I almost applied this patch, but my gut is having second thoughts about it.  I don't think this is the correct solution.  The correct solution would be to delay the encoding of the body part until the message generation phase, and use the requested linesep at that point.  That is, in 3.2 and 3.3 I've changed the paradigm from "always use \n and convert the final string at need" to "specify the linesep when flattening the message".

I'm uploading the completed patch here so I don't lose it, but I don't think I'm going to use it in this form.