Reproduction: Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models