[warrkan]

Export DINOv2 classifiers with the official facebookresearch/dinov2 ImageNet linear head (reg4 variants).
Resolves #655

[Copilot]

🟡 Changes recommended

The test checks only feature shape, leaving incorrect patch averaging or reversed feature concatenation undetected; it should assert the exact returned values.

1 open issue

• Assert exact CLS-plus-patch-mean feature values · New

What changed in this PR

This PR addresses #655 by initializing the converter’s DINOv2 classifier with the official ImageNet linear head, enabling meaningful predictions from exported models.

Changes:

• Loads the pretrained reg4 head and adapts backbone features to its expected input.
• Replaces placeholder label metadata with resolved class names and adds unit tests.
• Adds Ruff as a pinned lint dependency.

File	Description
model_converter/​uv.lock	Locks Ruff and the lint dependency group.
model_converter/​tests/​unit/​test_getitune.py	Tests head initialization, validation, routing, and label metadata.
model_converter/​src/​model_converter/​converters/​getitune.py	Integrates the official DINOv2 head and updates exported label metadata.
model_converter/​pyproject.toml	Adds the Ruff lint dependency group.

_{Review effort: Balanced}