llama.cpp chat templates fail on ROCm builds

This is a known friction point in the local LLM ecosystem — GGUF metadata embeds chat templates, but llama.cpp's `--chat-template auto` path has reliability gaps that disproportionately affect non-CUDA backends like ROCm.

• –The `--chat-template auto` flag is supposed to read template metadata from the GGUF file, but silently falls back to no template on parse failures, producing incoherent outputs
• –ROCm/HIP builds lag behind CUDA in community testing, meaning edge cases like this surface later and stay unfixed longer
• –The workaround (manually specifying `--chat-template chatml` for Qwen, etc.) is fragile and requires per-model-family knowledge that isn't documented centrally
• –The user's llama-runner wrapper — Makefile + TUI + model picker — is the kind of quality-of-life tooling that keeps filling the gap left by llama.cpp's CLI complexity
• –This is a community question post, not an announcement, but it surfaces a real usability gap in local inference tooling on AMD hardware