Avoid leaking raw <think> tag when reasoning is truncated

Bump default gen length/steps; render <think> as collapsible thinking panel

Add word/token position constraints

Fix Gradio 6 API: drop Chatbot type arg, move css to launch()

Delete files __pycache__/* with huggingface_hub

Add iLLaDA-8B-Instruct masked diffusion LLM demo

initial commit