I trained a 30M-param LLM from scratch and the scaling "floor" was a mirage