GPT-6 Luna Decisions

Add gpt-6-luna-decisions (GPT-6 Luna Decisions) as a decision helper/tool for my coding agent.
Keep my coding agent on its current chat model. GPT-6 Luna Decisions answers typed decision questions; it does not generate chat completions.

Write a small native HTTP helper for the OpenAI Decisions API shape, sent to my Experiential gateway:
- POST only to https://api.experientiallabs.ai/v1/decisions.
- Read the Bearer key only from EXPERIENTIAL_API_KEY. If missing, stop and ask me to set it locally; never ask me to paste or log it.
- Do not request an OpenAI API key or call api.openai.com.
- Send model "gpt-6-luna-decisions" exactly.

Send a JSON request with model, input (text, or messages whose content is input_text and inline base64 input_image parts), and questions (a list). Each question needs instructions and a type, and may carry a unique name:
- predicate: the answer is the probability the statement is true.
- choice: choices lists the allowed values with descriptions; the answer names one value.
- score: levels lists ordered labels with descriptions; the answer is a score.
Use a concise diff and test summary as input, not the entire repository.
Gateway limits: at most 32 questions per request, 64 options per choice, and 2-10 levels per score. These are platform admission bounds, not provider limits.
Request shape reference: https://developers.openai.com/api/docs/guides/decisions

Answers come back in question order; an unnamed question's answer has no name, and a refusal answer means the model declined that question. Read the answer at the review question's position, and show the decision and reported usage to me. Probabilities and confidence are not guarantees of correctness. Never execute a returned choice, merge a change, or bypass review automatically.
Use a bounded timeout, surface HTTP and transport errors, and do not add automatic retries or Idempotency-Key: an unknown outcome may already be charged. This endpoint does not stream and has no Chat Completions, Responses, or Messages facade.
Prepare the helper and a mocked test first. Ask before making a potentially charged live call; do not claim it worked without a real response. Tell me which files you changed.