Multi-Agent Collaborative Intelligence: Dual-Dial Control for Reliable LLM Reasoning