Reproducible research runs

Use Pinned Execution when a research run needs an explicit provider, model, API, and request. The completed execution record captures the settings and provider response so a result can be reviewed and reproduced.

Group related runs in a manifest and save a stable setup in an AI lockfile. The public endpoint is:

POST /v1/inference/pinned
GET  /v1/inference/pinned/{receipt_id}