The way build_annotations() makes `format-annotate-function' use the
same temp buffer " *Format Temp 0*" on each call is not reentrant.  If
TO-FN from the format-alist does anything that writes another buffer
with a buffer-file-format then the buffer which TO-FN is supposed to be
working on is clobbered.

foo.el below illustrates the problem.  Evaluating it writes a /tmp/x
containing

    THIS IS FORMAT ONE
    this is buffer yyy text

where I expected it to be

    THIS IS FORMAT TWO
    hello world

The `message's emitted by my-one-encode and my-two-encode show they get
the same " *Format Temp 0*" buffer.  The latter is what you get in
/tmp/x if taking away the "/tmp/y" blob within `my-two-encode'.

I expect this wouldn't arise often in practice, but it ought to be
reasonably easy to make build_annotations() (or wherever) do better in
the management of temporary buffers.  I expect it's a matter of finding
the right place in the write crunching that temp buffers used can be
killed.