Measure prompt fragility

Fragility only means something relative to noise: if two runs of the same prompt differ by 3 cases, a perturbation that moves 3 cases has shown nothing. So the check runs five things on the same 24 dev cases: the original, an identical repeat, and the three perturbations. The decision that matters most is measuring the noise floor directly with the repeat run, rather than assuming it from a formula, and comparing case flips rather than pass rates, because a pass rate can move several points from one flip on a case with three trials.