Same trained LoRA, three inference variants vs input and beat, plus the v2 LoRA (simple caption) with and without the lift prompt. Finding: the active lift/window-pull prompt unlocks the trained edit (white lift + cast removal); LoRA scale alone changes nothing.
