ins-pricing 0.4.3__py3-none-any.whl → 0.4.5__py3-none-any.whl

ins_pricing/cli/bayesopt_entry_runner.py

@@ -1,1442 +1,1442 @@
-"""
-CLI entry point generated from BayesOpt_AutoPricing.ipynb so the workflow can
-run non‑interactively (e.g., via torchrun).
-
-Example:
-    python -m torch.distributed.run --standalone --nproc_per_node=2 \\
-        ins_pricing/cli/BayesOpt_entry.py \\
-        --config-json ins_pricing/examples/modelling/config_template.json \\
-        --model-keys ft --max-evals 50 --use-ft-ddp
-"""
-
-from __future__ import annotations
-
-from pathlib import Path
-import sys
-
-if __package__ in {None, ""}:
-    repo_root = Path(__file__).resolve().parents[2]
-    if str(repo_root) not in sys.path:
-        sys.path.insert(0, str(repo_root))
-
-import argparse
-import hashlib
-import json
-import os
-from datetime import datetime
-from typing import Any, Dict, List, Optional
-
-import numpy as np
-import pandas as pd
-
-# Use unified import resolver to eliminate nested try/except chains
-from .utils.import_resolver import resolve_imports, setup_sys_path
-from .utils.evaluation_context import (
-    EvaluationContext,
-    TrainingContext,
-    ModelIdentity,
-    DataFingerprint,
-    CalibrationConfig,
-    ThresholdConfig,
-    BootstrapConfig,
-    ReportConfig,
-    RegistryConfig,
-)
-
-# Resolve all imports from a single location
-setup_sys_path()
-_imports = resolve_imports()
-
-ropt = _imports.bayesopt
-PLOT_MODEL_LABELS = _imports.PLOT_MODEL_LABELS
-PYTORCH_TRAINERS = _imports.PYTORCH_TRAINERS
-build_model_names = _imports.build_model_names
-dedupe_preserve_order = _imports.dedupe_preserve_order
-load_dataset = _imports.load_dataset
-parse_model_pairs = _imports.parse_model_pairs
-resolve_data_path = _imports.resolve_data_path
-resolve_path = _imports.resolve_path
-fingerprint_file = _imports.fingerprint_file
-coerce_dataset_types = _imports.coerce_dataset_types
-split_train_test = _imports.split_train_test
-
-add_config_json_arg = _imports.add_config_json_arg
-add_output_dir_arg = _imports.add_output_dir_arg
-resolve_and_load_config = _imports.resolve_and_load_config
-resolve_data_config = _imports.resolve_data_config
-resolve_report_config = _imports.resolve_report_config
-resolve_split_config = _imports.resolve_split_config
-resolve_runtime_config = _imports.resolve_runtime_config
-resolve_output_dirs = _imports.resolve_output_dirs
-
-bootstrap_ci = _imports.bootstrap_ci
-calibrate_predictions = _imports.calibrate_predictions
-eval_metrics_report = _imports.metrics_report
-select_threshold = _imports.select_threshold
-
-ModelArtifact = _imports.ModelArtifact
-ModelRegistry = _imports.ModelRegistry
-drift_psi_report = _imports.drift_psi_report
-group_metrics = _imports.group_metrics
-ReportPayload = _imports.ReportPayload
-write_report = _imports.write_report
-
-configure_run_logging = _imports.configure_run_logging
-plot_loss_curve_common = _imports.plot_loss_curve
-
-import matplotlib
-
-if os.name != "nt" and not os.environ.get("DISPLAY") and not os.environ.get("MPLBACKEND"):
-    matplotlib.use("Agg")
-import matplotlib.pyplot as plt
-
-
-def _parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Batch trainer generated from BayesOpt_AutoPricing notebook."
-    )
-    add_config_json_arg(
-        parser,
-        help_text="Path to the JSON config describing datasets and feature columns.",
-    )
-    parser.add_argument(
-        "--model-keys",
-        nargs="+",
-        default=["ft"],
-        choices=["glm", "xgb", "resn", "ft", "gnn", "all"],
-        help="Space-separated list of trainers to run (e.g., --model-keys glm xgb). Include 'all' to run every trainer.",
-    )
-    parser.add_argument(
-        "--stack-model-keys",
-        nargs="+",
-        default=None,
-        choices=["glm", "xgb", "resn", "ft", "gnn", "all"],
-        help=(
-            "Only used when ft_role != 'model' (FT runs as feature generator). "
-            "When provided (or when config defines stack_model_keys), these trainers run after FT features "
-            "are generated. Use 'all' to run every non-FT trainer."
-        ),
-    )
-    parser.add_argument(
-        "--max-evals",
-        type=int,
-        default=50,
-        help="Optuna trial count per dataset.",
-    )
-    parser.add_argument(
-        "--use-resn-ddp",
-        action="store_true",
-        help="Force ResNet trainer to use DistributedDataParallel.",
-    )
-    parser.add_argument(
-        "--use-ft-ddp",
-        action="store_true",
-        help="Force FT-Transformer trainer to use DistributedDataParallel.",
-    )
-    parser.add_argument(
-        "--use-resn-dp",
-        action="store_true",
-        help="Enable ResNet DataParallel fall-back regardless of config.",
-    )
-    parser.add_argument(
-        "--use-ft-dp",
-        action="store_true",
-        help="Enable FT-Transformer DataParallel fall-back regardless of config.",
-    )
-    parser.add_argument(
-        "--use-gnn-dp",
-        action="store_true",
-        help="Enable GNN DataParallel fall-back regardless of config.",
-    )
-    parser.add_argument(
-        "--use-gnn-ddp",
-        action="store_true",
-        help="Force GNN trainer to use DistributedDataParallel.",
-    )
-    parser.add_argument(
-        "--gnn-no-ann",
-        action="store_true",
-        help="Disable approximate k-NN for GNN graph construction and use exact search.",
-    )
-    parser.add_argument(
-        "--gnn-ann-threshold",
-        type=int,
-        default=None,
-        help="Row threshold above which approximate k-NN is preferred (overrides config).",
-    )
-    parser.add_argument(
-        "--gnn-graph-cache",
-        default=None,
-        help="Optional path to persist/load cached adjacency matrix for GNN.",
-    )
-    parser.add_argument(
-        "--gnn-max-gpu-nodes",
-        type=int,
-        default=None,
-        help="Overrides the maximum node count allowed for GPU k-NN graph construction.",
-    )
-    parser.add_argument(
-        "--gnn-gpu-mem-ratio",
-        type=float,
-        default=None,
-        help="Overrides the fraction of free GPU memory the k-NN builder may consume.",
-    )
-    parser.add_argument(
-        "--gnn-gpu-mem-overhead",
-        type=float,
-        default=None,
-        help="Overrides the temporary GPU memory overhead multiplier for k-NN estimation.",
-    )
-    add_output_dir_arg(
-        parser,
-        help_text="Override output root for models/results/plots.",
-    )
-    parser.add_argument(
-        "--plot-curves",
-        action="store_true",
-        help="Enable lift/diagnostic plots after training (config file may also request plotting).",
-    )
-    parser.add_argument(
-        "--ft-as-feature",
-        action="store_true",
-        help="Alias for --ft-role embedding (keep tuning, export embeddings; skip FT plots/SHAP).",
-    )
-    parser.add_argument(
-        "--ft-role",
-        default=None,
-        choices=["model", "embedding", "unsupervised_embedding"],
-        help="How to use FT: model (default), embedding (export pooling embeddings), or unsupervised_embedding.",
-    )
-    parser.add_argument(
-        "--ft-feature-prefix",
-        default="ft_feat",
-        help="Prefix used for generated FT features (columns: pred_<prefix>_0.. or pred_<prefix>).",
-    )
-    parser.add_argument(
-        "--reuse-best-params",
-        action="store_true",
-        help="Skip Optuna and reuse best_params saved in Results/versions or bestparams CSV when available.",
-    )
-    return parser.parse_args()
-
-
-def _plot_curves_for_model(model: ropt.BayesOptModel, trained_keys: List[str], cfg: Dict) -> None:
-    plot_cfg = cfg.get("plot", {})
-    legacy_lift_flags = {
-        "glm": cfg.get("plot_lift_glm", False),
-        "xgb": cfg.get("plot_lift_xgb", False),
-        "resn": cfg.get("plot_lift_resn", False),
-        "ft": cfg.get("plot_lift_ft", False),
-    }
-    plot_enabled = plot_cfg.get("enable", any(legacy_lift_flags.values()))
-    if not plot_enabled:
-        return
-
-    n_bins = int(plot_cfg.get("n_bins", 10))
-    oneway_enabled = plot_cfg.get("oneway", True)
-
-    available_models = dedupe_preserve_order(
-        [m for m in trained_keys if m in PLOT_MODEL_LABELS]
-    )
-
-    lift_models = plot_cfg.get("lift_models")
-    if lift_models is None:
-        lift_models = [
-            m for m, enabled in legacy_lift_flags.items() if enabled]
-        if not lift_models:
-            lift_models = available_models
-    lift_models = dedupe_preserve_order(
-        [m for m in lift_models if m in available_models]
-    )
-
-    if oneway_enabled:
-        oneway_pred = bool(plot_cfg.get("oneway_pred", False))
-        oneway_pred_models = plot_cfg.get("oneway_pred_models")
-        pred_plotted = False
-        if oneway_pred:
-            if oneway_pred_models is None:
-                oneway_pred_models = lift_models or available_models
-            oneway_pred_models = dedupe_preserve_order(
-                [m for m in oneway_pred_models if m in available_models]
-            )
-            for model_key in oneway_pred_models:
-                label, pred_nme = PLOT_MODEL_LABELS[model_key]
-                if pred_nme not in model.train_data.columns:
-                    print(
-                        f"[Oneway] Missing prediction column '{pred_nme}'; skip.",
-                        flush=True,
-                    )
-                    continue
-                model.plot_oneway(
-                    n_bins=n_bins,
-                    pred_col=pred_nme,
-                    pred_label=label,
-                    plot_subdir="oneway/post",
-                )
-                pred_plotted = True
-        if not oneway_pred or not pred_plotted:
-            model.plot_oneway(n_bins=n_bins, plot_subdir="oneway/post")
-
-    if not available_models:
-        return
-
-    for model_key in lift_models:
-        label, pred_nme = PLOT_MODEL_LABELS[model_key]
-        model.plot_lift(model_label=label, pred_nme=pred_nme, n_bins=n_bins)
-
-    if not plot_cfg.get("double_lift", True) or len(available_models) < 2:
-        return
-
-    raw_pairs = plot_cfg.get("double_lift_pairs")
-    if raw_pairs:
-        pairs = [
-            (a, b)
-            for a, b in parse_model_pairs(raw_pairs)
-            if a in available_models and b in available_models and a != b
-        ]
-    else:
-        pairs = [(a, b) for i, a in enumerate(available_models)
-                 for b in available_models[i + 1:]]
-
-    for first, second in pairs:
-        model.plot_dlift([first, second], n_bins=n_bins)
-
-
-def _plot_loss_curve_for_trainer(model_name: str, trainer) -> None:
-    model_obj = getattr(trainer, "model", None)
-    history = None
-    if model_obj is not None:
-        history = getattr(model_obj, "training_history", None)
-    if not history:
-        history = getattr(trainer, "training_history", None)
-    if not history:
-        return
-    train_hist = list(history.get("train") or [])
-    val_hist = list(history.get("val") or [])
-    if not train_hist and not val_hist:
-        return
-    try:
-        plot_dir = trainer.output.plot_path(
-            f"{model_name}/loss/loss_{model_name}_{trainer.model_name_prefix}.png"
-        )
-    except Exception:
-        default_dir = Path("plot") / model_name / "loss"
-        default_dir.mkdir(parents=True, exist_ok=True)
-        plot_dir = str(
-            default_dir / f"loss_{model_name}_{trainer.model_name_prefix}.png")
-    if plot_loss_curve_common is not None:
-        plot_loss_curve_common(
-            history=history,
-            title=f"{trainer.model_name_prefix} Loss Curve ({model_name})",
-            save_path=plot_dir,
-            show=False,
-        )
-    else:
-        epochs = range(1, max(len(train_hist), len(val_hist)) + 1)
-        fig, ax = plt.subplots(figsize=(8, 4))
-        if train_hist:
-            ax.plot(range(1, len(train_hist) + 1),
-                    train_hist, label="Train Loss", color="tab:blue")
-        if val_hist:
-            ax.plot(range(1, len(val_hist) + 1),
-                    val_hist, label="Validation Loss", color="tab:orange")
-        ax.set_xlabel("Epoch")
-        ax.set_ylabel("Weighted Loss")
-        ax.set_title(
-            f"{trainer.model_name_prefix} Loss Curve ({model_name})")
-        ax.grid(True, linestyle="--", alpha=0.3)
-        ax.legend()
-        plt.tight_layout()
-        plt.savefig(plot_dir, dpi=300)
-        plt.close(fig)
-    print(
-        f"[Plot] Saved loss curve for {model_name}/{trainer.label} -> {plot_dir}")
-
-
-def _sample_arrays(
-    y_true: np.ndarray,
-    y_pred: np.ndarray,
-    *,
-    max_rows: Optional[int],
-    seed: Optional[int],
-) -> tuple[np.ndarray, np.ndarray]:
-    if max_rows is None or max_rows <= 0:
-        return y_true, y_pred
-    n = len(y_true)
-    if n <= max_rows:
-        return y_true, y_pred
-    rng = np.random.default_rng(seed)
-    idx = rng.choice(n, size=int(max_rows), replace=False)
-    return y_true[idx], y_pred[idx]
-
-
-def _compute_psi_report(
-    model: ropt.BayesOptModel,
-    *,
-    features: Optional[List[str]],
-    bins: int,
-    strategy: str,
-) -> Optional[pd.DataFrame]:
-    if drift_psi_report is None:
-        return None
-    psi_features = features or list(getattr(model, "factor_nmes", []))
-    psi_features = [
-        f for f in psi_features if f in model.train_data.columns and f in model.test_data.columns]
-    if not psi_features:
-        return None
-    try:
-        return drift_psi_report(
-            model.train_data[psi_features],
-            model.test_data[psi_features],
-            features=psi_features,
-            bins=int(bins),
-            strategy=str(strategy),
-        )
-    except Exception as exc:
-        print(f"[Report] PSI computation failed: {exc}")
-        return None
-
-
-# --- Refactored helper functions for _evaluate_and_report ---
-
-
-def _apply_calibration(
-    y_true_train: np.ndarray,
-    y_pred_train: np.ndarray,
-    y_pred_test: np.ndarray,
-    calibration_cfg: Dict[str, Any],
-    model_name: str,
-    model_key: str,
-) -> tuple[np.ndarray, np.ndarray, Optional[Dict[str, Any]]]:
-    """Apply calibration to predictions for classification tasks.
