How AI Coding Agent conducts real repository evaluation: task set, pass rate, cost and regression baseline