fix: disable GLM thinking (reasoning model burns max_tokens on reasoning_content -> empty answers); bump max_tokens to 500

Cokro LLM comparison: GLM-4.7-Flash (free) vs DeepSeek v4-flash

initial commit