Cut WAN family compile 321.6s->34.8s and load waste: per-shape AOT executable cache, zero-exec warmup, VACE/animate cast+staged-put reuse, VAE decode AOT, prompt embeds cache, threaded 2.1 TE warm