Because I would like to parse my language for validity. Part of the
language definition allows arbitrary well-formed XML as data input -
so I guess my validity parse can only be "partial". If I use a DTD,
I need to escape the arbitrary XML (hence the CDATA).

It could already help to use ANY as content.
I personally would either try to use refererences to arbitrary content
rather than to put it in inline (and pull it in using document() during
a transformation), or if you validate only batch-wise rather than in
a DTD driven XML editor you could use a transformation which strips
the "arbitrary XML" before validation.