Fix final phase 1 and phase 2 inference and scoring issues

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

Initial submission: API Debug Environment