mohdel v3.9.0

Added

• AnswerResult.reasoningParts: the model's reasoning as encrypted
  reasoning parts ({ type, text, encrypted, id, model }). A caller that puts
  them at the head of the assistant message it records gets them replayed on
  the next call to the same model, and dropped on a call to any other.
• Responses-API providers (openai, chatgpt, meta, xai) request
  include: ['reasoning.encrypted_content'] when the model reasons (the
  catalog declares thinkingEffortLevels and the effort is not none), and
  return the reasoning items they receive as reasoningParts. Every other
  adapter drops an encrypted reasoning part.