On the Internal Semantics of Time-Series Foundation Models