We forced the model to answer in strict JSON, and accuracy on our reasoning-heavy task dropped. Why, and what would you change?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.