Restore a reproducible training and benchmark pipeline

[unes07]

Problem

The README says the training notebook is published, but the repository does not contain a complete pipeline that reproduces the packaged artifacts and reports.

Proposed work

• Publish the training script or notebook and any required configuration.
• Pin dependencies and random seeds.
• Record exact dataset IDs, versions, hashes, features, targets, preprocessing, and split assignments.
• Generate benchmark reports, registry metadata, and model artifacts from one documented command or workflow.

Acceptance criteria

• A clean environment can reproduce reports and artifacts from documented inputs.
• Runs record dependency versions, seeds, data hashes, and split identifiers.
• Generated outputs are deterministic within documented tolerances.
• Training code does not depend on unpublished local files.
• Tests cover pipeline configuration and registry generation.

Depends on open-turba/turba-data#5, open-turba/turba-data#6, and open-turba/turba-data#7.