Fix chat UI (raw-string JS) + CPU autodetect + flash-attn + /diag

Serve pepekberbulu Q4_K_M GGUF live via llama-server (fast CPU inference)

Use prebuilt llama.cpp binary (b9670) instead of compiling (fix OOM)

Deploy VibeThinker-3B -> GGUF Q4_K_M conversion+upload job

initial commit