Evidence-Preserving Multi-Agent LLMs for SOC Threat Hunting with Different Collaboration and Prompting Strategies