Struct VariantSelection 

pub struct VariantSelection {
    pub selected_variant_id: VariantId,
    pub validation_reports: Vec<RegressionMetricReport>,
    pub variant_validation_predictions: Vec<VariantValidationPredictions>,
}

Outcome of native variant selection: the winning variant plus EVERY scored variant’s cross-validation reports, each tagged with its own variant_id.

The reports are the per-fold + cross-fold-OOF-average VALIDATION (OOF) reports collected while ranking. They are emitted so a generated sweep can surface every variant’s CV score — not only the winner’s — to match the legacy per-variant num_predictions. These are REPORT-ONLY validation scores of non-selected models: they never feed any downstream training/feature path (no prediction blocks, no RegressionTargetRecords, no handles leave selection — see select_best_variant_by_cv), so the OOF/leakage invariants are unaffected.

Fields

selected_variant_id: VariantId

The winning variant, ranked by selection_metric. The SELECT DECISION is identical to the pre-existing behavior; validation_reports is purely additive context.

validation_reports: Vec<RegressionMetricReport>

Per-variant VALIDATION (OOF) reports for ALL ranked variants (winner included), each tagged with its variant_id. The cross-fold OOF average per producer is re-tagged with the variant id (its native form has variant_id = None); the per-fold reports already carry it.

variant_validation_predictions: Vec<VariantValidationPredictions>

Per-variant VALIDATION (OOF) PREDICTIONS for ALL ranked variants (winner included), captured from each variant’s transient FIT_CV RunContext BEFORE it is dropped, re-tagged with the variant’s id + content fingerprint. The scalar validation_reports above carry only the score; these carry the per-sample y_pred (+ id-matched y_true) so a host can fill a non-selected variant’s per-fold prediction rows, not just its CV score.

LEAKAGE: these are each variant’s OWN validation (OOF) predictions, re-tagged with that variant’s id (which prevents cross-variant mixing). They are surfaced for host persistence/display only — every transient CV run executes FIT_CV ONLY (no Final/Test/refit), so by construction this carries no train/refit predictions, and the captured blocks never feed a training/feature path or cross a requires_oof edge. This is strictly ADDITIVE — the same values the scalar reports were computed from, exposed per sample — analogous to the additive OOF-average block surfacing; no leakage validator is relaxed.

Trait Implementations

impl Clone for VariantSelection

fn clone(&self) -> VariantSelection

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for VariantSelection

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.