Record the preparation contract in model provenance

[pavsoss]

Summary

The model registry has always looked for a model card beside the artifacts to populate the training time, metrics and label set, but nothing ever wrote one — so those provenance fields were reported as empty for every deployed model. Separately, artifacts carried no record of how text must be prepared for them, meaning a future change to preparation could silently desynchronize from a model trained under the old rules.

Part of #1037.

Changes

Training now writes the provenance card after a successful save, recording when the model was trained, how it scored on the held-out split, its label set, and the text-preparation contract it was trained under. The card is written atomically and only after the artifacts land, so it can never describe a model that failed to persist or be read half-finished.

When a served model was trained under a different contract than the one in force, the mismatch is reported as the artifacts are loaded instead of quietly degrading predictions. It is surfaced rather than fatal: the model still answers, because taking the API down over a metadata disagreement is worse than serving a model an operator may already be retraining. Artifacts predating the card make no claim about their preparation and remain servable.

Test plan

• Matching, mismatching and unrecorded contract versions, and the version appearing in the reported provenance payload.
• A card written by training is read back through the registry.
• Card contents: recorded contract, holdout accuracy, training row count, label set and training timestamp.

Depends on

Stacked on #1038 and #1039. Their commits are included here, so this diff shrinks to just the provenance changes once they merge.

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.