Post-Grokking Collapse at the Representation-Readout Interface in Muon-Trained Transformers — ThinkLLM