-
-    Returns:
-        Tuple of (calibrated_train_preds, calibrated_test_preds, calibration_info)
-    """
-    cal_cfg = dict(calibration_cfg or {})
-    cal_enabled = bool(cal_cfg.get("enable", False) or cal_cfg.get("method"))
-
-    if not cal_enabled or calibrate_predictions is None:
-        return y_pred_train, y_pred_test, None
-
-    method = cal_cfg.get("method", "sigmoid")
-    max_rows = cal_cfg.get("max_rows")
-    seed = cal_cfg.get("seed")
-    y_cal, p_cal = _sample_arrays(
-        y_true_train, y_pred_train, max_rows=max_rows, seed=seed)
-
-    try:
-        calibrator = calibrate_predictions(y_cal, p_cal, method=method)
-        calibrated_train = calibrator.predict(y_pred_train)
-        calibrated_test = calibrator.predict(y_pred_test)
-        calibration_info = {"method": calibrator.method, "max_rows": max_rows}
-        return calibrated_train, calibrated_test, calibration_info
-    except Exception as exc:
-        print(f"[Report] Calibration failed for {model_name}/{model_key}: {exc}")
-        return y_pred_train, y_pred_test, None
-
-
-def _select_classification_threshold(
-    y_true_train: np.ndarray,
-    y_pred_train_eval: np.ndarray,
-    threshold_cfg: Dict[str, Any],
-) -> tuple[float, Optional[Dict[str, Any]]]:
-    """Select threshold for classification predictions.
-
-    Returns:
-        Tuple of (threshold_value, threshold_info)
-    """
-    thr_cfg = dict(threshold_cfg or {})
-    thr_enabled = bool(
-        thr_cfg.get("enable", False)
-        or thr_cfg.get("metric")
-        or thr_cfg.get("value") is not None
-    )
-
-    if thr_cfg.get("value") is not None:
-        threshold_value = float(thr_cfg["value"])
-        return threshold_value, {"threshold": threshold_value, "source": "fixed"}
-
-    if thr_enabled and select_threshold is not None:
-        max_rows = thr_cfg.get("max_rows")
-        seed = thr_cfg.get("seed")
-        y_thr, p_thr = _sample_arrays(
-            y_true_train, y_pred_train_eval, max_rows=max_rows, seed=seed)
-        threshold_info = select_threshold(
-            y_thr,
-            p_thr,
-            metric=thr_cfg.get("metric", "f1"),
-            min_positive_rate=thr_cfg.get("min_positive_rate"),
-            grid=thr_cfg.get("grid", 99),
-        )
-        return float(threshold_info.get("threshold", 0.5)), threshold_info
-
-    return 0.5, None
-
-
-def _compute_classification_metrics(
-    y_true_test: np.ndarray,
-    y_pred_test_eval: np.ndarray,
-    threshold_value: float,
-) -> Dict[str, Any]:
-    """Compute metrics for classification task."""
-    metrics = eval_metrics_report(
-        y_true_test,
-        y_pred_test_eval,
-        task_type="classification",
-        threshold=threshold_value,
-    )
-    precision = float(metrics.get("precision", 0.0))
-    recall = float(metrics.get("recall", 0.0))
-    f1 = 0.0 if (precision + recall) == 0 else 2 * precision * recall / (precision + recall)
-    metrics["f1"] = float(f1)
-    metrics["threshold"] = float(threshold_value)
-    return metrics
-
-
-def _compute_bootstrap_ci(
-    y_true_test: np.ndarray,
-    y_pred_test_eval: np.ndarray,
-    weight_test: Optional[np.ndarray],
-    metrics: Dict[str, Any],
-    bootstrap_cfg: Dict[str, Any],
-    task_type: str,
-) -> Dict[str, Dict[str, float]]:
-    """Compute bootstrap confidence intervals for metrics."""
-    if not bootstrap_cfg or not bool(bootstrap_cfg.get("enable", False)) or bootstrap_ci is None:
-        return {}
-
-    metric_names = bootstrap_cfg.get("metrics")
-    if not metric_names:
-        metric_names = [name for name in metrics.keys() if name != "threshold"]
-    n_samples = int(bootstrap_cfg.get("n_samples", 200))
-    ci = float(bootstrap_cfg.get("ci", 0.95))
-    seed = bootstrap_cfg.get("seed")
-
-    def _metric_fn(y_true, y_pred, weight=None):
-        vals = eval_metrics_report(
-            y_true,
-            y_pred,
-            task_type=task_type,
-            weight=weight,
-            threshold=metrics.get("threshold", 0.5),
-        )
-        if task_type == "classification":
-            prec = float(vals.get("precision", 0.0))
-            rec = float(vals.get("recall", 0.0))
-            vals["f1"] = 0.0 if (prec + rec) == 0 else 2 * prec * rec / (prec + rec)
-        return vals
-
-    bootstrap_results: Dict[str, Dict[str, float]] = {}
-    for name in metric_names:
-        if name not in metrics:
-            continue
-        ci_result = bootstrap_ci(
-            lambda y_t, y_p, w=None: float(_metric_fn(y_t, y_p, w).get(name, 0.0)),
-            y_true_test,
-            y_pred_test_eval,
-            weight=weight_test,
-            n_samples=n_samples,
-            ci=ci,
-            seed=seed,
-        )
-        bootstrap_results[str(name)] = ci_result
-
-    return bootstrap_results
-
-
-def _compute_validation_table(
-    model: ropt.BayesOptModel,
-    pred_col: str,
-    report_group_cols: Optional[List[str]],
-    weight_col: Optional[str],
-    model_name: str,
-    model_key: str,
-) -> Optional[pd.DataFrame]:
-    """Compute grouped validation metrics table."""
-    if not report_group_cols or group_metrics is None:
-        return None
-
-    available_groups = [
-        col for col in report_group_cols if col in model.test_data.columns
-    ]
-    if not available_groups:
-        return None
-
-    try:
-        validation_table = group_metrics(
-            model.test_data,
-            actual_col=model.resp_nme,
-            pred_col=pred_col,
-            group_cols=available_groups,
-            weight_col=weight_col if weight_col and weight_col in model.test_data.columns else None,
-        )
-        counts = (
-            model.test_data.groupby(available_groups, dropna=False)
-            .size()
-            .reset_index(name="count")
-        )
-        return validation_table.merge(counts, on=available_groups, how="left")
-    except Exception as exc:
-        print(f"[Report] group_metrics failed for {model_name}/{model_key}: {exc}")
-        return None
-
-
-def _compute_risk_trend(
-    model: ropt.BayesOptModel,
-    pred_col: str,
-    report_time_col: Optional[str],
-    report_time_freq: str,
-    report_time_ascending: bool,
-    weight_col: Optional[str],
-    model_name: str,
-    model_key: str,
-) -> Optional[pd.DataFrame]:
-    """Compute time-series risk trend metrics."""
-    if not report_time_col or group_metrics is None:
-        return None
-
-    if report_time_col not in model.test_data.columns:
-        return None
-
-    try:
-        time_df = model.test_data.copy()
-        time_series = pd.to_datetime(time_df[report_time_col], errors="coerce")
-        time_df = time_df.loc[time_series.notna()].copy()
-
-        if time_df.empty:
-            return None
-
-        time_df["_time_bucket"] = (
-            pd.to_datetime(time_df[report_time_col], errors="coerce")
-            .dt.to_period(report_time_freq)
-            .dt.to_timestamp()
-        )
-        risk_trend = group_metrics(
-            time_df,
-            actual_col=model.resp_nme,
-            pred_col=pred_col,
-            group_cols=["_time_bucket"],
-            weight_col=weight_col if weight_col and weight_col in time_df.columns else None,
-        )
-        counts = (
-            time_df.groupby("_time_bucket", dropna=False)
-            .size()
-            .reset_index(name="count")
-        )
-        risk_trend = risk_trend.merge(counts, on="_time_bucket", how="left")
-        risk_trend = risk_trend.sort_values(
-            "_time_bucket", ascending=bool(report_time_ascending)
-        ).reset_index(drop=True)
-        return risk_trend.rename(columns={"_time_bucket": report_time_col})
-    except Exception as exc:
-        print(f"[Report] time metrics failed for {model_name}/{model_key}: {exc}")
-        return None
-
-
-def _write_metrics_json(
-    report_root: Path,
-    model_name: str,
-    model_key: str,
-    version: str,
-    metrics: Dict[str, Any],
-    threshold_info: Optional[Dict[str, Any]],
-    calibration_info: Optional[Dict[str, Any]],
-    bootstrap_results: Dict[str, Dict[str, float]],
-    data_path: Path,
-    data_fingerprint: Dict[str, Any],
-    config_sha: str,
-    pred_col: str,
-    task_type: str,
-) -> Path:
-    """Write metrics to JSON file and return the path."""
-    metrics_payload = {
-        "model_name": model_name,
-        "model_key": model_key,
-        "model_version": version,
-        "metrics": metrics,
-        "threshold": threshold_info,
-        "calibration": calibration_info,
-        "bootstrap": bootstrap_results,
-        "data_path": str(data_path),
-        "data_fingerprint": data_fingerprint,
-        "config_sha256": config_sha,
-        "pred_col": pred_col,
-        "task_type": task_type,
-    }
-    metrics_path = report_root / f"{model_name}_{model_key}_metrics.json"
-    metrics_path.write_text(
-        json.dumps(metrics_payload, indent=2, ensure_ascii=True),
-        encoding="utf-8",
-    )
-    return metrics_path
-
-
-def _write_model_report(
-    report_root: Path,
-    model_name: str,
-    model_key: str,
-    version: str,
-    metrics: Dict[str, Any],
-    risk_trend: Optional[pd.DataFrame],
-    psi_report_df: Optional[pd.DataFrame],
-    validation_table: Optional[pd.DataFrame],
-    calibration_info: Optional[Dict[str, Any]],
-    threshold_info: Optional[Dict[str, Any]],
-    bootstrap_results: Dict[str, Dict[str, float]],
-    config_sha: str,
-    data_fingerprint: Dict[str, Any],
-) -> Optional[Path]:
-    """Write model report and return the path."""
-    if ReportPayload is None or write_report is None:
-        return None
-
-    notes_lines = [
-        f"- Config SHA256: {config_sha}",
-        f"- Data fingerprint: {data_fingerprint.get('sha256_prefix')}",
-    ]
-    if calibration_info:
-        notes_lines.append(f"- Calibration: {calibration_info.get('method')}")
-    if threshold_info:
-        notes_lines.append(f"- Threshold selection: {threshold_info}")
-    if bootstrap_results:
-        notes_lines.append("- Bootstrap: see metrics JSON for CI")
-
-    payload = ReportPayload(
-        model_name=f"{model_name}/{model_key}",
-        model_version=version,
-        metrics={k: float(v) for k, v in metrics.items()},
-        risk_trend=risk_trend,
-        drift_report=psi_report_df,
-        validation_table=validation_table,
-        extra_notes="\n".join(notes_lines),
-    )
-    return write_report(
-        payload,
-        report_root / f"{model_name}_{model_key}_report.md",
-    )
-
-
-def _register_model_to_registry(
-    model: ropt.BayesOptModel,
-    model_name: str,
-    model_key: str,
-    version: str,
-    metrics: Dict[str, Any],
-    task_type: str,
-    data_path: Path,
-    data_fingerprint: Dict[str, Any],
-    config_sha: str,
-    registry_path: Optional[str],
-    registry_tags: Dict[str, Any],
-    registry_status: str,
-    report_path: Optional[Path],
-    metrics_path: Path,
-    cfg: Dict[str, Any],
-) -> None:
-    """Register model artifacts to the model registry."""
-    if ModelRegistry is None or ModelArtifact is None:
-        return
-
-    registry = ModelRegistry(
-        registry_path
-        if registry_path
-        else Path(model.output_manager.result_dir) / "model_registry.json"
-    )
-
-    tags = {str(k): str(v) for k, v in (registry_tags or {}).items()}
-    tags.update({
-        "model_key": str(model_key),
-        "task_type": str(task_type),
-        "data_path": str(data_path),
-        "data_sha256_prefix": str(data_fingerprint.get("sha256_prefix", "")),
-        "data_size": str(data_fingerprint.get("size", "")),
-        "data_mtime": str(data_fingerprint.get("mtime", "")),
-        "config_sha256": str(config_sha),
-    })
-
-    artifacts = _collect_model_artifacts(
-        model, model_name, model_key, report_path, metrics_path, cfg
-    )
-
-    registry.register(
-        name=str(model_name),
-        version=version,
-        metrics={k: float(v) for k, v in metrics.items()},
-        tags=tags,
-        artifacts=artifacts,
-        status=str(registry_status or "candidate"),
-        notes=f"model_key={model_key}",
-    )
-
-
-def _collect_model_artifacts(
-    model: ropt.BayesOptModel,
-    model_name: str,
-    model_key: str,
-    report_path: Optional[Path],
-    metrics_path: Path,
-    cfg: Dict[str, Any],
-) -> List:
-    """Collect all model artifacts for registry."""
