fix(backbones): 修复 shufflenet 无法构建，并让换骨架不再踩配置合并的坑

- shufflenet_v2_x1_0 列在支持清单里却直接 AttributeError：ShuffleNetV2 既不是
  扁平 children 也没有 .features，需要按 conv1/maxpool/stage2..4/conv5 自己拼。
  该 bug 同时影响 SimCLR 与检测模型
- 新增参数化测试覆盖清单里的每个 arch：能前向、声明的 out_channels 与实际一致
- 配置支持 `_replace_: true`：子字典整块替换而不与 base 合并。换骨架时
  tiny_conv 的 widths/depths 会残留并传给新骨架，此前无论用 --set 还是 _base_
  都绕不开，只能改原文件
- Registry.build 提前校验字段，报错直接指出是哪些字段不被接受、该怎么办，
  取代裸 TypeError
- README 补充可选骨架表与三种改法（含不该用 --set 改 name 的说明）

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

tests/test_models.py

@@ -3,6 +3,7 @@ import pytest
 torch = pytest.importorskip("torch")
 
 from models import build_model  # noqa: E402
+from models.backbones.resnet import _SUPPORTED as _SUPPORTED_ARCHS  # noqa: E402
 from models.postprocess import decode_heatmap  # noqa: E402
 
 
@@ -90,3 +91,57 @@ def test_losses_decrease_on_a_single_batch(model_cfg):
         last = total
 
     assert last < first
+
+
+@pytest.mark.parametrize("arch", sorted(_SUPPORTED_ARCHS))
+def test_every_supported_torchvision_arch_actually_works(arch):
+    """清单里列出的每个骨架都必须真能前向，且声明的 out_channels 与实际一致。
+
+    回归测试：shufflenet_v2_x1_0 曾被列为支持，但它没有 `.features` 属性，
+    构造时直接 AttributeError。
+    """
+    from models.registry import BACKBONES
+
+    backbone = BACKBONES.build({"name": "torchvision", "arch": arch, "pretrained": False}).eval()
+    with torch.no_grad():
+        feat = backbone(torch.randn(1, 3, 64, 64))
+
+    assert feat.dim() == 4
+    assert feat.shape[1] == backbone.out_channels == _SUPPORTED_ARCHS[arch]
+    assert backbone.out_stride == 32
+
+
+def test_unsupported_arch_is_rejected():
+    from models.registry import BACKBONES
+
+    with pytest.raises(ValueError, match="不支持的骨架"):
+        BACKBONES.build({"name": "torchvision", "arch": "vit_b_16"})
+
+
+@pytest.mark.parametrize("arch", ["resnet18", "shufflenet_v2_x1_0", "mobilenet_v3_small"])
+def test_alternative_backbones_drive_simclr(arch):
+    """换骨架后 SimCLR 的投影头要能接上（out_channels 必须准）。"""
+    from models.heads.projection import ProjectionHead
+    from models.registry import BACKBONES
+
+    backbone = BACKBONES.build({"name": "torchvision", "arch": arch, "pretrained": False}).eval()
+    head = ProjectionHead(backbone.out_channels, 32, 16)
+    with torch.no_grad():
+        assert head(backbone(torch.randn(2, 3, 64, 64))).shape == (2, 16)
+
+
+def test_alternative_backbone_drives_defect_net():
+    """检测模型换骨架同样要能跑通（neck 的 in_channels 由 backbone 决定）。"""
+    model = build_model(
+        {
+            "name": "defect_net",
+            "num_classes": 3,
+            "backbone": {"name": "torchvision", "arch": "shufflenet_v2_x1_0", "pretrained": False},
+            "neck": {"name": "upsample", "channels": 16, "num_upsamples": 3},
+            "head": {"hidden_channels": 16},
+        }
+    ).eval()
+    assert model.out_stride == 4
+    with torch.no_grad():
+        out = model(torch.randn(1, 3, 64, 64))
+    assert out["heatmap"].shape == (1, 3, 16, 16)