wire Codette's cocoon memory from CodetteBrain bucket (read-only /data)

Upgrade to REAL orchestrated Codette on ZeroGPU (transformers backend for the llama.cpp pipeline)

fix generate(): apply_chat_template return_dict=True + **inputs (transformers 5 BatchEncoding)

gradio 6: drop removed ChatInterface(type=) arg

fix ZeroGPU: load adapters on CPU (peft infer_device saw hijacked cuda) before intercepted .to(cuda)

vendor 8 perspective adapters into the Space + load from local paths (token-free, no gated fetch)

Codette multi-perspective demo: 8 perspective LoRAs on Llama-3.1-8B (ZeroGPU)