Reject a repeated s1a option before the model is built
A later --option with the same key used to replace the description a System 1 model would see. The command still looked like two choices. Only the last text was kept, and the process did not say a description had been dropped. On the system1-agents checkout tested here, that replacement does not happen. s1a decide exits 2, names the repeated key, and prints neither description.
That exit is the request gate. It is not a score, and it is not a calibration result. A request that gets past it still has to find a key. A request that names a key still leaves the machine. The useful split is which of those failures you are looking at, because they do not mean the same thing and they do not happen in the same place.