Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry