Only map OpenAI reasoning effort to Claude thinking for models that support thinking and use budget tokens (not level-based thinking). Also add "xhigh" effort mapping and adjust minimal/low budgets, with new raw-payload conversion tests across protocols and models.