Skill Availability and Presentation Granularity in Large-Language-Model Agents: A Controlled SkillsBench Study | ArxivCSExplorer