Reconcile model-registry metrics with published leaderboards

[unes07]

Problem

Metrics recorded in the model registry differ substantially from values in the published leaderboards, making model selection and reported performance difficult to audit.

Proposed work

• Trace each registry metric and leaderboard row to its exact dataset, split, preprocessing, target, and artifact.
• Identify whether differences come from stale reports, different protocols, leakage, metric aggregation, or artifact mismatch.
• Establish one authoritative evaluation output and documented selection policy.
• Regenerate inconsistent reports, registry entries, and artifacts as needed.

Acceptance criteria

• Every published model metric is reproducible from the training pipeline.
• Registry and leaderboard values agree for the same evaluation protocol.
• Selection criteria are documented per crop and target.
• Zero-valued targets use appropriate metrics; unstable MAPE values are not presented without context.
• A regression test prevents report, registry, and artifact metadata from diverging.

Depends on #2.