v2.1 — finalized substack + poster (Group-Sobol H-S)

Finalized release for the Saraph Report substack post and the AGU Chapman SLR poster.

Headline change

• Group-Sobol is now the canonical SLR Hawkins-Sutton method, replacing the earlier TreeSHAP/Shapley attribution. Shapley under-counted the emissions axis ~3× (8.6% vs ~27–29% at 2150) because collinear cumulative-emissions features dilute per-feature credit; Group-Sobol decomposes grouped variance directly and is importance-weighted.
• Independent validation: a model-free 324,000-run balanced-factorial ANOVA reproduces the Sobol emissions/climate/internal shares to within ~2 percentage points at 2150 (emissions 27.0% ANOVA vs 28.9% Sobol), confirming the attribution is not a surrogate artifact.

Other changes

• Pulse-SLR figure drops the tipping-corrupted ensemble-mean line (median + 5–95% band retained); pulse GMST keeps its mean.
• Exceedance-table caption corrected to "FaIR v2.2.4 (v1.4.5 calibration)".
• Superseded TreeSHAP-era H-S outputs quarantined.
• .zenodo.json / CITATION.cff updated to describe the Group-Sobol method + ANOVA validator.

Data

Large ensembles (LHS-10k_s BRICK trajectories, ANOVA factorial outputs, FaIR cubes) are not in this repo; they are archived in the companion Zenodo dataset (10.5281/zenodo.20312325) and regenerable from the committed code. Every plotted figure value is backed by the small summary CSVs included here.