fix rambling: stop at <|eot_id|> (Llama-3.1 turn end) + cap tokens + anti-loop

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