Indeed, the "free" use of a variable is orthogonal to the "undeclared"
use of a variable, but that does in itself not mean anything.  A Usenet
posting may be factually incorrect, and it may be stupid.  But it may
be stupid without being incorrect, and it may be incorrect without
being stupid, and it may be both stupid and incorrect.  These two
properties are orthogonal.  I know this because I have both read and
writen many Usenet postings over the past 17 years.

Similarly, a reference to a variable may be either or both of "undeclared"
and "free".  Rather than muck around in the meaning of these terms, I'd
ask Erik the following focussed question:

Assume a freshly-booted ANS-compliant CL implementation.  What behavior,
if any, is guaranteed by execution of the following form?

 (print (progn (setq pie 22/7) pie))

Now, in every implementation I know, this will print an approximation of
PI, with (if in the interative top level) various prompting and whitespace
handling.  It might (especially in compiled code) or might not warn about
the undeclared variable PI.  None of this matters to me.

I want to know whether this behavior (modulo toplevel interaction,
prompting, and newlinification) is defined by the ANS.  If you claim that
it is, please cite the ANS section references.  I claim that the behavior
is either "is an error" or "is undefined".