From python's xmlrpclib, I send new page content to Confluence. xmlrpclib explicitly checks if the encoding is UTF-8 and if so, skips emitting the encoding in the prolog XML declaration.

In a storePage method, I post some UTF-8 encoded text. For example, 0xA0 (non-breaking space) is encoded in UTF-8 as 0xC2 0xA0. I have confirmed that these are the bytes on the wire. The OK response from the server has all UTF-8 characters changed to sequences of #65533 entities.

Glue's glue-config.xml has an <xmlEncodings> section near the bottom. Changing this to UTF-8 works around the problem. It's not clear to me that this section was ever doing the right thing for XML.