Steve,
I cannot help with NORMA supporting multiple submodels, but the ability to merge ORM2 models (expressed in plain text using my Constellation Query Language) is a fairly short-term goal of mine - i.e. before the end of the year. CQL is much more manageable in a large project than NORMA, because it uses plain text that makes it easy to apply differential revision management and other tools.
Even without the ability to break a definition into many separate models (CQL calls them vocabularies), it's possible of course to break one model into many files, and concatenate them before processing. I believe that I have a number of O(n^2) or worse algorithm loops which I may need to locate in order to make it feasible to effectively process a model of this size, but with your help, I'd certainly like to try. The SQL schemas I generate are of a generally comparable standard to those of NORMA, as of the release I have.
There's no need for you to write the CQL by hand; my tools can generate it from the NORMA files. If you wish to contact me privately, I'd be happy to receive a copy of your model (in confidence, of course) to see how my tools perform, and for you to see how effective CQL may be as an alternate representation of your model. The conversion to CQL requires that some simple rules are followed with respect to the existence and structure of fact type readings and role names, but the rules are not onerous.
The CQL tools are free open source software, and I can help you get started in using them.