feat(analyzer): composite sensitivity/uncertainty report

Andrew Dunnrequested to merge
p4-sensitivity-report into main

What

Adds a manifold sensitivity analysis (additive -- the normal analyze output and production scores are unchanged) that perturbs the centralized methodology weights and reports how robust the composite outputs are to those weight choices. Method: OECD/JRC uncertainty and sensitivity analysis (Saisana, Saltelli and Tarantola 2005).

Robustness finding (300-user data: 301 users, 20 teams, +/-20% perturbation, 80 runs)

  • Health score varies +/- 4 points (range 41 to 48 around a baseline of 45).
  • User ranking: mean Spearman 1.000 (worst 0.993; top decile 0.992).
  • Team ranking: mean Spearman 1.000 (worst 0.991).
  • Archetype churn: mean 0.5% (max 6.6%).

The composite outputs are robust to the weight choice. One caveat surfaced and is worth recording: DomainBalance (nominal weight 0.20) is as health-sensitive as DomainCoverage (nominal 0.35) -- a real nominal-vs-effective weight gap.

Merge request reports