Why your GPU reports 75 C while your VRAM is cooking at 105 C – the telemetry gap that kills LLM inference