log memory at the failing block instead of probing

release the world-decode cache before the block loop

probe VRAM headroom at the point of failure

disable expandable segments: they abort the ZeroGPU worker

probe: CUDA from a spawned thread inside a ZeroGPU worker

log worker tracebacks and the allocator config

surface worker errors instead of blocking on the frame queue

fix Gradio 6 API: theme/css on launch, drop show_download_button

LiveWan streaming demo on ZeroGPU

initial commit