Intervention bootstrap can omit reporting groups from individual replicates

Summary

bootstrapped_gmean_metric() samples all unique instances globally in the outer bootstrap and only groups by final_groups after resampling.

When final_groups contains multiple reporting groups, a bootstrap replicate can therefore omit a small group entirely. For example, with one instance in dataset A and 99 in dataset B, a size-100 global bootstrap omits A in roughly 37% of replicates. Those missing replicates are then absent from A's bootstrap rows, so A's final mean/std are computed only from the subset of replicates in which A happened to be sampled.

That conflicts with the documented/reporting contract that each reported group receives a bootstrap estimate for each replicate, and it differs from the scaffold's process/outcome bootstrap helper, which resamples instances within each dataset.

Expected behavior

Outer bootstrap sampling should preserve each reporting group and resample instances with replacement within final_groups, so every non-empty reporting group contributes one estimate per bootstrap replicate.

Suggested fix

Build the outer instance table with the reporting labels needed for final_groups, sample each reporting group independently while preserving its instance count, then continue the existing cross-fit and inner-bootstrap pipeline. Add a regression test with a highly imbalanced two-group input that verifies every bootstrap index contains both reporting groups.

Start at bootstrapped_gmean_metric() and compare its outer sampling with the scaffold’s process/outcome bootstrap helper, which resamples within each dataset. Add a regression test using a highly imbalanced two-group input, then verify that every bootstrap index contains both reporting groups and that the existing cross-fit and inner-bootstrap pipeline still runs.