Matched-fold event-model comparison

compare_event_models_grouped() compares the deterministic I-VT baseline, the probabilistic Random Forest event model, and the temporal-context MLP on identical participant/group-held-out folds.

Every learned model is fitted from scratch inside each training partition. The test-row positions are retained, allowing researchers to verify that all methods were scored on exactly the same samples rather than on independently generated splits.

Per-fold reporting includes accuracy, balanced accuracy, and macro-F1. The same held-out rows are also segmented into events and scored with event precision/recall/F1, temporal IoU, and onset/offset/duration error. Random Forest and ContextMLP additionally receive multiclass Brier score and expected calibration error (ECE). Calibration fields for I-VT are intentionally missing because a deterministic velocity threshold does not produce calibrated class probabilities.

The returned EventModelComparison contains:

• predictions: long-form row-level predictions for every model and validation fold;
• fold_metrics: matched fold-level discrimination and calibration metrics;
• summary: mean and standard deviation across folds for each model; and
• design: the sampling rate, grouping unit, thresholds, context window, and other validation settings required to reconstruct the comparison.

This function is benchmark infrastructure, not evidence that one model is superior. Scientific claims require expert-labelled empirical data, frozen splits, appropriate uncertainty around between-fold differences, and validation at the sampling rates for which suitability is claimed.