Build a paired upgrade gate for Claude Code runs

The system has three layers that never mix: graders that score one draft, statistics that compare two configurations on the same tasks, and a gate that applies rules written down before any candidate ran. The decision that matters most is the unit of comparison: each task's outcome is pass^3 (all three trials pass), and the two configurations are compared task by task, so only the tasks where they disagree carry information. That is what lets the gate see a category regression that a headline pass rate hides.