Update your chat template for dsv4 if you're using llama.cpp

Recent llama.cpp commits broke preserve_thinking behavior for chat templates baked into older DSV4 GGUF files, degrading the model in coding-agent use. Fix: pass --chat-template-file with an updated template.