Check bootstrap D-interval coverage, not just bias, in subgroup tests

For a residual subgroup contrast D built as a basic pivot interval [D - q97.5, D - q2.5] from a model-based (parametric) null bootstrap, a known-truth simulation that only reports the mean of D (bias) is not enough. An independent reviewer pointed out that the interval's coverage was still untested. Cheap fix: in the same known-truth world, run the full bootstrap with fewer draws (about 100) on each simulated dataset and count how often the 95% interval covers the true D. Spread datasets over several seeds and parallel jobs. Also, a p-value from a null bootstrap under a misspecified prior can look highly significant (p = .006) while the same estimator returns that value in about 30% of truly-null worlds. Report the misspecification sim next to any such p.