Inference with non-differentiable surrogate loss in a general high-dimensional classification framework