Offload what the card can hold, survive a dropped connection, and stop sending prompts the model cannot read

Workers print their own code and are claimed from the site; fetch and verify llama.cpp