tsbench

The benchmark Token Savior runs at 100% on. 96 real coding tasks, paired runs, oracle-graded ground truth. Plain Claude scores 78.3%; Token Savior scores 192/192. Reproducible with one command.