[DylanCosto]

Whole-cohort averages can hide differences in reconstruction sensitivity between participant groups. robustness-groups reuses a completed experiment and an explicit, exhaustive sample/group mapping to report each group's score errors, correlations, cutoff transitions and frozen-reference shifts.

All groups retain the original common-callable terms and full-cohort percentile thresholds; correlations use pairs within each group. The command verifies native output hashes and participant inventories, preserves unavailable reasons and reference overlap scope, and records memberships and fingerprints. The HTML report filters by group and scenario. Labels are researcher-supplied; the command does not infer ancestry or estimate phenotype prediction.

Validation includes hand-calculated global-cutoff and singleton cases, constant cohorts, invalid mappings, altered inputs and escaped labels; ten independent Python checker tests pass. On the completed 128-person public study, two artificial 64-person partitions pass 588 score-metric, 168 cutoff and 84 reference comparisons across 42 scenarios. This is reporting validation, not evidence of demographic differences. The release archive verifies 55 payload hashes and 52 local links, runs the extracted offline experiment/study/group checks, and rebuilds byte-identically. Formatting and Clippy pass.

Stacked on #5. The original genome-wide validation and paired-study evidence remain intact. No release tag or publication is included.