[Bug] Safety filter incorrectly triggers on legitimate performance optimization requests

Question

**Bug Description**
Asked Fable to:
"use the codebase memory to gain detailed knowledge on the structure of the code. your objective is to propose paths to improve performance in token generation on this machine (amd strix halo - gfx1151 - rocm7.2). main objective is performance gains for hip backend, but vulkan backend should also be considered"

Inside llama.cpp repo, safeguards flagged the message and switched to Opus

**Environment Info**
- Platform: linux
- Terminal: kitty
- Version: 2.1.