Did FP8 make the model dumber? A per-prompt regression check for quantized serving

FP8 gave us a clean 1.5x on Qwen3-8B serving throughput on an RTX PRO 6000 Blackwell (1,725 to 2,597...