Do Large Language Models Play Six Degrees of Separation? Measuring Topological Compression in Long-Context Manifolds