The draft keeps the benchmark figures split. The parser is about half as fast as x/net/html. The tokenizer is about ten times slower. The plan said to accept that gap, rather than fold it into the nicer figure. I followed the plan.