I use shelltestrunner for this: the input is a simple markdown document, the command runs pandoc with the filter, and the output is the expected markdown. Of course, it would be easy to use any other text-based input or output format that pandoc supports. I also use this to test the CSL style I developed: I can supply a reference to format in a YAML block in the input, and use `markdown-citations` as the output format to check the formatted result.

The only disadvantage is that testing is a bit slow, because pandoc is invoked once for every test case.