning the model browser for duplicate fact types to fix and then seeing if you can generate and save the relational model.
I should release the next official NORMA version within the next couple of weeks. If you're totally blocked on this, let me know and I can get you an intermediate build that fixes the problem.