Here is a patch of my quick hack, more for interest than any suggestion
it gets used. Although it does produce good output so long as you avoid
the BOM.

The full solution is beyond my (very weak) Python skills. The character
encoding is tied in with XML character substitution (& etc. and
hexadecimal representation of multibyte characters). I could disentangle
it, but I probably wouldn't produce optimal Python, or indeed anything
that wouldn't inspire mirth and/or incredulity.

NB. The workaround suggested by Fredrik Lundh doesn't solve our
particular problems, since the downsize to UTF-8 causes the multi-byte
characters to be represented in hex. Our software doesn't read those. (I
know that's our problem.)