Add the llama-cpp-classify classifier API. Each question becomes one chat prompt with single-token answer labels; llama-server's pre-sampling next-token log-probabilities give the softmax over those labels. Missing labels are retried with deeper readouts and then reported as errors.
