IMHO, the general rules of good coding practices should be relaxed when applied to test scripts and debugging. If anything, attempting to parse your way through the raw HTML makes your tests more dependant on implementation details compared to reading the datastructure alone. The things your tests will check are going to be some result that's inside that datastructure anyway, and going through the HTML will just create extra work. You don't necessarily have to check the entire datastructure, but just the pieces you're interested in.

What HTML::Template::Dumper will do is tie you to using HTML::Template. For the orginization I work for, this isn't much of a problem, because HTML::Template best fits the way we operate (e.g., strict seperation of code and data). Though this isn't an itch I personally have the need to scratch, I would like to see other modules for different template systems and work with the authors to integrate them where we can.