From Deceptive Outputs to Deceptive Mechanisms: A Causal Framework for Language-Model Deception Research