Pin diffusers to abot-world-modular@206121e08 (tiny VAE decode in the streaming preset)

Raise per-session ZeroGPU budget to 150 s

Disable SSR: the Node proxy does not forward the /ws WebSocket upgrade

Live front-end: gradio.Server + WebSocket paced frame stream (shell from acvlab/abot-world-interactive), engine = pipe.stream(action_source=...)

Give each block 6 s of ZeroGPU duration budget (measured ~5 s/block)

Run on ZeroGPU: spaces.GPU session generator + command queue for controls

Barebone interactive ABot-World on Modular Diffusers: one pipe.stream(action_source=...) loop