I followed this up with a complete best-of-three against Novice, again operated through Codex. Result: 2–1; the end screen reported 11, 31 and 15 turns and clearly unlocked Warden.
The longer test did expose meaningful decisions. A 13-point attack tied a hidden King wall and failed; spending an entire hand on attacks left nothing to defend with. In another turn, saving J+10 let me block a 16-point attack on my exposed King. The Joker explanation also made clear why a two-card attack matters.
One concrete wording fix: when I selected a 16-point King attack, the preview said “WIN!”, but the CPU then blocked it with exactly 16. The same thing happened with a 14-point preview blocked by 21. “Wins if unblocked” would avoid promising the result before the defender acts.
The 31-turn game included stretches of passing with low cards while waiting for enough attack value. That is the pacing segment I would watch with human testers; this agent-operated match cannot tell you whether they find that suspenseful or repetitive. Keyboard Up targeting worked once the game had focus, while the unlabeled wall-target issue from the first report remains separate. I still haven’t tested audio or mobile.