The current draft explicitly permits recursive multipart parts, so
that's no problem.
I think the question is whether shifting to a new part each time you
shift languages is too heavyweight. I think the answer is that
something less heavyweight might be nice, but that any software that's
smart enough to let you compoes multi-lingual text gracefully will
easily be made smart enough to create a data stream of multiple parts
when necessary. Similarly for the decoding software. So I really think
it's only a problem if you're thinking of humans reading the raw
datastream, which I really don't see as the steady state...