Fix: Gemini Live model → gemini-3.1-flash-live-preview, robust JSON parsing for exam generation

Fix Gemini Live: update model to gemini-2.5-flash-preview-native-audio-dialog, fix audio extraction from model_turn, add transcription config

Route ALL chat models through OpenRouter (free tier), add openai SDK, 6 models available

Fix critical bugs: move provider to InferenceClient constructor, migrate to google-genai SDK

Fix Gradio component wiring for Viz/Exam tabs, clean up full app

Add Gemini Live real-time bidirectional voice tab with native audio I/O

Fix exec builtins for visualization code execution, add pandas support

Improve robustness: handle DeepSeek think tags, better error handling, cleaner UI

Add comprehensive Study Assistant app with 6 tabs