Fix final phase 1 and phase 2 inference and scoring issues

readme

chore: push verified deployment fixes

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

Enhance episode state isolation and realistic grader hints for max scoring

Initial submission: API Debug Environment