“the simplest thing that can possibily (sic) work.”

Yeah, I agree, it's best to keep everything feature oriented for the foreseeable future since it's, in general, more important for reference implementations to be conformant rather than fast.

Not worrying too much about performance would also allow you to get a release out sooner rather than later ;)

We are implementing a .NET version XProc engine. Our way is very simple, all inputs and outputs are XmlDocument, that is, they are DOM based, of course DOM is easy to be cached. For a sequence of input documents, we create a temp XmlDocument with a special root element, then we can distinct a single document from a sequence of document. It makes our work extreme easy. My first version of XProc engine only takes one week, with processors like: identity, xinclude, xstl, xquery, load, store, choose, foreach etc. I find XProc very powerful. Our workflow engine is in fact built on it.

Passing around DOM trees definitely simplifies some things. Does your .NET implementation have a home page?