Attached is my proposal (absent the doc changes).  I'm not happy with it, though, because it means that you don't have the MIME-Version header until you do bytes(msg) (or otherwise generate it), *and* you get a MIME-Version header if you do str(part) for a subpart.  That makes debugging significantly confusing, but only with respect to whether or not you have a MIME-Version header.

I'm inclined to commit it, since this whole thing applies *only* to fully constructed messages, rather than parsed-and-manipulated messages.  I don't like it, but I haven't come up with a better solution so far.