Orthographic Constraint Satisfaction and Human Difficulty Alignment in Large Language Models