Is Token Confidence Just Chain Length? A Signal-Processing Audit of Reasoning Traces