I'll bite.  I think it's that you can get away with invalid values of
ID-declared attributes if you stick with a conservative
well-formed-enforcement only parser.  The doc. is saying it's on
balance a bad idea to take advantage of this.

In part, I presume, because many XML parsers actually fall into the
useful category acknowledged by the spec, i.e. not validating, but do
read the doctype, in which case they _must_ mark the types of
attributes.  This leads to a borderline contradictory situation with,
for example, as simple a case as <a id="3b"/>.