-    artifacts = []
-
-    # Trained model artifact
-    trainer = model.trainers.get(model_key)
-    if trainer is not None:
-        try:
-            model_path = trainer.output.model_path(trainer._get_model_filename())
-            if os.path.exists(model_path):
-                artifacts.append(ModelArtifact(path=model_path, description="trained model"))
-        except Exception:
-            pass
-
-    # Report artifact
-    if report_path is not None:
-        artifacts.append(ModelArtifact(path=str(report_path), description="model report"))
-
-    # Metrics JSON artifact
-    if metrics_path.exists():
-        artifacts.append(ModelArtifact(path=str(metrics_path), description="metrics json"))
-
-    # Preprocess artifacts
-    if bool(cfg.get("save_preprocess", False)):
-        artifact_path = cfg.get("preprocess_artifact_path")
-        if artifact_path:
-            preprocess_path = Path(str(artifact_path))
-            if not preprocess_path.is_absolute():
-                preprocess_path = Path(model.output_manager.result_dir) / preprocess_path
-        else:
-            preprocess_path = Path(model.output_manager.result_path(
-                f"{model.model_nme}_preprocess.json"
-            ))
-        if preprocess_path.exists():
-            artifacts.append(
-                ModelArtifact(path=str(preprocess_path), description="preprocess artifacts")
-            )
-
-    # Prediction cache artifacts
-    if bool(cfg.get("cache_predictions", False)):
-        cache_dir = cfg.get("prediction_cache_dir")
-        if cache_dir:
-            pred_root = Path(str(cache_dir))
-            if not pred_root.is_absolute():
-                pred_root = Path(model.output_manager.result_dir) / pred_root
-        else:
-            pred_root = Path(model.output_manager.result_dir) / "predictions"
-        ext = "csv" if str(cfg.get("prediction_cache_format", "parquet")).lower() == "csv" else "parquet"
-        for split_label in ("train", "test"):
-            pred_path = pred_root / f"{model_name}_{model_key}_{split_label}.{ext}"
-            if pred_path.exists():
-                artifacts.append(
-                    ModelArtifact(path=str(pred_path), description=f"predictions {split_label}")
-                )
-
-    return artifacts
-
-
-def _evaluate_and_report(
-    model: ropt.BayesOptModel,
-    *,
-    model_name: str,
-    model_key: str,
-    cfg: Dict[str, Any],
-    data_path: Path,
-    data_fingerprint: Dict[str, Any],
-    report_output_dir: Optional[str],
-    report_group_cols: Optional[List[str]],
-    report_time_col: Optional[str],
-    report_time_freq: str,
-    report_time_ascending: bool,
-    psi_report_df: Optional[pd.DataFrame],
-    calibration_cfg: Dict[str, Any],
-    threshold_cfg: Dict[str, Any],
-    bootstrap_cfg: Dict[str, Any],
-    register_model: bool,
-    registry_path: Optional[str],
-    registry_tags: Dict[str, Any],
-    registry_status: str,
-    run_id: str,
-    config_sha: str,
-) -> None:
-    """Evaluate model predictions and generate reports.
-
-    This function orchestrates the evaluation pipeline:
-    1. Extract predictions and ground truth
-    2. Apply calibration (for classification)
-    3. Select threshold (for classification)
-    4. Compute metrics
-    5. Compute bootstrap confidence intervals
-    6. Generate validation tables and risk trends
-    7. Write reports and register model
-    """
-    if eval_metrics_report is None:
-        print("[Report] Skip evaluation: metrics module unavailable.")
-        return
-
-    pred_col = PLOT_MODEL_LABELS.get(model_key, (None, f"pred_{model_key}"))[1]
-    if pred_col not in model.test_data.columns:
-        print(f"[Report] Missing prediction column '{pred_col}' for {model_name}/{model_key}; skip.")
-        return
-
-    # Extract predictions and weights
-    weight_col = getattr(model, "weight_nme", None)
-    y_true_train = model.train_data[model.resp_nme].to_numpy(dtype=float, copy=False)
-    y_true_test = model.test_data[model.resp_nme].to_numpy(dtype=float, copy=False)
-    y_pred_train = model.train_data[pred_col].to_numpy(dtype=float, copy=False)
-    y_pred_test = model.test_data[pred_col].to_numpy(dtype=float, copy=False)
-    weight_test = (
-        model.test_data[weight_col].to_numpy(dtype=float, copy=False)
-        if weight_col and weight_col in model.test_data.columns
-        else None
-    )
-
-    task_type = str(cfg.get("task_type", getattr(model, "task_type", "regression")))
-
-    # Process based on task type
-    if task_type == "classification":
-        y_pred_train = np.clip(y_pred_train, 0.0, 1.0)
-        y_pred_test = np.clip(y_pred_test, 0.0, 1.0)
-
-        y_pred_train_eval, y_pred_test_eval, calibration_info = _apply_calibration(
-            y_true_train, y_pred_train, y_pred_test, calibration_cfg, model_name, model_key
-        )
-        threshold_value, threshold_info = _select_classification_threshold(
-            y_true_train, y_pred_train_eval, threshold_cfg
-        )
-        metrics = _compute_classification_metrics(y_true_test, y_pred_test_eval, threshold_value)
-    else:
-        y_pred_test_eval = y_pred_test
-        calibration_info = None
-        threshold_info = None
-        metrics = eval_metrics_report(
-            y_true_test, y_pred_test_eval, task_type=task_type, weight=weight_test
-        )
-
-    # Compute bootstrap confidence intervals
-    bootstrap_results = _compute_bootstrap_ci(
-        y_true_test, y_pred_test_eval, weight_test, metrics, bootstrap_cfg, task_type
-    )
-
-    # Compute validation table and risk trend
-    validation_table = _compute_validation_table(
-        model, pred_col, report_group_cols, weight_col, model_name, model_key
-    )
-    risk_trend = _compute_risk_trend(
-        model, pred_col, report_time_col, report_time_freq,
-        report_time_ascending, weight_col, model_name, model_key
-    )
-
-    # Setup output directory
-    report_root = (
-        Path(report_output_dir)
-        if report_output_dir
-        else Path(model.output_manager.result_dir) / "reports"
-    )
-    report_root.mkdir(parents=True, exist_ok=True)
-    version = f"{model_key}_{run_id}"
-
-    # Write metrics JSON
-    metrics_path = _write_metrics_json(
-        report_root, model_name, model_key, version, metrics,
-        threshold_info, calibration_info, bootstrap_results,
-        data_path, data_fingerprint, config_sha, pred_col, task_type
-    )
-
-    # Write model report
-    report_path = _write_model_report(
-        report_root, model_name, model_key, version, metrics,
-        risk_trend, psi_report_df, validation_table,
-        calibration_info, threshold_info, bootstrap_results,
-        config_sha, data_fingerprint
-    )
-
-    # Register model
-    if register_model:
-        _register_model_to_registry(
-            model, model_name, model_key, version, metrics, task_type,
-            data_path, data_fingerprint, config_sha, registry_path,
-            registry_tags, registry_status, report_path, metrics_path, cfg
-        )
-
-
-def _evaluate_with_context(
-    model: ropt.BayesOptModel,
-    ctx: EvaluationContext,
-) -> None:
-    """Evaluate model predictions using context object.
-
-    This is a cleaner interface that uses the EvaluationContext dataclass
-    instead of 19+ individual parameters.
-    """
-    _evaluate_and_report(
-        model,
-        model_name=ctx.identity.model_name,
-        model_key=ctx.identity.model_key,
-        cfg=ctx.cfg,
-        data_path=ctx.data_path,
-        data_fingerprint=ctx.data_fingerprint.to_dict(),
-        report_output_dir=ctx.report.output_dir,
-        report_group_cols=ctx.report.group_cols,
-        report_time_col=ctx.report.time_col,
-        report_time_freq=ctx.report.time_freq,
-        report_time_ascending=ctx.report.time_ascending,
-        psi_report_df=ctx.psi_report_df,
-        calibration_cfg={
-            "enable": ctx.calibration.enable,
-            "method": ctx.calibration.method,
-            "max_rows": ctx.calibration.max_rows,
-            "seed": ctx.calibration.seed,
-        },
-        threshold_cfg={
-            "enable": ctx.threshold.enable,
-            "metric": ctx.threshold.metric,
-            "value": ctx.threshold.value,
-            "min_positive_rate": ctx.threshold.min_positive_rate,
-            "grid": ctx.threshold.grid,
-            "max_rows": ctx.threshold.max_rows,
-            "seed": ctx.threshold.seed,
-        },
-        bootstrap_cfg={
-            "enable": ctx.bootstrap.enable,
-            "metrics": ctx.bootstrap.metrics,
-            "n_samples": ctx.bootstrap.n_samples,
-            "ci": ctx.bootstrap.ci,
-            "seed": ctx.bootstrap.seed,
-        },
-        register_model=ctx.registry.register,
-        registry_path=ctx.registry.path,
-        registry_tags=ctx.registry.tags,
-        registry_status=ctx.registry.status,
-        run_id=ctx.run_id,
-        config_sha=ctx.config_sha,
-    )
-
-
-def _create_ddp_barrier(dist_ctx: TrainingContext):
-    """Create a DDP barrier function for distributed training synchronization."""
-    def _ddp_barrier(reason: str) -> None:
-        if not dist_ctx.is_distributed:
-            return
-        torch_mod = getattr(ropt, "torch", None)
-        dist_mod = getattr(torch_mod, "distributed", None)
-        if dist_mod is None:
-            return
-        try:
-            if not getattr(dist_mod, "is_available", lambda: False)():
-                return
-            if not dist_mod.is_initialized():
-                ddp_ok, _, _, _ = ropt.DistributedUtils.setup_ddp()
-                if not ddp_ok or not dist_mod.is_initialized():
-                    return
-            dist_mod.barrier()
-        except Exception as exc:
-            print(f"[DDP] barrier failed during {reason}: {exc}", flush=True)
-            raise
-    return _ddp_barrier
-
-
-def train_from_config(args: argparse.Namespace) -> None:
-    script_dir = Path(__file__).resolve().parents[1]
-    config_path, cfg = resolve_and_load_config(
-        args.config_json,
-        script_dir,
-        required_keys=["data_dir", "model_list",
-                       "model_categories", "target", "weight"],
-    )
-    plot_requested = bool(args.plot_curves or cfg.get("plot_curves", False))
-    config_sha = hashlib.sha256(config_path.read_bytes()).hexdigest()
-    run_id = datetime.utcnow().strftime("%Y%m%d_%H%M%S")
-
-    # Use TrainingContext for distributed training state
-    dist_ctx = TrainingContext.from_env()
-    dist_world_size = dist_ctx.world_size
-    dist_rank = dist_ctx.rank
-    dist_active = dist_ctx.is_distributed
-    is_main_process = dist_ctx.is_main_process
-    _ddp_barrier = _create_ddp_barrier(dist_ctx)
-
-    data_dir, data_format, data_path_template, dtype_map = resolve_data_config(
-        cfg,
-        config_path,
-        create_data_dir=True,
-    )
-    runtime_cfg = resolve_runtime_config(cfg)
-    ddp_min_rows = runtime_cfg["ddp_min_rows"]
-    bo_sample_limit = runtime_cfg["bo_sample_limit"]
-    cache_predictions = runtime_cfg["cache_predictions"]
-    prediction_cache_dir = runtime_cfg["prediction_cache_dir"]
-    prediction_cache_format = runtime_cfg["prediction_cache_format"]
-    report_cfg = resolve_report_config(cfg)
-    report_output_dir = report_cfg["report_output_dir"]
-    report_group_cols = report_cfg["report_group_cols"]
-    report_time_col = report_cfg["report_time_col"]
-    report_time_freq = report_cfg["report_time_freq"]
-    report_time_ascending = report_cfg["report_time_ascending"]
-    psi_bins = report_cfg["psi_bins"]
-    psi_strategy = report_cfg["psi_strategy"]
-    psi_features = report_cfg["psi_features"]
-    calibration_cfg = report_cfg["calibration_cfg"]
-    threshold_cfg = report_cfg["threshold_cfg"]
-    bootstrap_cfg = report_cfg["bootstrap_cfg"]
-    register_model = report_cfg["register_model"]
-    registry_path = report_cfg["registry_path"]
-    registry_tags = report_cfg["registry_tags"]
-    registry_status = report_cfg["registry_status"]
-    data_fingerprint_max_bytes = report_cfg["data_fingerprint_max_bytes"]
-    report_enabled = report_cfg["report_enabled"]
-
-    split_cfg = resolve_split_config(cfg)
-    prop_test = split_cfg["prop_test"]
-    holdout_ratio = split_cfg["holdout_ratio"]
-    val_ratio = split_cfg["val_ratio"]
-    split_strategy = split_cfg["split_strategy"]
-    split_group_col = split_cfg["split_group_col"]
-    split_time_col = split_cfg["split_time_col"]
-    split_time_ascending = split_cfg["split_time_ascending"]
-    cv_strategy = split_cfg["cv_strategy"]
-    cv_group_col = split_cfg["cv_group_col"]
-    cv_time_col = split_cfg["cv_time_col"]
-    cv_time_ascending = split_cfg["cv_time_ascending"]
-    cv_splits = split_cfg["cv_splits"]
-    ft_oof_folds = split_cfg["ft_oof_folds"]
-    ft_oof_strategy = split_cfg["ft_oof_strategy"]
-    ft_oof_shuffle = split_cfg["ft_oof_shuffle"]
-    save_preprocess = runtime_cfg["save_preprocess"]
-    preprocess_artifact_path = runtime_cfg["preprocess_artifact_path"]
-    rand_seed = runtime_cfg["rand_seed"]
-    epochs = runtime_cfg["epochs"]
-    output_cfg = resolve_output_dirs(
-        cfg,
-        config_path,
-        output_override=args.output_dir,
-    )
-    output_dir = output_cfg["output_dir"]
-    reuse_best_params = bool(
-        args.reuse_best_params or runtime_cfg["reuse_best_params"])
-    xgb_max_depth_max = runtime_cfg["xgb_max_depth_max"]
-    xgb_n_estimators_max = runtime_cfg["xgb_n_estimators_max"]
-    optuna_storage = runtime_cfg["optuna_storage"]
-    optuna_study_prefix = runtime_cfg["optuna_study_prefix"]
-    best_params_files = runtime_cfg["best_params_files"]
-    plot_path_style = runtime_cfg["plot_path_style"]
-
-    model_names = build_model_names(
-        cfg["model_list"], cfg["model_categories"])
-    if not model_names:
-        raise ValueError(
-            "No model names generated from model_list/model_categories.")
