Auto-correct EVERY broken GDScript block in place (capped at MAX_FIX_PASSES)

ZeroGPU: keep model GPU-resident (canonical pattern)

Load the LLM once at startup instead of per ZeroGPU call

Hardcode chat memory to 4 turns (lock history_turns slider)

Add bounded multi-turn chat memory + turns slider (app.py)

GDScript RAG assistant: app + corpus (index added later via Colab)