State the proved half-ULP accuracy bound; drop the and-what-is-not panel

Per-layer prefill progress bar so the long prefill wait is visible

Persistent in-memory runner; token-by-token streaming; theme, diagram, educational panels

Audit pass: scope claims to what is proved (trig via libm, trusted live boundary), credit SmolLM2, cut defensive slowness framing

Pin gradio 5.49.1 coherent stack; disable SSR (no node in slim image)

Verified SmolLM2 Space: Docker build, runner, weights, app

initial commit