Pooled five-fold consistency: foundation models pooled like CNN/CLIP; 9-model McNemar (36 pairs); per-class pooled; pooled ensemble report

Add numeric confusion matrix CSVs

Add per-class precision/recall/F1/support CSV

Add academic polish: inference example, hparams, CITATION.cff, dataset figures, per-class CSVs

Add post-hoc analysis: confusion matrices, ROC, reliability, per-class F1, accuracy ranking