The fact that so many tweaks were needed on that bit of code is certainly indicative that it's overly complex. For now, I've essentially made the code clear enough so that at least further changes will try to cope with all the different situations:
- HTTPException wrapping a raw exception, which may eventually be non-unicode
- HTTPException wrapping (the message of) a TracError, which may eventually be markup (need to avoid regression on #5066)
- HTTPException wrapping a simple string (unicode or not?)
