fix(simclr): 续训时改变总轮数不再让学习率卡在 0

resume 会恢复调度器状态，其中包含存档时的 T_max。原计划 10 轮的 cosine 跑完后
学习率已降到 0，此时把 epochs 改成 30 接着练，学习率仍是 0——训练照常进行、
日志照常打印，但一步都学不动。是静默失败，比报错更难发现。

- 总轮数与存档一致时行为不变，直接恢复调度器状态
- 不一致时按新的总轮数重建调度器，并快进到当前 epoch；重建前先把 lr 复位回
  initial_lr，因为 cosine 的链式实现基于当前 lr 递推，从 0 出发只会一路留在 0
- 实测续训到 30 轮时 lr 恢复为 7.50e-04，与 cosine 在 10/30 处的理论值一致
- 补三个回归测试：改变总轮数后的学习率、总轮数不变时直接恢复状态、
  save_interval 落的中间盘可直接续训
- 修正一个旧测试：它断言续训后 lr 仍等于旧计划末尾的 0，编码的是错误行为

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

tests/test_simclr.py

@@ -239,7 +239,6 @@ def test_resume_continues_from_the_saved_epoch(tiny_image_dir, tmp_path):
     first = _make_trainer(tiny_image_dir, work_dir, epochs=2)
     first.fit()
     steps_after_2 = first.global_step
-    lr_after_2 = first.optimizer.param_groups[0]["lr"]
 
     resumed = SimCLRTrainer(
         BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
@@ -253,7 +252,8 @@ def test_resume_continues_from_the_saved_epoch(tiny_image_dir, tmp_path):
     )
     assert resumed.start_epoch == 2                    # 从第 3 轮开始
     assert resumed.global_step == steps_after_2        # step 计数接上
-    assert resumed.optimizer.param_groups[0]["lr"] == pytest.approx(lr_after_2)  # 学习率进度接上
+    # 总轮数从 2 改成了 5，学习率按新计划重算（若沿用旧的 T_max，此处会是 0）
+    assert resumed.optimizer.param_groups[0]["lr"] > 1e-5
 
     resumed.fit()
     state = torch.load(work_dir / "simclr_last.pth", map_location="cpu", weights_only=False)
@@ -261,6 +261,84 @@ def test_resume_continues_from_the_saved_epoch(tiny_image_dir, tmp_path):
     assert resumed.global_step > steps_after_2
 
 
+def test_resume_with_a_longer_schedule_rebuilds_the_lr_curve(tiny_image_dir, tmp_path):
+    """续训时调大 epochs，学习率必须按新的总轮数重算。
+
+    回归测试：此前直接恢复调度器状态，会沿用存档里的 T_max。原计划 10 轮的 cosine
+    跑完后 lr 已经是 0，这时把 epochs 改成 30 接着练，lr 仍是 0——训练照跑但一步都学不动。
+    """
+    import math
+
+    work_dir = tmp_path / "run"
+    first = _make_trainer(tiny_image_dir, work_dir, epochs=10)
+    first.fit()
+    assert first.optimizer.param_groups[0]["lr"] == pytest.approx(0.0, abs=1e-9)  # cosine 已走完
+
+    resumed = SimCLRTrainer(
+        BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
+        first.dataloader,
+        {"epochs": 30, "amp": False, "log_interval": 999, "save_interval": 1,
+         "scheduler": {"name": "cosine"}, "optimizer": {"name": "adamw", "lr": 1e-3},
+         "projection": {"hidden_channels": 16, "out_channels": 8},
+         "swanlab": {"enabled": False}},
+        torch.device("cpu"),
+        work_dir,
+        resume=work_dir / "simclr_last.pth",
+    )
+
+    assert resumed.scheduler.T_max == 30
+    lr = resumed.optimizer.param_groups[0]["lr"]
+    assert lr > 1e-5, "续训后学习率仍是 0，等于白跑"
+    assert lr == pytest.approx(1e-3 * (1 + math.cos(math.pi * 10 / 30)) / 2, rel=1e-3)
+
+
+def test_resume_with_same_schedule_restores_state(tiny_image_dir, tmp_path):
+    """总轮数不变时应当直接恢复调度器状态，不重建。"""
+    work_dir = tmp_path / "run"
+    _make_trainer(tiny_image_dir, work_dir, epochs=6).fit()
+
+    resumed = SimCLRTrainer(
+        BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
+        _make_trainer(tiny_image_dir, work_dir, epochs=1).dataloader,
+        {"epochs": 6, "amp": False, "log_interval": 999, "save_interval": 1,
+         "scheduler": {"name": "cosine"},
+         "projection": {"hidden_channels": 16, "out_channels": 8},
+         "swanlab": {"enabled": False}},
+        torch.device("cpu"),
+        work_dir,
+        resume=work_dir / "simclr_last.pth",
+    )
+    assert resumed.scheduler.T_max == 6
+    assert resumed.scheduler.last_epoch == 6
+
+
+def test_intermediate_checkpoint_is_resumable(tiny_image_dir, tmp_path):
+    """save_interval 落的中间盘必须能直接拿来续训（不必等训练正常结束）。"""
+    work_dir = tmp_path / "run"
+    trainer = _make_trainer(tiny_image_dir, work_dir, epochs=10)
+    # 只手动跑到第 4 轮就落盘，模拟被 kill 掉
+    for epoch in range(4):
+        trainer.train_one_epoch(epoch)
+        trainer.scheduler.step()
+    trainer._checkpoint(3)
+
+    resumed = SimCLRTrainer(
+        BACKBONES.build({"name": "tiny_conv", "widths": [8, 16], "depths": [1, 1]}),
+        trainer.dataloader,
+        {"epochs": 10, "amp": False, "log_interval": 999, "save_interval": 1,
+         "scheduler": {"name": "cosine"},
+         "projection": {"hidden_channels": 16, "out_channels": 8},
+         "swanlab": {"enabled": False}},
+        torch.device("cpu"),
+        work_dir,
+        resume=work_dir / "simclr_last.pth",
+    )
+    assert resumed.start_epoch == 4
+    resumed.fit()
+    state = torch.load(work_dir / "simclr_last.pth", map_location="cpu", weights_only=False)
+    assert state["epoch"] == 9
+
+
 def test_resume_rejects_backbone_only_file(tiny_image_dir, tmp_path):
     """simclr_backbone.pth 只有骨架，不能当续训状态用，要给出可操作的报错。"""
     work_dir = tmp_path / "run"