AutoRLBench v2 (solution level)

The repair level of AutoRLBench's second task list, which 24 of its 32 failed reinforcement-learning runs support. The agent repairs a writable copy of the run's workspace and submits a report without launching training; the harness reruns the experiment on the repaired workspace and grades only that fresh run against a metric threshold, and repairs that game the reward or rely on fetched upstream material fail. Each model runs each task once.