ConfigParser.set() doesn't allow non-string arguments
for 'value' any
more.  This breaks Docutils code.  I submit that the
behavior should
not have been changed, rather that the documentation
needed updating.
I volunteer to undo the change and update the
documentation.

ConfigParser.py rev. 1.65 implements the fix detailed
in bug report
810843 (http://python.org/sf/810843):

    Fixed using methods (2) and (3): ConfigParser.set()
now checks
    that the value is some flavor of string, and raises
TypeError if
    it isn't.  This is documented behavior; regression
tests have been
    added.

"This is documented behavior": where was this behavior
documented
before the bug fix?  I couldn't find it.  If it wasn't
documented,
wouldn't method 4 (from the bug report, below) have
been more
appropriate?

    (4) Stating in the documentation that the raw parameter
    should be used when an option's value might not be of
    type string (this might be the intended usage, but it
    isn't made clear).

IOW, perhaps it wasn't a code bug but rather an
omission in the docs.
By adding the restriction and raising TypeError, it
breaks Docutils,
which subclasses ConfigParser.  If the string-only
restriction *was*
documented and I just missed it, I'll accept that the
Docutils code
was at fault (it's not the first time ;-) and rework
it.  But if that
restriction wasn't documented, I don't think
ConfigParser's behavior
should change.

See also
<http://article.gmane.org/gmane.text.docutils.devel/2073>.