Yeah! I am gathering the schema is quite extensive.
Update: I put the process on a machine and gave it 60 GB of memory with the xsdvi approach -- no luck, it stabilized at 50 GB memory usage, but ended up crashing at some point in the night after it allocated more. I doubt this approach would have been useful because I'm sure it would be produce a big unusable, unrenderable mess.
Next idea I have is to map out the class relationships from one of the client libraries generated from the schema: C#, python, Java. Those class structures are probably analogous to the schema but less complex.