I missed that one too. Thank you for pointing it out.

But after modifying my test code to take account of it I saw no difference in the results.

It turns out that it is the last bullet of the sub-subsection "Restrictions on Portable Programs" that is the key factor in the issue I am interested in here (last bullet of page 144 in AMOP, or here as pointed by Steve). It basically states that the defmethod on slot-value-using-class must happen before any call to make-instance on a class the method uses as a specializer. (And then goes on to motivate this definition order restriction).

If I modify my test code to comply with this rule (see slot_mop_strict.lsp here attached) then sbcl stops being the odd man out and behaves like clisp and ACL (which I think is the proper behavior).

Further, if I agree to stop tickling the system where it does not like to be tickled and define a class tracked_class as a subclass of standard-class to be used as a metaclass of the classes I want to track (see slot_mop_strict_tracked_class.lsp here attached), then all of sbcl, clisp, ccl and ACL show the same behavior, properly calling slot-value-using-class both from slot-value or from the slot accessor. Only LispWorks persists in its (erroneous, I think) behavior of calling s-v-u-c only from slot-value.

I think there is a consensus emerging here.

Thank you all for your help on this matter with a special emphasis toward Steve.