Tighten clean_output role-prefix regex; debug-flag for raw output

Add identity-anchored system prompt + chat history support

Fix: reduce GPU duration 120s->60s, add error handling for quota exceeded

Deploy with ZeroGPU - full bf16, no quantization

Switch to ZeroGPU A100 - full bf16, no quantization

Initial MINDI 1.5 Space with 4-bit quantization

initial commit