v1.1: repetition_penalty + greedy 默认 + 能力边界声明 (闲聊 OOD 退化缓解)

Fix: robust duration estimate, cap 120s (ZeroGPU adds safety margin)

Fix: capture ponder_log reference without falsy-empty fallback

Fix: normalize Chatbot content parts (Gradio 6); move theme to launch()

Fix: drop Chatbot type kwarg (removed in Gradio 6)

Add Ponder-MoE 2B chat demo (ZeroGPU, streaming + ponder diagnostics)