The mapping from objects to EQL specializers on those objects should be weak (if the specializer isn't otherwise referenced, it and the corresponding object can be GCed.)
There's another issue here. A defining macro like DEFMETHOD can expand into code which has arbitrary implementation-dependent execution-time effects (keeping track of source information or other debugging info) as well as the defined effects (ENSURE-METHOD or something equivalent, etc.) There's at least some argument that says that the benefits of retaining debugging information about otherwise unreferenced obsolete methods (and having that retention prevent the methods and/or their specializers from being GCed) outweigh the disadvantages of doing so. (I'm not sure that I believe that; I don't have too much good to say about how CCL handles source-location recording, but there's at least an argument there.)
Rather than (ab)using a defining-macro (DEFMETHOD) to do method redefinition declaratively at runtime, it should be possible to do this procedurally (via ENSURE-METHOD) and not worry about those other effects. It wasn't possible prior to r13980 in the trunk; ENSURE-METHOD was calling CCL::RECORD-SOURCE-FILE, and that should really happen elsewhere in the expansion of DEFMETHOD.
In other words, if your example used ENSURE-METHOD, I think that its memory utilization should be relatively stable and AFAICT it is in the CCL trunk as of r13980.) If the example uses DEFMETHOD, it's not totally unreasonable that environmental/debugging info accrues, and that information costs something; someone who finds that information more useful than I do might consider that to be a reasonable tradeoff in many cases.