Fix final phase 1 and phase 2 inference and scoring issues

readme

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

Implement strict judge feedback: ambiguous task descriptions, server-side payload clues, increased exploration steps

Enhance episode state isolation and realistic grader hints for max scoring

fix: Dockerfile to root, port 7860, README frontmatter

Initial submission: API Debug Environment