Statistical Order in Representation Learning: Sufficiency, Architectural Blindness, and Generalization Bounds via Maximum Entropy