187 live prompts, 27 bugs: what testing my local AI agent against a real 7B model taught me