Add /evaluate endpoint: numerical position MSE model vs Pymunk

Swap featured: dominos -> pyramid (vetted via live API)

Demos: side-by-side Model vs Pymunk ground truth + Featured list

Fix slow/frozen-objects on dense scenes (30-40 obj)

Drop show_copy_button (removed in gradio 6 API)

Fix canvas->state sync + add velocity vector editor

Interactive canvas editor + fix lost collisions

UI: turn the scene-JSON box into a proper editor (taller, interactive, accordion)

Editable scene (JSON) + full-context rollout

Drop broken speculative configs; keep flash_attn (proven ~+36%) + plain fallback

Auto-select fastest working llama.cpp backend (flash+spec -> spec -> flash -> plain)

Fix speculative decoding: logits_all=True for the draft scores buffer

debug: surface generation traceback to status to diagnose ValueError

Speed up: flash attention + prompt-lookup speculative decoding

Speed up serving + allow 200 frames

Raise frame slider to 120 (default 60) — GPU is ~0.6s/frame, fits the 180s window

Run on ZeroGPU: spaces @GPU + CUDA llama.cpp (n_gpu_layers=-1)

Fix CPU perf: pin n_threads=2 (cpu-basic), tighter token budget

Rebuild as native Gradio app: server-side LFM2 (GGUF) rollout

Bake React routes into Gradio app before launch via _app=

Disable Spaces hot-reload so React UI serves (fixes 7861 + hijack)

Serve React UI via Gradio's own launched server (cpu-basic)

Restore real app: FastAPI(React+API) + mounted Gradio on cpu-basic

diag: vanilla gradio app to discover HF launch mechanism

Fix gradio Space startup: override demo.launch() to serve the full app

Fix gradio Space startup: serve at import time on GRADIO_SERVER_PORT

Redeploy as gradio SDK Space (no custom Docker)