TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs | Read Paper on Bytez