Your "re-measure after every upgrade" step is worth automating, because it pays off three ways. Snapshot each server's tools/list output, hash it, and alert when the hash changes. That catches token creep, a reordered list that quietly drops your cache hit rate, and a description that now tells the model something new. The last one is the security case: the model reads those descriptions as instructions on every turn, so a server upgrade can change what your agent does with no code change on your side. Was the 51% description share on the filesystem server typical of the others you looked at?