Sampling-based planners: RRT vs. RRT* and the role of optimality

You are tasked with planning a path for a robot arm in a high-dimensional configuration space (e.g., a 7-DOF manipulator) cluttered with obstacles. A colleague proposes using RRT (Rapidly-exploring Random Tree); you suggest RRT* instead. Explain how RRT works, what specific property RRT* adds and the mechanism by which it achieves it, what asymptotic guarantees each provides, and discuss the practical trade-offs of choosing RRT* over RRT in this scenario. When (if ever) would plain RRT still be the better choice?