xai_engine.py: load property_means from per-ensemble JSON; classify_scenario uses manuscript dual criteria (ratio>=0.70 OR mean signed direction)

xai_engine.py: Fast/Full XAI toggle + skip +0.00 cavity labels + About Ensembles and XAI explainer

Replace per-node occlusion with gradient x input attribution (O(1) backward, ~150x faster)

Cap per-pore occlusion XAI to 40 largest pores (free CPU compute budget)

Initial PoreGCN web tool deployment