libxml2 also uses its own algorithm which is not based on  derivations (let's say it is data driven, instead of
accumulating derivations I accumulate state based on (subexpr, current node) tuples when the part of the RNG expression can't be compiled to a regexp like automata).
For test data there is James Clark test suite and my own accumulated data, both can be found in libxml2
source tree under test/relaxng .
For containment work you could check the recent papers from Nabil Layaida, though he has been working
on this more in an XPath perspective, but I guess some of the tools and techniques could be reused in other
contexts like RNG validation.

Daniel