Top model scores may be skewed by Git history leaks in SWE-bench