Fix final phase 1 and phase 2 inference and scoring issues

readme

improve: hard tasks max_steps=10, fix reward gap, clearer descriptions

Add Expert tier tasks to evaluate multi-step UUID stateful tracking and OpenAPI discovery