Ordering the pg_dump output by name within classes instead of OID sounds
good to me, too.
Also, something that might be easier for comparing schemata between
databases: rather than dumping the database, have you tried using PostgreSQL
Autodoc (http://www.rbt.ca/autodoc/) which just outputs the schema in a
variety of formats (including XML for dia and DocBook). It just seems that
if you're only concerned with structure, and not content of user tables,
this could at least shorten the amount of data to be compared, if not get
you into a space where there's already a tool to do all the work (I don't
know if there's an XML-smart diff.)