-
-    results: Dict[str, ropt.BayesOptModel] = {}
-    trained_keys_by_model: Dict[str, List[str]] = {}
-
-    for model_name in model_names:
-        # Per-dataset training loop: load data, split train/test, and train requested models.
-        data_path = resolve_data_path(
-            data_dir,
-            model_name,
-            data_format=data_format,
-            path_template=data_path_template,
-        )
-        if not data_path.exists():
-            raise FileNotFoundError(f"Missing dataset: {data_path}")
-        data_fingerprint = {"path": str(data_path)}
-        if report_enabled and is_main_process:
-            data_fingerprint = fingerprint_file(
-                data_path,
-                max_bytes=data_fingerprint_max_bytes,
-            )
-
-        print(f"\n=== Processing model {model_name} ===")
-        raw = load_dataset(
-            data_path,
-            data_format=data_format,
-            dtype_map=dtype_map,
-            low_memory=False,
-        )
-        raw = coerce_dataset_types(raw)
-
-        train_df, test_df = split_train_test(
-            raw,
-            holdout_ratio=holdout_ratio,
-            strategy=split_strategy,
-            group_col=split_group_col,
-            time_col=split_time_col,
-            time_ascending=split_time_ascending,
-            rand_seed=rand_seed,
-            reset_index_mode="time_group",
-            ratio_label="holdout_ratio",
-        )
-
-        use_resn_dp = args.use_resn_dp or cfg.get(
-            "use_resn_data_parallel", False)
-        use_ft_dp = args.use_ft_dp or cfg.get("use_ft_data_parallel", True)
-        dataset_rows = len(raw)
-        ddp_enabled = bool(dist_active and (dataset_rows >= int(ddp_min_rows)))
-        use_resn_ddp = (args.use_resn_ddp or cfg.get(
-            "use_resn_ddp", False)) and ddp_enabled
-        use_ft_ddp = (args.use_ft_ddp or cfg.get(
-            "use_ft_ddp", False)) and ddp_enabled
-        use_gnn_dp = args.use_gnn_dp or cfg.get("use_gnn_data_parallel", False)
-        use_gnn_ddp = (args.use_gnn_ddp or cfg.get(
-            "use_gnn_ddp", False)) and ddp_enabled
-        gnn_use_ann = cfg.get("gnn_use_approx_knn", True)
-        if args.gnn_no_ann:
-            gnn_use_ann = False
-        gnn_threshold = args.gnn_ann_threshold if args.gnn_ann_threshold is not None else cfg.get(
-            "gnn_approx_knn_threshold", 50000)
-        gnn_graph_cache = args.gnn_graph_cache or cfg.get("gnn_graph_cache")
-        if isinstance(gnn_graph_cache, str) and gnn_graph_cache.strip():
-            resolved_cache = resolve_path(gnn_graph_cache, config_path.parent)
-            if resolved_cache is not None:
-                gnn_graph_cache = str(resolved_cache)
-        gnn_max_gpu_nodes = args.gnn_max_gpu_nodes if args.gnn_max_gpu_nodes is not None else cfg.get(
-            "gnn_max_gpu_knn_nodes", 200000)
-        gnn_gpu_mem_ratio = args.gnn_gpu_mem_ratio if args.gnn_gpu_mem_ratio is not None else cfg.get(
-            "gnn_knn_gpu_mem_ratio", 0.9)
-        gnn_gpu_mem_overhead = args.gnn_gpu_mem_overhead if args.gnn_gpu_mem_overhead is not None else cfg.get(
-            "gnn_knn_gpu_mem_overhead", 2.0)
-
-        binary_target = cfg.get("binary_target") or cfg.get("binary_resp_nme")
-        task_type = str(cfg.get("task_type", "regression"))
-        feature_list = cfg.get("feature_list")
-        categorical_features = cfg.get("categorical_features")
-        use_gpu = bool(cfg.get("use_gpu", True))
-        region_province_col = cfg.get("region_province_col")
-        region_city_col = cfg.get("region_city_col")
-        region_effect_alpha = cfg.get("region_effect_alpha")
-        geo_feature_nmes = cfg.get("geo_feature_nmes")
-        geo_token_hidden_dim = cfg.get("geo_token_hidden_dim")
-        geo_token_layers = cfg.get("geo_token_layers")
-        geo_token_dropout = cfg.get("geo_token_dropout")
-        geo_token_k_neighbors = cfg.get("geo_token_k_neighbors")
-        geo_token_learning_rate = cfg.get("geo_token_learning_rate")
-        geo_token_epochs = cfg.get("geo_token_epochs")
-
-        ft_role = args.ft_role or cfg.get("ft_role", "model")
-        if args.ft_as_feature and args.ft_role is None:
-            # Keep legacy behavior as a convenience alias only when the config
-            # didn't already request a non-default FT role.
-            if str(cfg.get("ft_role", "model")) == "model":
-                ft_role = "embedding"
-        ft_feature_prefix = str(
-            cfg.get("ft_feature_prefix", args.ft_feature_prefix))
-        ft_num_numeric_tokens = cfg.get("ft_num_numeric_tokens")
-
-        config_fields = getattr(ropt.BayesOptConfig,
-                                "__dataclass_fields__", {})
-        allowed_config_keys = set(config_fields.keys())
-        config_payload = {k: v for k,
-                          v in cfg.items() if k in allowed_config_keys}
-        config_payload.update({
-            "model_nme": model_name,
-            "resp_nme": cfg["target"],
-            "weight_nme": cfg["weight"],
-            "factor_nmes": feature_list,
-            "task_type": task_type,
-            "binary_resp_nme": binary_target,
-            "cate_list": categorical_features,
-            "prop_test": val_ratio,
-            "rand_seed": rand_seed,
-            "epochs": epochs,
-            "use_gpu": use_gpu,
-            "use_resn_data_parallel": use_resn_dp,
-            "use_ft_data_parallel": use_ft_dp,
-            "use_gnn_data_parallel": use_gnn_dp,
-            "use_resn_ddp": use_resn_ddp,
-            "use_ft_ddp": use_ft_ddp,
-            "use_gnn_ddp": use_gnn_ddp,
-            "output_dir": output_dir,
-            "xgb_max_depth_max": xgb_max_depth_max,
-            "xgb_n_estimators_max": xgb_n_estimators_max,
-            "resn_weight_decay": cfg.get("resn_weight_decay"),
-            "final_ensemble": bool(cfg.get("final_ensemble", False)),
-            "final_ensemble_k": int(cfg.get("final_ensemble_k", 3)),
-            "final_refit": bool(cfg.get("final_refit", True)),
-            "optuna_storage": optuna_storage,
-            "optuna_study_prefix": optuna_study_prefix,
-            "best_params_files": best_params_files,
-            "gnn_use_approx_knn": gnn_use_ann,
-            "gnn_approx_knn_threshold": gnn_threshold,
-            "gnn_graph_cache": gnn_graph_cache,
-            "gnn_max_gpu_knn_nodes": gnn_max_gpu_nodes,
-            "gnn_knn_gpu_mem_ratio": gnn_gpu_mem_ratio,
-            "gnn_knn_gpu_mem_overhead": gnn_gpu_mem_overhead,
-            "region_province_col": region_province_col,
-            "region_city_col": region_city_col,
-            "region_effect_alpha": region_effect_alpha,
-            "geo_feature_nmes": geo_feature_nmes,
-            "geo_token_hidden_dim": geo_token_hidden_dim,
-            "geo_token_layers": geo_token_layers,
-            "geo_token_dropout": geo_token_dropout,
-            "geo_token_k_neighbors": geo_token_k_neighbors,
-            "geo_token_learning_rate": geo_token_learning_rate,
-            "geo_token_epochs": geo_token_epochs,
-            "ft_role": ft_role,
-            "ft_feature_prefix": ft_feature_prefix,
-            "ft_num_numeric_tokens": ft_num_numeric_tokens,
-            "reuse_best_params": reuse_best_params,
-            "bo_sample_limit": bo_sample_limit,
-            "cache_predictions": cache_predictions,
-            "prediction_cache_dir": prediction_cache_dir,
-            "prediction_cache_format": prediction_cache_format,
-            "cv_strategy": cv_strategy or split_strategy,
-            "cv_group_col": cv_group_col or split_group_col,
-            "cv_time_col": cv_time_col or split_time_col,
-            "cv_time_ascending": cv_time_ascending,
-            "cv_splits": cv_splits,
-            "ft_oof_folds": ft_oof_folds,
-            "ft_oof_strategy": ft_oof_strategy,
-            "ft_oof_shuffle": ft_oof_shuffle,
-            "save_preprocess": save_preprocess,
-            "preprocess_artifact_path": preprocess_artifact_path,
-            "plot_path_style": plot_path_style or "nested",
-        })
-        config_payload = {
-            k: v for k, v in config_payload.items() if v is not None}
-        config = ropt.BayesOptConfig(**config_payload)
-        model = ropt.BayesOptModel(train_df, test_df, config=config)
-
-        if plot_requested:
-            plot_cfg = cfg.get("plot", {})
-            legacy_lift_flags = {
-                "glm": cfg.get("plot_lift_glm", False),
-                "xgb": cfg.get("plot_lift_xgb", False),
-                "resn": cfg.get("plot_lift_resn", False),
-                "ft": cfg.get("plot_lift_ft", False),
-            }
-            plot_enabled = plot_cfg.get(
-                "enable", any(legacy_lift_flags.values()))
-            if plot_enabled and plot_cfg.get("pre_oneway", False) and plot_cfg.get("oneway", True):
-                n_bins = int(plot_cfg.get("n_bins", 10))
-                model.plot_oneway(n_bins=n_bins, plot_subdir="oneway/pre")
-
-        if "all" in args.model_keys:
-            requested_keys = ["glm", "xgb", "resn", "ft", "gnn"]
-        else:
-            requested_keys = args.model_keys
-        requested_keys = dedupe_preserve_order(requested_keys)
-
-        if ft_role != "model":
-            requested_keys = [k for k in requested_keys if k != "ft"]
-            if not requested_keys:
-                stack_keys = args.stack_model_keys or cfg.get(
-                    "stack_model_keys")
-                if stack_keys:
-                    if "all" in stack_keys:
-                        requested_keys = ["glm", "xgb", "resn", "gnn"]
-                    else:
-                        requested_keys = [k for k in stack_keys if k != "ft"]
-                    requested_keys = dedupe_preserve_order(requested_keys)
-            if dist_active and ddp_enabled:
-                ft_trainer = model.trainers.get("ft")
-                if ft_trainer is None:
-                    raise ValueError("FT trainer is not available.")
-                ft_trainer_uses_ddp = bool(
-                    getattr(ft_trainer, "enable_distributed_optuna", False))
-                if not ft_trainer_uses_ddp:
-                    raise ValueError(
-                        "FT embedding under torchrun requires enabling FT DDP (use --use-ft-ddp or set use_ft_ddp=true)."
-                    )
-        missing = [key for key in requested_keys if key not in model.trainers]
-        if missing:
-            raise ValueError(
-                f"Trainer(s) {missing} not available for {model_name}")
-
-        executed_keys: List[str] = []
-        if ft_role != "model":
-            if dist_active and not ddp_enabled:
-                _ddp_barrier("start_ft_embedding")
-                if dist_rank != 0:
-                    _ddp_barrier("finish_ft_embedding")
-                    continue
-            print(
-                f"Optimizing ft as {ft_role} for {model_name} (max_evals={args.max_evals})")
-            model.optimize_model("ft", max_evals=args.max_evals)
-            model.trainers["ft"].save()
-            if getattr(ropt, "torch", None) is not None and ropt.torch.cuda.is_available():
-                ropt.free_cuda()
-            if dist_active and not ddp_enabled:
-                _ddp_barrier("finish_ft_embedding")
-        for key in requested_keys:
-            trainer = model.trainers[key]
-            trainer_uses_ddp = bool(
-                getattr(trainer, "enable_distributed_optuna", False))
-            if dist_active and not trainer_uses_ddp:
-                if dist_rank != 0:
-                    print(
-                        f"[Rank {dist_rank}] Skip {model_name}/{key} because trainer is not DDP-enabled."
-                    )
-                _ddp_barrier(f"start_non_ddp_{model_name}_{key}")
-                if dist_rank != 0:
-                    _ddp_barrier(f"finish_non_ddp_{model_name}_{key}")
-                    continue
-
-            print(
-                f"Optimizing {key} for {model_name} (max_evals={args.max_evals})")
-            model.optimize_model(key, max_evals=args.max_evals)
-            model.trainers[key].save()
-            _plot_loss_curve_for_trainer(model_name, model.trainers[key])
-            if key in PYTORCH_TRAINERS:
-                ropt.free_cuda()
-            if dist_active and not trainer_uses_ddp:
-                _ddp_barrier(f"finish_non_ddp_{model_name}_{key}")
-            executed_keys.append(key)
-
-        if not executed_keys:
-            continue
-
-        results[model_name] = model
-        trained_keys_by_model[model_name] = executed_keys
-        if report_enabled and is_main_process:
-            psi_report_df = _compute_psi_report(
-                model,
-                features=psi_features,
-                bins=psi_bins,
-                strategy=str(psi_strategy),
-            )
-            for key in executed_keys:
-                _evaluate_and_report(
-                    model,
-                    model_name=model_name,
-                    model_key=key,
-                    cfg=cfg,
-                    data_path=data_path,
-                    data_fingerprint=data_fingerprint,
-                    report_output_dir=report_output_dir,
-                    report_group_cols=report_group_cols,
-                    report_time_col=report_time_col,
-                    report_time_freq=str(report_time_freq),
-                    report_time_ascending=bool(report_time_ascending),
-                    psi_report_df=psi_report_df,
-                    calibration_cfg=calibration_cfg,
-                    threshold_cfg=threshold_cfg,
-                    bootstrap_cfg=bootstrap_cfg,
-                    register_model=register_model,
-                    registry_path=registry_path,
-                    registry_tags=registry_tags,
-                    registry_status=registry_status,
-                    run_id=run_id,
-                    config_sha=config_sha,
-                )
-
-    if not plot_requested:
-        return
-
-    for name, model in results.items():
-        _plot_curves_for_model(
-            model,
-            trained_keys_by_model.get(name, []),
-            cfg,
-        )
-
-
-def main() -> None:
-    if configure_run_logging:
-        configure_run_logging(prefix="bayesopt_entry")
-    args = _parse_args()
-    train_from_config(args)
-
-
-if __name__ == "__main__":
-    main()
+"""
+CLI entry point generated from BayesOpt_AutoPricing.ipynb so the workflow can
+run non‑interactively (e.g., via torchrun).
