Function run_oof_classification 

pub fn run_oof_classification(
    model: &TabFMModel,
    x_train_raw: &[Vec<Value>],
    y_train_codes: &[f64],
    cat_mask: &[bool],
    n_classes: usize,
    configs: &[MemberConfig],
    p: &EnsembleParams,
    num_folds: usize,
) -> Result<Vec<Vec<Vec<f64>>>>

Runs the full n_estimators-member ensemble on each of num_folds folds (fold’s validation rows as query, the rest as context), assembling [member][original_train_row][class] un-shifted OOF logits (every training row appears in exactly one fold’s validation set).

A flattened-across-folds version (grouping folds by (train_size, val_size) shape and running all fold×member tasks through one rayon pass per shape group) was tried here, on the theory that 5 sequential per-fold rounds (default num_folds_for_cv=5) waste parallelism when each round only chunks n_estimators members into a handful of rayon chunks. Measured worse in every configuration tried (same or ~25% slower with default chunking, ~75% slower forcing one batch per shape group) — each fold’s own predict_batch call already saturates Accelerate’s internal BLAS threading on this machine, so adding a rayon layer across folds/shape-groups on top oversubscribes rather than helping (the same class of issue Round 1 documented for thread count). Reverted; kept simple.