My inclination at the moment is to go the small dependency but possibly fragile route of pretty much just directly comparing output with expected output. Is there a better solution that I have overlooked?

Personally I would add the dependency. Better diagnostics for me outweigh any extra installation pain for the user :)

Other ideas:

• Don't package the dependency and SKIP the tests if it's not installed.
• Abstract out the comparison to a separate function and call HTML::Diff if that's installed, otherwise do a straight comparison. Take a look at the Test::Differences docs for an example.