+
+Example:
+    python -m torch.distributed.run --standalone --nproc_per_node=2 \\
+        ins_pricing/cli/BayesOpt_entry.py \\
+        --config-json examples/config_template.json \\
+        --model-keys ft --max-evals 50 --use-ft-ddp
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+import sys
+
+if __package__ in {None, ""}:
+    repo_root = Path(__file__).resolve().parents[2]
+    if str(repo_root) not in sys.path:
+        sys.path.insert(0, str(repo_root))
+
+import argparse
+import hashlib
+import json
+import os
+from datetime import datetime
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+import pandas as pd
+
+# Use unified import resolver to eliminate nested try/except chains
+from .utils.import_resolver import resolve_imports, setup_sys_path
+from .utils.evaluation_context import (
+    EvaluationContext,
+    TrainingContext,
+    ModelIdentity,
+    DataFingerprint,
+    CalibrationConfig,
+    ThresholdConfig,
+    BootstrapConfig,
+    ReportConfig,
+    RegistryConfig,
+)
+
+# Resolve all imports from a single location
+setup_sys_path()
+_imports = resolve_imports()
+
+ropt = _imports.bayesopt
+PLOT_MODEL_LABELS = _imports.PLOT_MODEL_LABELS
+PYTORCH_TRAINERS = _imports.PYTORCH_TRAINERS
+build_model_names = _imports.build_model_names
+dedupe_preserve_order = _imports.dedupe_preserve_order
+load_dataset = _imports.load_dataset
+parse_model_pairs = _imports.parse_model_pairs
+resolve_data_path = _imports.resolve_data_path
+resolve_path = _imports.resolve_path
+fingerprint_file = _imports.fingerprint_file
+coerce_dataset_types = _imports.coerce_dataset_types
+split_train_test = _imports.split_train_test
+
+add_config_json_arg = _imports.add_config_json_arg
+add_output_dir_arg = _imports.add_output_dir_arg
+resolve_and_load_config = _imports.resolve_and_load_config
+resolve_data_config = _imports.resolve_data_config
+resolve_report_config = _imports.resolve_report_config
+resolve_split_config = _imports.resolve_split_config
+resolve_runtime_config = _imports.resolve_runtime_config
+resolve_output_dirs = _imports.resolve_output_dirs
+
+bootstrap_ci = _imports.bootstrap_ci
+calibrate_predictions = _imports.calibrate_predictions
+eval_metrics_report = _imports.metrics_report
+select_threshold = _imports.select_threshold
+
+ModelArtifact = _imports.ModelArtifact
+ModelRegistry = _imports.ModelRegistry
+drift_psi_report = _imports.drift_psi_report
+group_metrics = _imports.group_metrics
+ReportPayload = _imports.ReportPayload
+write_report = _imports.write_report
+
+configure_run_logging = _imports.configure_run_logging
+plot_loss_curve_common = _imports.plot_loss_curve
+
+import matplotlib
+
+if os.name != "nt" and not os.environ.get("DISPLAY") and not os.environ.get("MPLBACKEND"):
+    matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Batch trainer generated from BayesOpt_AutoPricing notebook."
+    )
+    add_config_json_arg(
+        parser,
+        help_text="Path to the JSON config describing datasets and feature columns.",
+    )
+    parser.add_argument(
+        "--model-keys",
+        nargs="+",
+        default=["ft"],
+        choices=["glm", "xgb", "resn", "ft", "gnn", "all"],
+        help="Space-separated list of trainers to run (e.g., --model-keys glm xgb). Include 'all' to run every trainer.",
+    )
+    parser.add_argument(
+        "--stack-model-keys",
+        nargs="+",
+        default=None,
+        choices=["glm", "xgb", "resn", "ft", "gnn", "all"],
+        help=(
+            "Only used when ft_role != 'model' (FT runs as feature generator). "
+            "When provided (or when config defines stack_model_keys), these trainers run after FT features "
+            "are generated. Use 'all' to run every non-FT trainer."
+        ),
+    )
+    parser.add_argument(
+        "--max-evals",
+        type=int,
+        default=50,
+        help="Optuna trial count per dataset.",
+    )
+    parser.add_argument(
+        "--use-resn-ddp",
+        action="store_true",
+        help="Force ResNet trainer to use DistributedDataParallel.",
+    )
+    parser.add_argument(
+        "--use-ft-ddp",
+        action="store_true",
+        help="Force FT-Transformer trainer to use DistributedDataParallel.",
+    )
+    parser.add_argument(
+        "--use-resn-dp",
+        action="store_true",
+        help="Enable ResNet DataParallel fall-back regardless of config.",
+    )
+    parser.add_argument(
+        "--use-ft-dp",
+        action="store_true",
+        help="Enable FT-Transformer DataParallel fall-back regardless of config.",
+    )
+    parser.add_argument(
+        "--use-gnn-dp",
+        action="store_true",
+        help="Enable GNN DataParallel fall-back regardless of config.",
+    )
+    parser.add_argument(
+        "--use-gnn-ddp",
+        action="store_true",
+        help="Force GNN trainer to use DistributedDataParallel.",
+    )
+    parser.add_argument(
+        "--gnn-no-ann",
+        action="store_true",
+        help="Disable approximate k-NN for GNN graph construction and use exact search.",
+    )
+    parser.add_argument(
+        "--gnn-ann-threshold",
+        type=int,
+        default=None,
+        help="Row threshold above which approximate k-NN is preferred (overrides config).",
+    )
+    parser.add_argument(
+        "--gnn-graph-cache",
+        default=None,
+        help="Optional path to persist/load cached adjacency matrix for GNN.",
+    )
+    parser.add_argument(
+        "--gnn-max-gpu-nodes",
+        type=int,
+        default=None,
+        help="Overrides the maximum node count allowed for GPU k-NN graph construction.",
+    )
+    parser.add_argument(
+        "--gnn-gpu-mem-ratio",
+        type=float,
+        default=None,
+        help="Overrides the fraction of free GPU memory the k-NN builder may consume.",
+    )
+    parser.add_argument(
+        "--gnn-gpu-mem-overhead",
+        type=float,
+        default=None,
+        help="Overrides the temporary GPU memory overhead multiplier for k-NN estimation.",
+    )
+    add_output_dir_arg(
+        parser,
+        help_text="Override output root for models/results/plots.",
+    )
+    parser.add_argument(
+        "--plot-curves",
+        action="store_true",
+        help="Enable lift/diagnostic plots after training (config file may also request plotting).",
+    )
+    parser.add_argument(
+        "--ft-as-feature",
+        action="store_true",
+        help="Alias for --ft-role embedding (keep tuning, export embeddings; skip FT plots/SHAP).",
+    )
+    parser.add_argument(
+        "--ft-role",
+        default=None,
+        choices=["model", "embedding", "unsupervised_embedding"],
+        help="How to use FT: model (default), embedding (export pooling embeddings), or unsupervised_embedding.",
+    )
+    parser.add_argument(
+        "--ft-feature-prefix",
+        default="ft_feat",
+        help="Prefix used for generated FT features (columns: pred_<prefix>_0.. or pred_<prefix>).",
+    )
+    parser.add_argument(
+        "--reuse-best-params",
+        action="store_true",
+        help="Skip Optuna and reuse best_params saved in Results/versions or bestparams CSV when available.",
+    )
+    return parser.parse_args()
+
+
+def _plot_curves_for_model(model: ropt.BayesOptModel, trained_keys: List[str], cfg: Dict) -> None:
+    plot_cfg = cfg.get("plot", {})
+    legacy_lift_flags = {
+        "glm": cfg.get("plot_lift_glm", False),
+        "xgb": cfg.get("plot_lift_xgb", False),
+        "resn": cfg.get("plot_lift_resn", False),
+        "ft": cfg.get("plot_lift_ft", False),
+    }
+    plot_enabled = plot_cfg.get("enable", any(legacy_lift_flags.values()))
+    if not plot_enabled:
+        return
+
+    n_bins = int(plot_cfg.get("n_bins", 10))
+    oneway_enabled = plot_cfg.get("oneway", True)
+
+    available_models = dedupe_preserve_order(
+        [m for m in trained_keys if m in PLOT_MODEL_LABELS]
+    )
+
+    lift_models = plot_cfg.get("lift_models")
+    if lift_models is None:
+        lift_models = [
+            m for m, enabled in legacy_lift_flags.items() if enabled]
+        if not lift_models:
+            lift_models = available_models
+    lift_models = dedupe_preserve_order(
+        [m for m in lift_models if m in available_models]
+    )
+
+    if oneway_enabled:
+        oneway_pred = bool(plot_cfg.get("oneway_pred", False))
+        oneway_pred_models = plot_cfg.get("oneway_pred_models")
+        pred_plotted = False
+        if oneway_pred:
+            if oneway_pred_models is None:
+                oneway_pred_models = lift_models or available_models
+            oneway_pred_models = dedupe_preserve_order(
+                [m for m in oneway_pred_models if m in available_models]
+            )
+            for model_key in oneway_pred_models:
+                label, pred_nme = PLOT_MODEL_LABELS[model_key]
+                if pred_nme not in model.train_data.columns:
+                    print(
+                        f"[Oneway] Missing prediction column '{pred_nme}'; skip.",
+                        flush=True,
+                    )
+                    continue
+                model.plot_oneway(
+                    n_bins=n_bins,
+                    pred_col=pred_nme,
+                    pred_label=label,
+                    plot_subdir="oneway/post",
+                )
+                pred_plotted = True
+        if not oneway_pred or not pred_plotted:
+            model.plot_oneway(n_bins=n_bins, plot_subdir="oneway/post")
+
+    if not available_models:
+        return
+
+    for model_key in lift_models:
+        label, pred_nme = PLOT_MODEL_LABELS[model_key]
+        model.plot_lift(model_label=label, pred_nme=pred_nme, n_bins=n_bins)
+
+    if not plot_cfg.get("double_lift", True) or len(available_models) < 2:
+        return
+
+    raw_pairs = plot_cfg.get("double_lift_pairs")
+    if raw_pairs:
+        pairs = [
+            (a, b)
+            for a, b in parse_model_pairs(raw_pairs)
+            if a in available_models and b in available_models and a != b
+        ]
+    else:
+        pairs = [(a, b) for i, a in enumerate(available_models)
+                 for b in available_models[i + 1:]]
+
+    for first, second in pairs:
+        model.plot_dlift([first, second], n_bins=n_bins)
+
+
+def _plot_loss_curve_for_trainer(model_name: str, trainer) -> None:
+    model_obj = getattr(trainer, "model", None)
+    history = None
+    if model_obj is not None:
+        history = getattr(model_obj, "training_history", None)
+    if not history:
+        history = getattr(trainer, "training_history", None)
+    if not history:
+        return
+    train_hist = list(history.get("train") or [])
+    val_hist = list(history.get("val") or [])
+    if not train_hist and not val_hist:
+        return
+    try:
+        plot_dir = trainer.output.plot_path(
+            f"{model_name}/loss/loss_{model_name}_{trainer.model_name_prefix}.png"
+        )
+    except Exception:
+        default_dir = Path("plot") / model_name / "loss"
+        default_dir.mkdir(parents=True, exist_ok=True)
+        plot_dir = str(
+            default_dir / f"loss_{model_name}_{trainer.model_name_prefix}.png")
+    if plot_loss_curve_common is not None:
+        plot_loss_curve_common(
+            history=history,
+            title=f"{trainer.model_name_prefix} Loss Curve ({model_name})",
+            save_path=plot_dir,
+            show=False,
+        )
+    else:
+        epochs = range(1, max(len(train_hist), len(val_hist)) + 1)
+        fig, ax = plt.subplots(figsize=(8, 4))
+        if train_hist:
+            ax.plot(range(1, len(train_hist) + 1),
+                    train_hist, label="Train Loss", color="tab:blue")
+        if val_hist:
+            ax.plot(range(1, len(val_hist) + 1),
+                    val_hist, label="Validation Loss", color="tab:orange")
+        ax.set_xlabel("Epoch")
+        ax.set_ylabel("Weighted Loss")
+        ax.set_title(
+            f"{trainer.model_name_prefix} Loss Curve ({model_name})")
+        ax.grid(True, linestyle="--", alpha=0.3)
+        ax.legend()
+        plt.tight_layout()
+        plt.savefig(plot_dir, dpi=300)
+        plt.close(fig)
+    print(
+        f"[Plot] Saved loss curve for {model_name}/{trainer.label} -> {plot_dir}")
+
+
+def _sample_arrays(
+    y_true: np.ndarray,
+    y_pred: np.ndarray,
+    *,
+    max_rows: Optional[int],
+    seed: Optional[int],
+) -> tuple[np.ndarray, np.ndarray]:
+    if max_rows is None or max_rows <= 0:
+        return y_true, y_pred
+    n = len(y_true)
+    if n <= max_rows:
+        return y_true, y_pred
+    rng = np.random.default_rng(seed)
+    idx = rng.choice(n, size=int(max_rows), replace=False)
+    return y_true[idx], y_pred[idx]
+
+
+def _compute_psi_report(
+    model: ropt.BayesOptModel,
+    *,
+    features: Optional[List[str]],
+    bins: int,
+    strategy: str,
+) -> Optional[pd.DataFrame]:
+    if drift_psi_report is None:
+        return None
+    psi_features = features or list(getattr(model, "factor_nmes", []))
+    psi_features = [
+        f for f in psi_features if f in model.train_data.columns and f in model.test_data.columns]
+    if not psi_features:
+        return None
+    try:
+        return drift_psi_report(
+            model.train_data[psi_features],
+            model.test_data[psi_features],
+            features=psi_features,
+            bins=int(bins),
+            strategy=str(strategy),
+        )
+    except Exception as exc:
+        print(f"[Report] PSI computation failed: {exc}")
+        return None
+
+
+# --- Refactored helper functions for _evaluate_and_report ---
+
+
+def _apply_calibration(
+    y_true_train: np.ndarray,
+    y_pred_train: np.ndarray,
+    y_pred_test: np.ndarray,
+    calibration_cfg: Dict[str, Any],
+    model_name: str,
+    model_key: str,
+) -> tuple[np.ndarray, np.ndarray, Optional[Dict[str, Any]]]:
+    """Apply calibration to predictions for classification tasks.
