# 续写同一文件：training_project/src/readingtorch/engine.py
def evaluate(checkpoint, data):
    # 选定 best 后单独读取测试集，训练循环里没有查询测试指标。
    saved = load_checkpoint(checkpoint)
    task = saved["task"]
    bundle = load_bundle(task, data, saved["seed"])
    if saved["source_hash"] != bundle.source_hash:
        raise ValueError("评估数据与训练记录不一致")
    model = build_model(task)
    model.load_state_dict(saved["model"])
    loss_fn = nn.MSELoss() if task == "regression" else nn.CrossEntropyLoss()
    scale = bundle.preprocess["y_scale"].item() if task == "regression" else 1.0
    metrics = one_epoch(
        model, make_loader(bundle, "test"), loss_fn, torch.device("cpu"), task, scale
    )
    if task == "regression":
        # 标准化后的 0 对应训练目标均值，这就是不训练模型的基线。
        metrics["baseline_rmse"] = (
            bundle.y[bundle.splits["test"]].square().mean().sqrt().item() * scale
        )
    else:
        metrics["baseline_accuracy"] = 1 / 3
    return {
        "task": task,
        "checkpoint_epoch": saved["epoch"],
        "samples": len(bundle.splits["test"]),
        **metrics,
    }