One Token Embedding Is Enough to Deadlock Your Large Reasoning Model | Read Paper on Bytez