Why a diffusion model cannot simply keep its keys, what a fixed-size cache buys, and how a training gap shrank