A new benchmark claim puts a hard number on a familiar fear: when AI writes code, the problem is often not speed, but the steady return of old weaknesses at industrial scale.