The quota fallback question also has a concurrency side. If several editor requests begin just below the monthly limit, each can independently appear eligible and collectively exceed it before their token usage is recorded.
I would reserve an estimated allowance when a request starts, reconcile it against reported usage, and show the selected provider before the editor accepts the result. That makes the local fallback visible at the point where output quality can change, while giving administrators a budget whose behavior stays understandable under a busy publishing session.