A 1500ms budget plus the one 100ms poll interval the trailing sleep can overshoot is exactly 1600ms, which was the bound — but the test measures a whole process around that, so startup pushed real runs to 1602ms. Carry another interval. Two intervals of overshoot, the regression this guards, still trips it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>