The model and the data that made it share a snapshot

Six weeks after shipping v7, someone asks: which rows trained it, with which preprocessing, on which library versions?

You don't check a registry. You restore the snapshot named train-v7. Before the training script ran, you POSTed /api/v1/containers/{id}/snapshots and gave it that alias. The snapshot captured the filesystem: the exact parquet files, the requirements.txt and the installed packages, the feature-engineering notebook at the commit that ran, the config YAML with the random seed. After training, a second snapshot, train-v7-done, added the weights to the same frozen world.

The answer to "what made v7" is a container you can boot. Re-run the script in it and you get v7 again.

Later snapshots store only the changed blocks, so twenty model versions don't mean twenty copies of the dataset.

Model lineage isn't a registry you maintain. It's a filesystem you froze.