[kukirin-ui]

• start.sh: --flash-attn on (newer llama.cpp needs a value; the bare flag made the GPU tier fail and fall back to CPU). Added a GPU tier without Flash-Attention/quantized KV before the CPU fallbacks. Default THREADS=2 when GPU offload is on.
• setup.sh + start.sh: detect the real GPU name from llama-server --list-devices (first Vulkan0: <name> ( line, e.g. Samsung Xclipse 960) and save it in ~/.nullcloud-profile.env.
• setup.sh: installs packaged llama-cpp + llama-cpp-backend-vulkan + vulkan-loader-android (removes vulkan-loader-generic); builds from source only with BUILD_VULKAN=1 or if no llama-server. GPU-first wording.
• README: measured llama-bench table and screenshot (Samsung Xclipse 960, Qwen3 4B Q4_K_M, Vulkan, -ngl 99, 2/4/8 threads); THREADS=2 recommended. Troubleshooting for CPU backend and --flash-attn errors.
• Web UI hardware card shows the measured numbers.
• verify-all.sh passes.