SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces | ArxivCSExplorer