When migrating an entire MLflow instance using bulk tools (export-all then import-all), registered model tags are preserved, but per-version tags on model versions are not preserved.

Observed behavior

• Registered model tags are present after import.
• Model-version tags for many/all versions are missing after import.

Expected behavior

• Both registered model tags and model-version tags should be preserved.

Confirmed code path

• Full-instance export goes through export_all.py, which delegates to export_models.py.
• Each model export calls export_model.py, which fetches versions via list_model_versions.
• In non-UC path, list_model_versions returns search results directly at model_utils.py (or latest versions at model_utils.py), without per-version population via get_model_version.
• In UC path, versions are explicitly populated via get_model_version at model_utils.py.
• Import side does pass src_vr tags into create_model_version (import_model_version.py, import_model_version.py), so missing tags are likely already absent in exported metadata for this path.

Why this affects OSS full-instance migration

• Whole-instance migration uses bulk export/import flow (not single export-model-version by default).
• The non-UC export path for model versions differs from UC and can propagate incomplete version metadata.
• Therefore, model-version tag preservation is not guaranteed in OSS bulk migration and can fail in practice.

Reproduction (high level)

1. Create a registered model with version tags on source OSS MLflow.
2. Run export-all.
3. Run import-all into destination OSS MLflow.
4. Inspect destination model versions: tags are missing, while model-level tags remain.

Potential fix direction

• In non-UC list_model_versions, populate each version with get_model_version (same strategy already used for UC in model_utils.py) before serialization.