Foundation: opt-in live Jev backend and response recorder #64

Goal

Let a reader with their own key run any recipe against Jev, and capture those responses as recorded fixtures.

Scope

• LiveBackend wrapping typesafe_sdk.TypeSafeClient.system_one, available only with the live extra, JEV_COOKBOOK_LIVE=1, and TYPESAFE_API_KEY.
• The model is passed explicitly per run. The model string the API returns and the usage block are kept on the result.
• A request budget guard: a maximum number of calls per run, enforced in code, with a clear error when exceeded.
• A recorder that writes responses into a recipe's fixtures with provenance recorded, the returned model string, and the date.
• A documented path for alternative backends through the System One Adapter (S08), needed by recipes 50 and 55. Decide whether that is an optional extra and write the decision down.
• The key never appears in logs, exceptions, fixtures, or notebook output.

Acceptance criteria

• Tests run with the HTTP layer mocked; CI makes no network call and needs no key
• Missing key or missing opt-in gives a clear, actionable error
• A test shows the budget guard stops a run
• A test shows recorded fixtures replay through ReplayBackend byte for byte
• The pull request states whether a real live call was made during development. If none was, it says so
• Lint and tests pass in CI, or locally with the output pasted in the pull request if CI does not exist yet

Dependencies

• Blocked by Foundation: decision backends with typed answers, replay, and scripted modes #63 (decision backends with typed answers, replay, and scripted modes)
• Unblocks Foundation: reader documentation for getting started, offline and live modes, and the glossary #72
• Unblocks the recipes labelled needs: live run
• Tracked in Epic: Foundation #74

References

• Recipe contract
• Catalog
• TypeSafe documentation index for tools: https://docs.typesafe.ai/llms.txt
• Python SDK: https://docs.typesafe.ai/sdk/python