test: 单元测试与端到端冒烟

88 个测试覆盖配置继承、注册表、GPU 守卫、YOLO 格式解析、数据集、模型形状、
标签模式、聚类原语与 SimCLR 损失，以及 train.py / eval.py / 聚类脚本的入口接线。

含两条回归测试：
- PCA 白化放大噪声方向导致簇塌陷（默认 whiten=False 的依据）
- classes 写成 data.yaml 路径时 num_classes 必须是类别数而非路径字符串长度

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_train_smoke.py

@@ -0,0 +1,137 @@
+"""train.py / eval.py 的端到端冒烟：YOLO 数据目录 -> 训练 -> 评估。
+
+覆盖入口脚本的接线（classes 解析、label_mode 下发、GPU 守卫），
+所以在 CPU 上跑，`require_cuda` 用 monkeypatch 放行。
+"""
+
+import sys
+
+import pytest
+
+pytest.importorskip("torch")
+pytest.importorskip("yaml")
+pytest.importorskip("PIL")
+
+import numpy as np  # noqa: E402
+import yaml  # noqa: E402
+from PIL import Image  # noqa: E402
+
+import eval as eval_mod  # noqa: E402
+import train as train_mod  # noqa: E402
+from datasets import write_data_yaml  # noqa: E402
+from utils.device import NoNvidiaGpuError  # noqa: E402
+
+CLASSES = ["scratch", "bubble"]
+TRANSFORMS = [
+    {"name": "Resize", "size": [32, 32]},
+    {"name": "ToTensor"},
+    {"name": "Normalize", "mean": [0.5, 0.5, 0.5], "std": [0.5, 0.5, 0.5]},
+]
+
+
+@pytest.fixture
+def project(tmp_path):
+    """造一个最小的 YOLO 数据集 + 配置，classes 用 data.yaml 的路径给。"""
+    for split in ("train", "val"):
+        images = tmp_path / "images" / split
+        labels = tmp_path / "labels" / split
+        images.mkdir(parents=True)
+        labels.mkdir(parents=True)
+        for i in range(2):
+            Image.fromarray(np.zeros((32, 32, 3), dtype=np.uint8)).save(images / f"{i}.png")
+            content = "0 0.5 0.5 0.25 0.25\n" if i else ""
+            (labels / f"{i}.txt").write_text(content, encoding="utf-8")
+
+    data_yaml = write_data_yaml(tmp_path / "data.yaml", CLASSES, root=tmp_path)
+
+    def dataset(split):
+        return {
+            "dataset": {"name": "film_defect", "img_dir": str(tmp_path / "images" / split),
+                        "transforms": TRANSFORMS},
+            "loader": {"batch_size": 2, "num_workers": 0, "shuffle": split == "train"},
+        }
+
+    config = {
+        "classes": str(data_yaml),
+        "label_mode": "multilabel",
+        "image_size": [32, 32],
+        "device": "cpu",
+        "seed": 0,
+        "model": {
+            "name": "defect_net",
+            "backbone": {"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]},
+            "neck": {"name": "upsample", "channels": 8, "num_upsamples": 1},
+            "head": {"hidden_channels": 8},
+        },
+        "data": {"train": dataset("train"), "val": dataset("val")},
+        "train": {
+            "epochs": 1,
+            "amp": False,
+            "log_interval": 10,
+            "optimizer": {"name": "adamw", "lr": 0.001},
+            "swanlab": {"enabled": False},
+        },
+    }
+    config_path = tmp_path / "config.yaml"
+    config_path.write_text(yaml.safe_dump(config, allow_unicode=True), encoding="utf-8")
+    return config_path, tmp_path
+
+
+def run(module, argv, monkeypatch):
+    monkeypatch.setattr(sys, "argv", argv)
+    return module.main()
+
+
+def test_train_then_eval(project, monkeypatch, tmp_path):
+    config_path, root = project
+    work_dir = tmp_path / "runs"
+    monkeypatch.setattr(train_mod, "require_cuda", lambda: ["FAKE-GPU"])
+
+    assert run(train_mod, ["train.py", "--config", str(config_path), "--work-dir", str(work_dir)],
+               monkeypatch) == 0
+    assert (work_dir / "best.pth").exists()
+    assert (work_dir / "config.yaml").exists()
+
+    metrics_path = tmp_path / "metrics.json"
+    assert run(
+        eval_mod,
+        ["eval.py", "--config", str(config_path), "--checkpoint", str(work_dir / "best.pth"),
+         "--set", "device=cpu", "--out", str(metrics_path)],
+        monkeypatch,
+    ) == 0
+    assert metrics_path.exists()
+
+
+def test_classes_from_data_yaml_sizes_the_heads(project, monkeypatch, tmp_path):
+    """classes 写成 data.yaml 路径时，num_classes 必须是 2，而不是路径字符串的长度。"""
+    config_path, _ = project
+    work_dir = tmp_path / "runs2"
+    captured = {}
+
+    real_build = train_mod.build_model
+
+    def spy(cfg):
+        captured["cfg"] = dict(cfg)
+        return real_build(cfg)
+
+    monkeypatch.setattr(train_mod, "build_model", spy)
+    monkeypatch.setattr(train_mod, "require_cuda", lambda: ["FAKE-GPU"])
+
+    run(train_mod, ["train.py", "--config", str(config_path), "--work-dir", str(work_dir)], monkeypatch)
+
+    assert captured["cfg"]["num_classes"] == len(CLASSES)
+    assert captured["cfg"]["label_mode"] == "multilabel"
+
+
+def test_training_is_blocked_without_nvidia_gpu(project, monkeypatch, tmp_path):
+    config_path, _ = project
+
+    def refuse():
+        raise NoNvidiaGpuError("没有 Nvidia GPU")
+
+    monkeypatch.setattr(train_mod, "require_cuda", refuse)
+    assert run(
+        train_mod,
+        ["train.py", "--config", str(config_path), "--work-dir", str(tmp_path / "runs3")],
+        monkeypatch,
+    ) == 1