I think the "why not write it all in Haskell" argument is a bit spurious. Only the engineers on the team knew best the advantages and disadvantages of the various approaches. Presumably there were very good reasons to integrate it the way that they did, whether driven by architecture, performance, knowledge, library, smallest-effective-change, or other reasons. It's clearly worked out very well in the most basic sense (it does the job). And on top of that, we now have bug fixes and other improvements submitted to ghc. Seems like a win-win.