Give KPlanes a linear head when use_hybrid_mlp is off

src/quantem/core/ml/models/kplanes.py

@@ -243,6 +243,12 @@ def __init__(
             nn.init.zeros_(out.bias)
             layers.append(out)
             self.sigma_net = nn.Sequential(*layers)
+        else:
+            # Linear head fallback, matching KPlanesTILTED._build_sigma_net and
+            # CPTilted: forward/get_params reference sigma_net unconditionally.
+            self.sigma_net = nn.Linear(self.feature_dim, 1, bias=True)
+            nn.init.normal_(self.sigma_net.weight, std=0.01)
+            nn.init.zeros_(self.sigma_net.bias)
 
     def get_densities(self, coords: torch.Tensor):
         """Computes and returns densities"""

tests/ml/test_kplanes.py

@@ -18,3 +18,24 @@ def test_anisotropic_raises(self):
             KPlanes(M_features=2, resolution=(16, 16, 8))
         with pytest.raises(ValueError, match="isotropic"):
             KPlanesTILTED(M_features=2, T=2, resolution=(16, 8, 16))
+
+
+class TestDefaultHeadConstruction:
+    """Regression: KPlanes(use_hybrid_mlp=False) -- the constructor default --
+    built no sigma_net at all, so forward / get_params / ObjectTensorDecomp
+    .from_model crashed with AttributeError. KPlanesTILTED and CPTilted both
+    fall back to a linear head; plain KPlanes must do the same."""
+
+    def test_default_get_params(self):
+        model = KPlanes(M_features=2, resolution=(8, 8, 8))
+        params = model.get_params()
+        assert set(params) == set(model.param_keys)
+        assert all(len(v) > 0 for v in params.values())
+
+    def test_default_forward(self):
+        import torch
+
+        model = KPlanes(M_features=2, resolution=(8, 8, 8))
+        out = model(torch.rand(5, 3) * 2 - 1)
+        assert out.shape == (5, 1)
+        assert torch.isfinite(out).all()