Coding-agent benchmarks start from a clean checkout. Production hands an agent a retry, a resumed session, or another agent’s branch. This paper rebuilt 198 starting points from 111 tasks — each the state an agent left after declaring the task done while hidden tests still failed — and gave a fresh agent only the original request and that state, with no hint the attempt had failed. Every one of five frontier agents fell 20.6 to 53.5 points below its clean-start score, and the ranking inverted. The failure is not weaker coding: agents read the inherited diff as the definition of the task, keep 86–98% of its lines, and submit with exactly the tests that were already failing.