Add interactive live-play mode: steer the world model frame by frame with streaming output

Calibrate ZeroGPU duration to measured 6.85 s/step; add cost note

Give the fixed-prompt text-encoder stub a real buffer so diffusers can read device/dtype

Gradio 6 theme/css placement; JSON state so the two-stage chain is client-drivable

Marionette world-model demo: dynamics + geometry bridge + observation diffusion

initial commit