+
+    Returns:
+        Tuple of (calibrated_train_preds, calibrated_test_preds, calibration_info)
+    """
+    cal_cfg = dict(calibration_cfg or {})
+    cal_enabled = bool(cal_cfg.get("enable", False) or cal_cfg.get("method"))
+
+    if not cal_enabled or calibrate_predictions is None:
+        return y_pred_train, y_pred_test, None
+
+    method = cal_cfg.get("method", "sigmoid")
+    max_rows = cal_cfg.get("max_rows")
+    seed = cal_cfg.get("seed")
+    y_cal, p_cal = _sample_arrays(
+        y_true_train, y_pred_train, max_rows=max_rows, seed=seed)
+
+    try:
+        calibrator = calibrate_predictions(y_cal, p_cal, method=method)
+        calibrated_train = calibrator.predict(y_pred_train)
+        calibrated_test = calibrator.predict(y_pred_test)
+        calibration_info = {"method": calibrator.method, "max_rows": max_rows}
+        return calibrated_train, calibrated_test, calibration_info
+    except Exception as exc:
+        print(f"[Report] Calibration failed for {model_name}/{model_key}: {exc}")
+        return y_pred_train, y_pred_test, None
+
+
+def _select_classification_threshold(
+    y_true_train: np.ndarray,
+    y_pred_train_eval: np.ndarray,
+    threshold_cfg: Dict[str, Any],
+) -> tuple[float, Optional[Dict[str, Any]]]:
+    """Select threshold for classification predictions.
+
+    Returns:
+        Tuple of (threshold_value, threshold_info)
+    """
+    thr_cfg = dict(threshold_cfg or {})
+    thr_enabled = bool(
+        thr_cfg.get("enable", False)
+        or thr_cfg.get("metric")
+        or thr_cfg.get("value") is not None
+    )
+
+    if thr_cfg.get("value") is not None:
+        threshold_value = float(thr_cfg["value"])
+        return threshold_value, {"threshold": threshold_value, "source": "fixed"}
+
+    if thr_enabled and select_threshold is not None:
+        max_rows = thr_cfg.get("max_rows")
+        seed = thr_cfg.get("seed")
+        y_thr, p_thr = _sample_arrays(
+            y_true_train, y_pred_train_eval, max_rows=max_rows, seed=seed)
+        threshold_info = select_threshold(
+            y_thr,
+            p_thr,
+            metric=thr_cfg.get("metric", "f1"),
+            min_positive_rate=thr_cfg.get("min_positive_rate"),
+            grid=thr_cfg.get("grid", 99),
+        )
+        return float(threshold_info.get("threshold", 0.5)), threshold_info
+
+    return 0.5, None
+
+
+def _compute_classification_metrics(
+    y_true_test: np.ndarray,
+    y_pred_test_eval: np.ndarray,
+    threshold_value: float,
+) -> Dict[str, Any]:
+    """Compute metrics for classification task."""
+    metrics = eval_metrics_report(
+        y_true_test,
+        y_pred_test_eval,
+        task_type="classification",
+        threshold=threshold_value,
+    )
+    precision = float(metrics.get("precision", 0.0))
+    recall = float(metrics.get("recall", 0.0))
+    f1 = 0.0 if (precision + recall) == 0 else 2 * precision * recall / (precision + recall)
+    metrics["f1"] = float(f1)
+    metrics["threshold"] = float(threshold_value)
+    return metrics
+
+
+def _compute_bootstrap_ci(
+    y_true_test: np.ndarray,
+    y_pred_test_eval: np.ndarray,
+    weight_test: Optional[np.ndarray],
+    metrics: Dict[str, Any],
+    bootstrap_cfg: Dict[str, Any],
+    task_type: str,
+) -> Dict[str, Dict[str, float]]:
+    """Compute bootstrap confidence intervals for metrics."""
+    if not bootstrap_cfg or not bool(bootstrap_cfg.get("enable", False)) or bootstrap_ci is None:
+        return {}
+
+    metric_names = bootstrap_cfg.get("metrics")
+    if not metric_names:
+        metric_names = [name for name in metrics.keys() if name != "threshold"]
+    n_samples = int(bootstrap_cfg.get("n_samples", 200))
+    ci = float(bootstrap_cfg.get("ci", 0.95))
+    seed = bootstrap_cfg.get("seed")
+
+    def _metric_fn(y_true, y_pred, weight=None):
+        vals = eval_metrics_report(
+            y_true,
+            y_pred,
+            task_type=task_type,
+            weight=weight,
+            threshold=metrics.get("threshold", 0.5),
+        )
+        if task_type == "classification":
+            prec = float(vals.get("precision", 0.0))
+            rec = float(vals.get("recall", 0.0))
+            vals["f1"] = 0.0 if (prec + rec) == 0 else 2 * prec * rec / (prec + rec)
+        return vals
+
+    bootstrap_results: Dict[str, Dict[str, float]] = {}
+    for name in metric_names:
+        if name not in metrics:
+            continue
+        ci_result = bootstrap_ci(
+            lambda y_t, y_p, w=None: float(_metric_fn(y_t, y_p, w).get(name, 0.0)),
+            y_true_test,
+            y_pred_test_eval,
+            weight=weight_test,
+            n_samples=n_samples,
+            ci=ci,
+            seed=seed,
+        )
+        bootstrap_results[str(name)] = ci_result
+
+    return bootstrap_results
+
+
+def _compute_validation_table(
+    model: ropt.BayesOptModel,
+    pred_col: str,
+    report_group_cols: Optional[List[str]],
+    weight_col: Optional[str],
+    model_name: str,
+    model_key: str,
+) -> Optional[pd.DataFrame]:
+    """Compute grouped validation metrics table."""
+    if not report_group_cols or group_metrics is None:
+        return None
+
+    available_groups = [
+        col for col in report_group_cols if col in model.test_data.columns
+    ]
+    if not available_groups:
+        return None
+
+    try:
+        validation_table = group_metrics(
+            model.test_data,
+            actual_col=model.resp_nme,
+            pred_col=pred_col,
+            group_cols=available_groups,
+            weight_col=weight_col if weight_col and weight_col in model.test_data.columns else None,
+        )
+        counts = (
+            model.test_data.groupby(available_groups, dropna=False)
+            .size()
+            .reset_index(name="count")
+        )
+        return validation_table.merge(counts, on=available_groups, how="left")
+    except Exception as exc:
+        print(f"[Report] group_metrics failed for {model_name}/{model_key}: {exc}")
+        return None
+
+
+def _compute_risk_trend(
+    model: ropt.BayesOptModel,
+    pred_col: str,
+    report_time_col: Optional[str],
+    report_time_freq: str,
+    report_time_ascending: bool,
+    weight_col: Optional[str],
+    model_name: str,
+    model_key: str,
+) -> Optional[pd.DataFrame]:
+    """Compute time-series risk trend metrics."""
+    if not report_time_col or group_metrics is None:
+        return None
+
+    if report_time_col not in model.test_data.columns:
+        return None
+
+    try:
+        time_df = model.test_data.copy()
+        time_series = pd.to_datetime(time_df[report_time_col], errors="coerce")
+        time_df = time_df.loc[time_series.notna()].copy()
+
+        if time_df.empty:
+            return None
+
+        time_df["_time_bucket"] = (
+            pd.to_datetime(time_df[report_time_col], errors="coerce")
+            .dt.to_period(report_time_freq)
+            .dt.to_timestamp()
+        )
+        risk_trend = group_metrics(
+            time_df,
+            actual_col=model.resp_nme,
+            pred_col=pred_col,
+            group_cols=["_time_bucket"],
+            weight_col=weight_col if weight_col and weight_col in time_df.columns else None,
+        )
+        counts = (
+            time_df.groupby("_time_bucket", dropna=False)
+            .size()
+            .reset_index(name="count")
+        )
+        risk_trend = risk_trend.merge(counts, on="_time_bucket", how="left")
+        risk_trend = risk_trend.sort_values(
+            "_time_bucket", ascending=bool(report_time_ascending)
+        ).reset_index(drop=True)
+        return risk_trend.rename(columns={"_time_bucket": report_time_col})
+    except Exception as exc:
+        print(f"[Report] time metrics failed for {model_name}/{model_key}: {exc}")
+        return None
+
+
+def _write_metrics_json(
+    report_root: Path,
+    model_name: str,
+    model_key: str,
+    version: str,
+    metrics: Dict[str, Any],
+    threshold_info: Optional[Dict[str, Any]],
+    calibration_info: Optional[Dict[str, Any]],
+    bootstrap_results: Dict[str, Dict[str, float]],
+    data_path: Path,
+    data_fingerprint: Dict[str, Any],
+    config_sha: str,
+    pred_col: str,
+    task_type: str,
+) -> Path:
+    """Write metrics to JSON file and return the path."""
+    metrics_payload = {
+        "model_name": model_name,
+        "model_key": model_key,
+        "model_version": version,
+        "metrics": metrics,
+        "threshold": threshold_info,
+        "calibration": calibration_info,
+        "bootstrap": bootstrap_results,
+        "data_path": str(data_path),
+        "data_fingerprint": data_fingerprint,
+        "config_sha256": config_sha,
+        "pred_col": pred_col,
+        "task_type": task_type,
+    }
+    metrics_path = report_root / f"{model_name}_{model_key}_metrics.json"
+    metrics_path.write_text(
+        json.dumps(metrics_payload, indent=2, ensure_ascii=True),
+        encoding="utf-8",
+    )
+    return metrics_path
+
+
+def _write_model_report(
+    report_root: Path,
+    model_name: str,
+    model_key: str,
+    version: str,
+    metrics: Dict[str, Any],
+    risk_trend: Optional[pd.DataFrame],
+    psi_report_df: Optional[pd.DataFrame],
+    validation_table: Optional[pd.DataFrame],
+    calibration_info: Optional[Dict[str, Any]],
+    threshold_info: Optional[Dict[str, Any]],
+    bootstrap_results: Dict[str, Dict[str, float]],
+    config_sha: str,
+    data_fingerprint: Dict[str, Any],
+) -> Optional[Path]:
+    """Write model report and return the path."""
+    if ReportPayload is None or write_report is None:
+        return None
+
+    notes_lines = [
+        f"- Config SHA256: {config_sha}",
+        f"- Data fingerprint: {data_fingerprint.get('sha256_prefix')}",
+    ]
+    if calibration_info:
+        notes_lines.append(f"- Calibration: {calibration_info.get('method')}")
+    if threshold_info:
+        notes_lines.append(f"- Threshold selection: {threshold_info}")
+    if bootstrap_results:
+        notes_lines.append("- Bootstrap: see metrics JSON for CI")
+
+    payload = ReportPayload(
+        model_name=f"{model_name}/{model_key}",
+        model_version=version,
+        metrics={k: float(v) for k, v in metrics.items()},
+        risk_trend=risk_trend,
+        drift_report=psi_report_df,
+        validation_table=validation_table,
+        extra_notes="\n".join(notes_lines),
+    )
+    return write_report(
+        payload,
+        report_root / f"{model_name}_{model_key}_report.md",
+    )
+
+
+def _register_model_to_registry(
+    model: ropt.BayesOptModel,
+    model_name: str,
+    model_key: str,
+    version: str,
+    metrics: Dict[str, Any],
+    task_type: str,
+    data_path: Path,
+    data_fingerprint: Dict[str, Any],
+    config_sha: str,
+    registry_path: Optional[str],
+    registry_tags: Dict[str, Any],
+    registry_status: str,
+    report_path: Optional[Path],
+    metrics_path: Path,
+    cfg: Dict[str, Any],
+) -> None:
+    """Register model artifacts to the model registry."""
+    if ModelRegistry is None or ModelArtifact is None:
+        return
+
+    registry = ModelRegistry(
+        registry_path
+        if registry_path
+        else Path(model.output_manager.result_dir) / "model_registry.json"
+    )
+
+    tags = {str(k): str(v) for k, v in (registry_tags or {}).items()}
+    tags.update({
+        "model_key": str(model_key),
+        "task_type": str(task_type),
+        "data_path": str(data_path),
+        "data_sha256_prefix": str(data_fingerprint.get("sha256_prefix", "")),
+        "data_size": str(data_fingerprint.get("size", "")),
+        "data_mtime": str(data_fingerprint.get("mtime", "")),
+        "config_sha256": str(config_sha),
+    })
+
+    artifacts = _collect_model_artifacts(
+        model, model_name, model_key, report_path, metrics_path, cfg
+    )
+
+    registry.register(
+        name=str(model_name),
+        version=version,
+        metrics={k: float(v) for k, v in metrics.items()},
+        tags=tags,
+        artifacts=artifacts,
+        status=str(registry_status or "candidate"),
+        notes=f"model_key={model_key}",
+    )
+
+
+def _collect_model_artifacts(
+    model: ropt.BayesOptModel,
+    model_name: str,
+    model_key: str,
+    report_path: Optional[Path],
+    metrics_path: Path,
+    cfg: Dict[str, Any],
+) -> List:
+    """Collect all model artifacts for registry."""
