Fix tensor_embedding.distance_proj for stock (PyG) matgl: split the fused Warp distance_proj [3*units] into distance_proj1/2/3 [units] (exact reshape). The fused form silently random-initialized on non-Warp matgl (load_state_dict strict=False), breaking the model; the split loads cleanly on PyG matgl and merges back under Warp. Si bulk modulus restored to the DFT reference.

state.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:249376319ee027c921b8fd46a6cd0dc9f6f6d1a41feeae41372cee06ecad8873
-size 4295077
+oid sha256:68084b0ff4f39da55b0f865409076045972ba6d29b34686ffded3f7d2a2ccd9f
+size 4291464