Halve @spaces.GPU duration 60s → 30s to save quota per call

System prompt: explicit greeting + short-followup rules; never introduce self

Slash system prompt: short, answer directly, no identity preamble

Ignore Gradio session history + bake identity facts into system prompt

Defensive history shape + voice-rewrite of system prompt

Fix: handle both dict + tuple history shapes in format_mistral_prompt

Bootstrap: llm.py