Did this change to the coding-agent harness make the code better or worse? Paired, repeated, isolated evaluation with a verdict you can trust. - View it on GitHub