Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks | ArxivCSExplorer