Function run_members_classification 

pub fn run_members_classification(
    model: &TabFMModel,
    x_train_raw: &[Vec<Value>],
    y_train_codes: &[f64],
    x_query_raw: &[Vec<Value>],
    cat_mask: &[bool],
    n_classes: usize,
    configs: &[MemberConfig],
    outlier_threshold: f64,
    batch_size: Option<usize>,
) -> Result<Vec<Vec<Vec<f64>>>>

Runs every ensemble member’s forward pass for classification, given a train/query row split (query rows may be real held-out test rows, or an OOF fold’s validation rows). Members are grouped into batch_size-sized chunks (default: DEFAULT_BATCH_CHUNK_SIZE) and each chunk runs as a single TabFMModel::predict_batch call — sharing the fixed cost of the model’s deepest stage (24-block ICL) across the whole chunk instead of paying it once per member. Chunks run in parallel via rayon, combining with Round 1’s parallelism. Returns [member][query_row][class] logits, already un-shifted back to original class order.