cte base64 I think (see below).

Basically, set_payload should be putting the surrogateescape encoded utf-8 into the _payload (which it should now be doing), and probably calling set_charset.  The cte will at that point be 8bit, but when as_string calls Generator, it will get converted to 7bit clean by doing (I think) a base64 encode and emitting that as the CTE.  I have to look through the code to remind myself how it all works, which I haven't had time for yet, which is why I haven't tried to make a fix myself yet.

(Yes, this is poor design, but we are dealing with a long line of legacy code and API here...)