fix(trainer): 检测训练续训以配置为准，并恢复学习率调度进度

train.py --resume 走的是与 SimCLR 不同的一套代码，同样的问题在这里更严重：

- load_checkpoint 直接 optimizer.load_state_dict，把 lr/weight_decay 连同动量缓冲
  一起还原成存档值：yaml 写 lr=5e-4，实际仍按检查点的 1e-3 训
- save_checkpoint 根本没保存调度器状态，续训时 scheduler.last_epoch 从 0 开始，
  而 optimizer 的 lr 又来自检查点，两者互相矛盾，cosine 曲线完全错乱

改为与 SimCLR 一致的「配置文件是唯一真相源」：

- save_checkpoint 补上 scheduler 状态、global_step 与当时的 train 配置快照
- 新增 Trainer.resume()：加载优化器状态后按组写回配置里的超参（只留动量缓冲，
  并保住 BN/bias 那组 weight_decay=0 的结构），调度器按当前配置重建再快进到已训 epoch
- 配置与检查点不一致时逐项打日志说明按当前配置执行
- train.py 改用 trainer.resume()，不再自己拼恢复逻辑

实测续训到 10 轮时 lr=3.273e-04，与 cosine 在 4/10 处的理论值精确一致；
weight_decay、T_max、调度进度、动量缓冲均正确。

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

tests/test_train_smoke.py

@@ -13,6 +13,7 @@ pytest.importorskip("yaml")
 pytest.importorskip("PIL")
 
 import numpy as np  # noqa: E402
+import torch  # noqa: E402
 import yaml  # noqa: E402
 from PIL import Image  # noqa: E402
 
@@ -150,3 +151,67 @@ def test_device_flag_overrides_config(project, monkeypatch, tmp_path):
 
 def config_path_of(project):
     return project[0]
+
+
+def test_resume_takes_hyperparams_from_config(project, monkeypatch, tmp_path):
+    """检测训练续训：lr / weight_decay / 总轮数以 yaml 为准，检查点只提供进度。
+
+    回归测试：load_checkpoint 直接 optimizer.load_state_dict，把超参连同动量一起还原，
+    改了配置再续训会被静默覆盖；而且此前根本没保存调度器状态，学习率调度会从头开始。
+    """
+    import math
+
+    import yaml as yaml_mod
+
+    config_path, root = project
+    cfg = yaml_mod.safe_load(config_path.read_text(encoding="utf-8"))
+    cfg["train"].update({"epochs": 4, "optimizer": {"name": "adamw", "lr": 1e-3, "weight_decay": 1e-4},
+                         "scheduler": {"name": "cosine"}})
+    config_path.write_text(yaml_mod.safe_dump(cfg, allow_unicode=True), encoding="utf-8")
+
+    work_dir = tmp_path / "runs"
+    assert run(train_mod, ["train.py", "--config", str(config_path), "--work-dir", str(work_dir)],
+               monkeypatch) == 0
+
+    captured = {}
+    real_fit = train_mod.Trainer.fit
+
+    def probe(self):
+        captured["groups"] = [dict(g) for g in self.optimizer.param_groups]
+        captured["scheduler"] = self.scheduler
+        captured["start_epoch"] = self.start_epoch
+        return 0.0
+
+    monkeypatch.setattr(train_mod.Trainer, "fit", probe)
+    assert run(
+        train_mod,
+        ["train.py", "--config", str(config_path), "--work-dir", str(work_dir),
+         "--resume", str(work_dir / "last.pth"), "--set",
+         "train.epochs=10", "train.optimizer.lr=5e-4", "train.optimizer.weight_decay=0.05"],
+        monkeypatch,
+    ) == 0
+    monkeypatch.setattr(train_mod.Trainer, "fit", real_fit)
+
+    groups = captured["groups"]
+    assert groups[0]["initial_lr"] == pytest.approx(5e-4)     # 不是检查点里的 1e-3
+    assert groups[0]["weight_decay"] == pytest.approx(0.05)   # 不是检查点里的 1e-4
+    assert groups[1]["weight_decay"] == 0.0                   # BN/bias 分组结构没被冲掉
+
+    scheduler = captured["scheduler"]
+    assert scheduler.T_max == 10
+    assert scheduler.last_epoch == captured["start_epoch"] == 4  # 调度进度已快进，不是从 0 开始
+    assert groups[0]["lr"] == pytest.approx(5e-4 * (1 + math.cos(math.pi * 4 / 10)) / 2)
+
+
+def test_checkpoint_carries_scheduler_and_progress(project, monkeypatch, tmp_path):
+    """检查点要带上调度器状态与 global_step，否则续训无从恢复调度进度。"""
+    config_path, _ = project
+    work_dir = tmp_path / "runs"
+    assert run(train_mod, ["train.py", "--config", str(config_path), "--work-dir", str(work_dir)],
+               monkeypatch) == 0
+
+    state = torch.load(work_dir / "last.pth", map_location="cpu", weights_only=False)
+    expected = {"model", "optimizer", "scheduler", "epoch", "global_step", "best_metric", "train_cfg"}
+    assert expected <= set(state)
+    assert state["global_step"] > 0
+    assert state["train_cfg"]["optimizer"]["lr"] is not None