Dual-signature export: decode_chat / prefill_128_chat (no LoRA, no classifier) + decode_classifier / prefill_128_classifier (with LoRA + classifier head). iOS Metal fused-attention shader should now match the chat graph.

model.litertlm

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:675e9249f05be4486da6143f1d03db70be424e226ff592b953f4dbf7c26fb1e8
-size 4155919584
+oid sha256:b16df2aeb387a1a020914e87b39f5db6cf843457945a4c07661d44eff3a540ef
+size 5791075552