When selser runs (html2wt) on modified HTML, we have to figure out whether the serialized wikitext is correct or not. However, there is no original wikitext to compare against since the DOM has been modified.
Currently, parserTests runs wt2wt on the same modified HTML and compares selser output with wt2wt output. wt2wt uses the regular full serializer. This pass/failure detection implicitly assumes that the output from wt2wt is correct. However, this is flawed -- it might actually be the case that selser output is correct and wt2wt is incorrect since wt2wt might be normalizing wikitext whereas selser wont (which is the whole point of selser, so to speak). This was not as much of an issue till a while back since wt2wt was normalizing less and parserTests used to run in rt-testing mode which used src-based rting for templates and the like. However, this is becoming more of a problem as we move towards greater normalization (ex: bug 51004), and when the regular serializer no longer run in rt-testing mode.
So, we need to figure out a more robust solution to this problem.