+    artifacts = []
+
+    # Trained model artifact
+    trainer = model.trainers.get(model_key)
+    if trainer is not None:
+        try:
+            model_path = trainer.output.model_path(trainer._get_model_filename())
+            if os.path.exists(model_path):
+                artifacts.append(ModelArtifact(path=model_path, description="trained model"))
+        except Exception:
+            pass
+
+    # Report artifact
+    if report_path is not None:
+        artifacts.append(ModelArtifact(path=str(report_path), description="model report"))
+
+    # Metrics JSON artifact
+    if metrics_path.exists():
+        artifacts.append(ModelArtifact(path=str(metrics_path), description="metrics json"))
+
+    # Preprocess artifacts
+    if bool(cfg.get("save_preprocess", False)):
+        artifact_path = cfg.get("preprocess_artifact_path")
+        if artifact_path:
+            preprocess_path = Path(str(artifact_path))
+            if not preprocess_path.is_absolute():
+                preprocess_path = Path(model.output_manager.result_dir) / preprocess_path
+        else:
+            preprocess_path = Path(model.output_manager.result_path(
+                f"{model.model_nme}_preprocess.json"
+            ))
+        if preprocess_path.exists():
+            artifacts.append(
+                ModelArtifact(path=str(preprocess_path), description="preprocess artifacts")
+            )
+
+    # Prediction cache artifacts
+    if bool(cfg.get("cache_predictions", False)):
+        cache_dir = cfg.get("prediction_cache_dir")
+        if cache_dir:
+            pred_root = Path(str(cache_dir))
+            if not pred_root.is_absolute():
+                pred_root = Path(model.output_manager.result_dir) / pred_root
+        else:
+            pred_root = Path(model.output_manager.result_dir) / "predictions"
+        ext = "csv" if str(cfg.get("prediction_cache_format", "parquet")).lower() == "csv" else "parquet"
+        for split_label in ("train", "test"):
+            pred_path = pred_root / f"{model_name}_{model_key}_{split_label}.{ext}"
+            if pred_path.exists():
+                artifacts.append(
+                    ModelArtifact(path=str(pred_path), description=f"predictions {split_label}")
+                )
+
+    return artifacts
+
+
+def _evaluate_and_report(
+    model: ropt.BayesOptModel,
+    *,
+    model_name: str,
+    model_key: str,
+    cfg: Dict[str, Any],
+    data_path: Path,
+    data_fingerprint: Dict[str, Any],
+    report_output_dir: Optional[str],
+    report_group_cols: Optional[List[str]],
+    report_time_col: Optional[str],
+    report_time_freq: str,
+    report_time_ascending: bool,
+    psi_report_df: Optional[pd.DataFrame],
+    calibration_cfg: Dict[str, Any],
+    threshold_cfg: Dict[str, Any],
+    bootstrap_cfg: Dict[str, Any],
+    register_model: bool,
+    registry_path: Optional[str],
+    registry_tags: Dict[str, Any],
+    registry_status: str,
+    run_id: str,
+    config_sha: str,
+) -> None:
+    """Evaluate model predictions and generate reports.
+
+    This function orchestrates the evaluation pipeline:
+    1. Extract predictions and ground truth
+    2. Apply calibration (for classification)
+    3. Select threshold (for classification)
+    4. Compute metrics
+    5. Compute bootstrap confidence intervals
+    6. Generate validation tables and risk trends
+    7. Write reports and register model
+    """
+    if eval_metrics_report is None:
+        print("[Report] Skip evaluation: metrics module unavailable.")
+        return
+
+    pred_col = PLOT_MODEL_LABELS.get(model_key, (None, f"pred_{model_key}"))[1]
+    if pred_col not in model.test_data.columns:
+        print(f"[Report] Missing prediction column '{pred_col}' for {model_name}/{model_key}; skip.")
+        return
+
+    # Extract predictions and weights
+    weight_col = getattr(model, "weight_nme", None)
+    y_true_train = model.train_data[model.resp_nme].to_numpy(dtype=float, copy=False)
+    y_true_test = model.test_data[model.resp_nme].to_numpy(dtype=float, copy=False)
+    y_pred_train = model.train_data[pred_col].to_numpy(dtype=float, copy=False)
+    y_pred_test = model.test_data[pred_col].to_numpy(dtype=float, copy=False)
+    weight_test = (
+        model.test_data[weight_col].to_numpy(dtype=float, copy=False)
+        if weight_col and weight_col in model.test_data.columns
+        else None
+    )
+
+    task_type = str(cfg.get("task_type", getattr(model, "task_type", "regression")))
+
+    # Process based on task type
+    if task_type == "classification":
+        y_pred_train = np.clip(y_pred_train, 0.0, 1.0)
+        y_pred_test = np.clip(y_pred_test, 0.0, 1.0)
+
+        y_pred_train_eval, y_pred_test_eval, calibration_info = _apply_calibration(
+            y_true_train, y_pred_train, y_pred_test, calibration_cfg, model_name, model_key
+        )
+        threshold_value, threshold_info = _select_classification_threshold(
+            y_true_train, y_pred_train_eval, threshold_cfg
+        )
+        metrics = _compute_classification_metrics(y_true_test, y_pred_test_eval, threshold_value)
+    else:
+        y_pred_test_eval = y_pred_test
+        calibration_info = None
+        threshold_info = None
+        metrics = eval_metrics_report(
+            y_true_test, y_pred_test_eval, task_type=task_type, weight=weight_test
+        )
+
+    # Compute bootstrap confidence intervals
+    bootstrap_results = _compute_bootstrap_ci(
+        y_true_test, y_pred_test_eval, weight_test, metrics, bootstrap_cfg, task_type
+    )
+
+    # Compute validation table and risk trend
+    validation_table = _compute_validation_table(
+        model, pred_col, report_group_cols, weight_col, model_name, model_key
+    )
+    risk_trend = _compute_risk_trend(
+        model, pred_col, report_time_col, report_time_freq,
+        report_time_ascending, weight_col, model_name, model_key
+    )
+
+    # Setup output directory
+    report_root = (
+        Path(report_output_dir)
+        if report_output_dir
+        else Path(model.output_manager.result_dir) / "reports"
+    )
+    report_root.mkdir(parents=True, exist_ok=True)
+    version = f"{model_key}_{run_id}"
+
+    # Write metrics JSON
+    metrics_path = _write_metrics_json(
+        report_root, model_name, model_key, version, metrics,
+        threshold_info, calibration_info, bootstrap_results,
+        data_path, data_fingerprint, config_sha, pred_col, task_type
+    )
+
+    # Write model report
+    report_path = _write_model_report(
+        report_root, model_name, model_key, version, metrics,
+        risk_trend, psi_report_df, validation_table,
+        calibration_info, threshold_info, bootstrap_results,
+        config_sha, data_fingerprint
+    )
+
+    # Register model
+    if register_model:
+        _register_model_to_registry(
+            model, model_name, model_key, version, metrics, task_type,
+            data_path, data_fingerprint, config_sha, registry_path,
+            registry_tags, registry_status, report_path, metrics_path, cfg
+        )
+
+
+def _evaluate_with_context(
+    model: ropt.BayesOptModel,
+    ctx: EvaluationContext,
+) -> None:
+    """Evaluate model predictions using context object.
+
+    This is a cleaner interface that uses the EvaluationContext dataclass
+    instead of 19+ individual parameters.
+    """
+    _evaluate_and_report(
+        model,
+        model_name=ctx.identity.model_name,
+        model_key=ctx.identity.model_key,
+        cfg=ctx.cfg,
+        data_path=ctx.data_path,
+        data_fingerprint=ctx.data_fingerprint.to_dict(),
+        report_output_dir=ctx.report.output_dir,
+        report_group_cols=ctx.report.group_cols,
+        report_time_col=ctx.report.time_col,
+        report_time_freq=ctx.report.time_freq,
+        report_time_ascending=ctx.report.time_ascending,
+        psi_report_df=ctx.psi_report_df,
+        calibration_cfg={
+            "enable": ctx.calibration.enable,
+            "method": ctx.calibration.method,
+            "max_rows": ctx.calibration.max_rows,
+            "seed": ctx.calibration.seed,
+        },
+        threshold_cfg={
+            "enable": ctx.threshold.enable,
+            "metric": ctx.threshold.metric,
+            "value": ctx.threshold.value,
+            "min_positive_rate": ctx.threshold.min_positive_rate,
+            "grid": ctx.threshold.grid,
+            "max_rows": ctx.threshold.max_rows,
+            "seed": ctx.threshold.seed,
+        },
+        bootstrap_cfg={
+            "enable": ctx.bootstrap.enable,
+            "metrics": ctx.bootstrap.metrics,
+            "n_samples": ctx.bootstrap.n_samples,
+            "ci": ctx.bootstrap.ci,
+            "seed": ctx.bootstrap.seed,
+        },
+        register_model=ctx.registry.register,
+        registry_path=ctx.registry.path,
+        registry_tags=ctx.registry.tags,
+        registry_status=ctx.registry.status,
+        run_id=ctx.run_id,
+        config_sha=ctx.config_sha,
+    )
+
+
+def _create_ddp_barrier(dist_ctx: TrainingContext):
+    """Create a DDP barrier function for distributed training synchronization."""
+    def _ddp_barrier(reason: str) -> None:
+        if not dist_ctx.is_distributed:
+            return
+        torch_mod = getattr(ropt, "torch", None)
+        dist_mod = getattr(torch_mod, "distributed", None)
+        if dist_mod is None:
+            return
+        try:
+            if not getattr(dist_mod, "is_available", lambda: False)():
+                return
+            if not dist_mod.is_initialized():
+                ddp_ok, _, _, _ = ropt.DistributedUtils.setup_ddp()
+                if not ddp_ok or not dist_mod.is_initialized():
+                    return
+            dist_mod.barrier()
+        except Exception as exc:
+            print(f"[DDP] barrier failed during {reason}: {exc}", flush=True)
+            raise
+    return _ddp_barrier
+
+
+def train_from_config(args: argparse.Namespace) -> None:
+    script_dir = Path(__file__).resolve().parents[1]
+    config_path, cfg = resolve_and_load_config(
+        args.config_json,
+        script_dir,
+        required_keys=["data_dir", "model_list",
+                       "model_categories", "target", "weight"],
+    )
+    plot_requested = bool(args.plot_curves or cfg.get("plot_curves", False))
+    config_sha = hashlib.sha256(config_path.read_bytes()).hexdigest()
+    run_id = datetime.utcnow().strftime("%Y%m%d_%H%M%S")
+
+    # Use TrainingContext for distributed training state
+    dist_ctx = TrainingContext.from_env()
+    dist_world_size = dist_ctx.world_size
+    dist_rank = dist_ctx.rank
+    dist_active = dist_ctx.is_distributed
+    is_main_process = dist_ctx.is_main_process
+    _ddp_barrier = _create_ddp_barrier(dist_ctx)
+
+    data_dir, data_format, data_path_template, dtype_map = resolve_data_config(
+        cfg,
+        config_path,
+        create_data_dir=True,
+    )
+    runtime_cfg = resolve_runtime_config(cfg)
+    ddp_min_rows = runtime_cfg["ddp_min_rows"]
+    bo_sample_limit = runtime_cfg["bo_sample_limit"]
+    cache_predictions = runtime_cfg["cache_predictions"]
+    prediction_cache_dir = runtime_cfg["prediction_cache_dir"]
+    prediction_cache_format = runtime_cfg["prediction_cache_format"]
+    report_cfg = resolve_report_config(cfg)
+    report_output_dir = report_cfg["report_output_dir"]
+    report_group_cols = report_cfg["report_group_cols"]
+    report_time_col = report_cfg["report_time_col"]
+    report_time_freq = report_cfg["report_time_freq"]
+    report_time_ascending = report_cfg["report_time_ascending"]
+    psi_bins = report_cfg["psi_bins"]
+    psi_strategy = report_cfg["psi_strategy"]
+    psi_features = report_cfg["psi_features"]
+    calibration_cfg = report_cfg["calibration_cfg"]
+    threshold_cfg = report_cfg["threshold_cfg"]
+    bootstrap_cfg = report_cfg["bootstrap_cfg"]
+    register_model = report_cfg["register_model"]
+    registry_path = report_cfg["registry_path"]
+    registry_tags = report_cfg["registry_tags"]
+    registry_status = report_cfg["registry_status"]
+    data_fingerprint_max_bytes = report_cfg["data_fingerprint_max_bytes"]
+    report_enabled = report_cfg["report_enabled"]
+
+    split_cfg = resolve_split_config(cfg)
+    prop_test = split_cfg["prop_test"]
+    holdout_ratio = split_cfg["holdout_ratio"]
+    val_ratio = split_cfg["val_ratio"]
+    split_strategy = split_cfg["split_strategy"]
+    split_group_col = split_cfg["split_group_col"]
+    split_time_col = split_cfg["split_time_col"]
+    split_time_ascending = split_cfg["split_time_ascending"]
+    cv_strategy = split_cfg["cv_strategy"]
+    cv_group_col = split_cfg["cv_group_col"]
+    cv_time_col = split_cfg["cv_time_col"]
+    cv_time_ascending = split_cfg["cv_time_ascending"]
+    cv_splits = split_cfg["cv_splits"]
+    ft_oof_folds = split_cfg["ft_oof_folds"]
+    ft_oof_strategy = split_cfg["ft_oof_strategy"]
+    ft_oof_shuffle = split_cfg["ft_oof_shuffle"]
+    save_preprocess = runtime_cfg["save_preprocess"]
+    preprocess_artifact_path = runtime_cfg["preprocess_artifact_path"]
+    rand_seed = runtime_cfg["rand_seed"]
+    epochs = runtime_cfg["epochs"]
+    output_cfg = resolve_output_dirs(
+        cfg,
+        config_path,
+        output_override=args.output_dir,
+    )
+    output_dir = output_cfg["output_dir"]
+    reuse_best_params = bool(
+        args.reuse_best_params or runtime_cfg["reuse_best_params"])
+    xgb_max_depth_max = runtime_cfg["xgb_max_depth_max"]
+    xgb_n_estimators_max = runtime_cfg["xgb_n_estimators_max"]
+    optuna_storage = runtime_cfg["optuna_storage"]
+    optuna_study_prefix = runtime_cfg["optuna_study_prefix"]
+    best_params_files = runtime_cfg["best_params_files"]
+    plot_path_style = runtime_cfg["plot_path_style"]
+
+    model_names = build_model_names(
+        cfg["model_list"], cfg["model_categories"])
+    if not model_names:
+        raise ValueError(
+            "No model names generated from model_list/model_categories.")
+
+    results: Dict[str, ropt.BayesOptModel] = {}
+    trained_keys_by_model: Dict[str, List[str]] = {}
+
+    for model_name in model_names:
+        # Per-dataset training loop: load data, split train/test, and train requested models.
