Submitted by Qi HU - SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces sssr-lab 0 3