Write choose(scenario) implementing the decision chain (request-time input → realtime; staleness ok AND enumerable → batch; else the hybrid) and run the four scenarios, printing f"{name:<26} -> {shape}". Only ONE of these genuinely needs live inference.