+        data_path = resolve_data_path(
+            data_dir,
+            model_name,
+            data_format=data_format,
+            path_template=data_path_template,
+        )
+        if not data_path.exists():
+            raise FileNotFoundError(f"Missing dataset: {data_path}")
+        data_fingerprint = {"path": str(data_path)}
+        if report_enabled and is_main_process:
+            data_fingerprint = fingerprint_file(
+                data_path,
+                max_bytes=data_fingerprint_max_bytes,
+            )
+
+        print(f"\n=== Processing model {model_name} ===")
+        raw = load_dataset(
+            data_path,
+            data_format=data_format,
+            dtype_map=dtype_map,
+            low_memory=False,
+        )
+        raw = coerce_dataset_types(raw)
+
+        train_df, test_df = split_train_test(
+            raw,
+            holdout_ratio=holdout_ratio,
+            strategy=split_strategy,
+            group_col=split_group_col,
+            time_col=split_time_col,
+            time_ascending=split_time_ascending,
+            rand_seed=rand_seed,
+            reset_index_mode="time_group",
+            ratio_label="holdout_ratio",
+        )
+
+        use_resn_dp = args.use_resn_dp or cfg.get(
+            "use_resn_data_parallel", False)
+        use_ft_dp = args.use_ft_dp or cfg.get("use_ft_data_parallel", True)
+        dataset_rows = len(raw)
+        ddp_enabled = bool(dist_active and (dataset_rows >= int(ddp_min_rows)))
+        use_resn_ddp = (args.use_resn_ddp or cfg.get(
+            "use_resn_ddp", False)) and ddp_enabled
+        use_ft_ddp = (args.use_ft_ddp or cfg.get(
+            "use_ft_ddp", False)) and ddp_enabled
+        use_gnn_dp = args.use_gnn_dp or cfg.get("use_gnn_data_parallel", False)
+        use_gnn_ddp = (args.use_gnn_ddp or cfg.get(
+            "use_gnn_ddp", False)) and ddp_enabled
+        gnn_use_ann = cfg.get("gnn_use_approx_knn", True)
+        if args.gnn_no_ann:
+            gnn_use_ann = False
+        gnn_threshold = args.gnn_ann_threshold if args.gnn_ann_threshold is not None else cfg.get(
+            "gnn_approx_knn_threshold", 50000)
+        gnn_graph_cache = args.gnn_graph_cache or cfg.get("gnn_graph_cache")
+        if isinstance(gnn_graph_cache, str) and gnn_graph_cache.strip():
+            resolved_cache = resolve_path(gnn_graph_cache, config_path.parent)
+            if resolved_cache is not None:
+                gnn_graph_cache = str(resolved_cache)
+        gnn_max_gpu_nodes = args.gnn_max_gpu_nodes if args.gnn_max_gpu_nodes is not None else cfg.get(
+            "gnn_max_gpu_knn_nodes", 200000)
+        gnn_gpu_mem_ratio = args.gnn_gpu_mem_ratio if args.gnn_gpu_mem_ratio is not None else cfg.get(
+            "gnn_knn_gpu_mem_ratio", 0.9)
+        gnn_gpu_mem_overhead = args.gnn_gpu_mem_overhead if args.gnn_gpu_mem_overhead is not None else cfg.get(
+            "gnn_knn_gpu_mem_overhead", 2.0)
+
+        binary_target = cfg.get("binary_target") or cfg.get("binary_resp_nme")
+        task_type = str(cfg.get("task_type", "regression"))
+        feature_list = cfg.get("feature_list")
+        categorical_features = cfg.get("categorical_features")
+        use_gpu = bool(cfg.get("use_gpu", True))
+        region_province_col = cfg.get("region_province_col")
+        region_city_col = cfg.get("region_city_col")
+        region_effect_alpha = cfg.get("region_effect_alpha")
+        geo_feature_nmes = cfg.get("geo_feature_nmes")
+        geo_token_hidden_dim = cfg.get("geo_token_hidden_dim")
+        geo_token_layers = cfg.get("geo_token_layers")
+        geo_token_dropout = cfg.get("geo_token_dropout")
+        geo_token_k_neighbors = cfg.get("geo_token_k_neighbors")
+        geo_token_learning_rate = cfg.get("geo_token_learning_rate")
+        geo_token_epochs = cfg.get("geo_token_epochs")
+
+        ft_role = args.ft_role or cfg.get("ft_role", "model")
+        if args.ft_as_feature and args.ft_role is None:
+            # Keep legacy behavior as a convenience alias only when the config
+            # didn't already request a non-default FT role.
+            if str(cfg.get("ft_role", "model")) == "model":
+                ft_role = "embedding"
+        ft_feature_prefix = str(
+            cfg.get("ft_feature_prefix", args.ft_feature_prefix))
+        ft_num_numeric_tokens = cfg.get("ft_num_numeric_tokens")
+
+        config_fields = getattr(ropt.BayesOptConfig,
+                                "__dataclass_fields__", {})
+        allowed_config_keys = set(config_fields.keys())
+        config_payload = {k: v for k,
+                          v in cfg.items() if k in allowed_config_keys}
+        config_payload.update({
+            "model_nme": model_name,
+            "resp_nme": cfg["target"],
+            "weight_nme": cfg["weight"],
+            "factor_nmes": feature_list,
+            "task_type": task_type,
+            "binary_resp_nme": binary_target,
+            "cate_list": categorical_features,
+            "prop_test": val_ratio,
+            "rand_seed": rand_seed,
+            "epochs": epochs,
+            "use_gpu": use_gpu,
+            "use_resn_data_parallel": use_resn_dp,
+            "use_ft_data_parallel": use_ft_dp,
+            "use_gnn_data_parallel": use_gnn_dp,
+            "use_resn_ddp": use_resn_ddp,
+            "use_ft_ddp": use_ft_ddp,
+            "use_gnn_ddp": use_gnn_ddp,
+            "output_dir": output_dir,
+            "xgb_max_depth_max": xgb_max_depth_max,
+            "xgb_n_estimators_max": xgb_n_estimators_max,
+            "resn_weight_decay": cfg.get("resn_weight_decay"),
+            "final_ensemble": bool(cfg.get("final_ensemble", False)),
+            "final_ensemble_k": int(cfg.get("final_ensemble_k", 3)),
+            "final_refit": bool(cfg.get("final_refit", True)),
+            "optuna_storage": optuna_storage,
+            "optuna_study_prefix": optuna_study_prefix,
+            "best_params_files": best_params_files,
+            "gnn_use_approx_knn": gnn_use_ann,
+            "gnn_approx_knn_threshold": gnn_threshold,
+            "gnn_graph_cache": gnn_graph_cache,
+            "gnn_max_gpu_knn_nodes": gnn_max_gpu_nodes,
+            "gnn_knn_gpu_mem_ratio": gnn_gpu_mem_ratio,
+            "gnn_knn_gpu_mem_overhead": gnn_gpu_mem_overhead,
+            "region_province_col": region_province_col,
+            "region_city_col": region_city_col,
+            "region_effect_alpha": region_effect_alpha,
+            "geo_feature_nmes": geo_feature_nmes,
+            "geo_token_hidden_dim": geo_token_hidden_dim,
+            "geo_token_layers": geo_token_layers,
+            "geo_token_dropout": geo_token_dropout,
+            "geo_token_k_neighbors": geo_token_k_neighbors,
+            "geo_token_learning_rate": geo_token_learning_rate,
+            "geo_token_epochs": geo_token_epochs,
+            "ft_role": ft_role,
+            "ft_feature_prefix": ft_feature_prefix,
+            "ft_num_numeric_tokens": ft_num_numeric_tokens,
+            "reuse_best_params": reuse_best_params,
+            "bo_sample_limit": bo_sample_limit,
+            "cache_predictions": cache_predictions,
+            "prediction_cache_dir": prediction_cache_dir,
+            "prediction_cache_format": prediction_cache_format,
+            "cv_strategy": cv_strategy or split_strategy,
+            "cv_group_col": cv_group_col or split_group_col,
+            "cv_time_col": cv_time_col or split_time_col,
+            "cv_time_ascending": cv_time_ascending,
+            "cv_splits": cv_splits,
+            "ft_oof_folds": ft_oof_folds,
+            "ft_oof_strategy": ft_oof_strategy,
+            "ft_oof_shuffle": ft_oof_shuffle,
+            "save_preprocess": save_preprocess,
+            "preprocess_artifact_path": preprocess_artifact_path,
+            "plot_path_style": plot_path_style or "nested",
+        })
+        config_payload = {
+            k: v for k, v in config_payload.items() if v is not None}
+        config = ropt.BayesOptConfig(**config_payload)
+        model = ropt.BayesOptModel(train_df, test_df, config=config)
+
+        if plot_requested:
+            plot_cfg = cfg.get("plot", {})
+            legacy_lift_flags = {
+                "glm": cfg.get("plot_lift_glm", False),
+                "xgb": cfg.get("plot_lift_xgb", False),
+                "resn": cfg.get("plot_lift_resn", False),
+                "ft": cfg.get("plot_lift_ft", False),
+            }
+            plot_enabled = plot_cfg.get(
+                "enable", any(legacy_lift_flags.values()))
+            if plot_enabled and plot_cfg.get("pre_oneway", False) and plot_cfg.get("oneway", True):
+                n_bins = int(plot_cfg.get("n_bins", 10))
+                model.plot_oneway(n_bins=n_bins, plot_subdir="oneway/pre")
+
+        if "all" in args.model_keys:
+            requested_keys = ["glm", "xgb", "resn", "ft", "gnn"]
+        else:
+            requested_keys = args.model_keys
+        requested_keys = dedupe_preserve_order(requested_keys)
+
+        if ft_role != "model":
+            requested_keys = [k for k in requested_keys if k != "ft"]
+            if not requested_keys:
+                stack_keys = args.stack_model_keys or cfg.get(
+                    "stack_model_keys")
+                if stack_keys:
+                    if "all" in stack_keys:
+                        requested_keys = ["glm", "xgb", "resn", "gnn"]
+                    else:
+                        requested_keys = [k for k in stack_keys if k != "ft"]
+                    requested_keys = dedupe_preserve_order(requested_keys)
+            if dist_active and ddp_enabled:
+                ft_trainer = model.trainers.get("ft")
+                if ft_trainer is None:
+                    raise ValueError("FT trainer is not available.")
+                ft_trainer_uses_ddp = bool(
+                    getattr(ft_trainer, "enable_distributed_optuna", False))
+                if not ft_trainer_uses_ddp:
+                    raise ValueError(
+                        "FT embedding under torchrun requires enabling FT DDP (use --use-ft-ddp or set use_ft_ddp=true)."
+                    )
+        missing = [key for key in requested_keys if key not in model.trainers]
+        if missing:
+            raise ValueError(
+                f"Trainer(s) {missing} not available for {model_name}")
+
+        executed_keys: List[str] = []
+        if ft_role != "model":
+            if dist_active and not ddp_enabled:
+                _ddp_barrier("start_ft_embedding")
+                if dist_rank != 0:
+                    _ddp_barrier("finish_ft_embedding")
+                    continue
+            print(
+                f"Optimizing ft as {ft_role} for {model_name} (max_evals={args.max_evals})")
+            model.optimize_model("ft", max_evals=args.max_evals)
+            model.trainers["ft"].save()
+            if getattr(ropt, "torch", None) is not None and ropt.torch.cuda.is_available():
+                ropt.free_cuda()
+            if dist_active and not ddp_enabled:
+                _ddp_barrier("finish_ft_embedding")
+        for key in requested_keys:
+            trainer = model.trainers[key]
+            trainer_uses_ddp = bool(
+                getattr(trainer, "enable_distributed_optuna", False))
+            if dist_active and not trainer_uses_ddp:
+                if dist_rank != 0:
+                    print(
+                        f"[Rank {dist_rank}] Skip {model_name}/{key} because trainer is not DDP-enabled."
+                    )
+                _ddp_barrier(f"start_non_ddp_{model_name}_{key}")
+                if dist_rank != 0:
+                    _ddp_barrier(f"finish_non_ddp_{model_name}_{key}")
+                    continue
+
+            print(
+                f"Optimizing {key} for {model_name} (max_evals={args.max_evals})")
+            model.optimize_model(key, max_evals=args.max_evals)
+            model.trainers[key].save()
+            _plot_loss_curve_for_trainer(model_name, model.trainers[key])
+            if key in PYTORCH_TRAINERS:
+                ropt.free_cuda()
+            if dist_active and not trainer_uses_ddp:
+                _ddp_barrier(f"finish_non_ddp_{model_name}_{key}")
+            executed_keys.append(key)
+
+        if not executed_keys:
+            continue
+
+        results[model_name] = model
+        trained_keys_by_model[model_name] = executed_keys
+        if report_enabled and is_main_process:
+            psi_report_df = _compute_psi_report(
+                model,
+                features=psi_features,
+                bins=psi_bins,
+                strategy=str(psi_strategy),
+            )
+            for key in executed_keys:
+                _evaluate_and_report(
+                    model,
+                    model_name=model_name,
+                    model_key=key,
+                    cfg=cfg,
+                    data_path=data_path,
+                    data_fingerprint=data_fingerprint,
+                    report_output_dir=report_output_dir,
+                    report_group_cols=report_group_cols,
+                    report_time_col=report_time_col,
+                    report_time_freq=str(report_time_freq),
+                    report_time_ascending=bool(report_time_ascending),
+                    psi_report_df=psi_report_df,
+                    calibration_cfg=calibration_cfg,
+                    threshold_cfg=threshold_cfg,
+                    bootstrap_cfg=bootstrap_cfg,
+                    register_model=register_model,
+                    registry_path=registry_path,
+                    registry_tags=registry_tags,
+                    registry_status=registry_status,
+                    run_id=run_id,
+                    config_sha=config_sha,
+                )
+
+    if not plot_requested:
+        return
+
+    for name, model in results.items():
+        _plot_curves_for_model(
+            model,
+            trained_keys_by_model.get(name, []),
+            cfg,
+        )
+
+
+def main() -> None:
+    if configure_run_logging:
+        configure_run_logging(prefix="bayesopt_entry")
+    args = _parse_args()
+    train_from_config(args)
+
+
+if __name__ == "__main__":
+    main()