[its-amann]

This change does not add a model or a dataset, so the dataset and model checklists do not apply.

New tasks need a concrete license. The license field is already typed as Licenses | StrURL | None, so pydantic rejects anything that is not a known license id or a URL. The only unclear values that get through are None and "not specified", and this PR targets those two.

What changed:

• TaskMetadata gets has_unclear_license(), which is true when the license is None or "not specified", and _validate_license(), which raises a ValueError in that case.
• _validate_metadata() does not call _validate_license(). None and "not specified" still pass metadata validation, so existing task files keep loading and nothing new starts failing at runtime.
• The policy is enforced in the tests instead. test_all_metadata_is_filled_and_valid calls _assert_license_policy(task) for every task after the historic datasets check. A task on the exception list must still have an unclear license (otherwise it should be removed from the list), and any other task must pass _validate_license().
• test_unclear_licenses_limited_to_existing_tasks runs the same check over all tasks, including superseded, aggregate, beta and private ones, and fails if the exception list names a task that no longer exists.
• Unit tests in test_task_metadata.py cover None and "not specified" passing _validate_metadata(), the same two values being rejected by _validate_license(), and pydantic rejecting a value like "unknown".

552 existing tasks that use "not specified" are listed in tests/test_tasks/_tasks_with_unclear_license.py. Their license values are unchanged. The file says not to add new tasks and to remove a task once it has a real license.

The dataset contribution guide and the license docstring now say a concrete license is required for new tasks, and the dataset PR checklist has an item for it.

Fixes #5348

@KennethEnevoldsen @isaac-chung this follows what you both suggested on the issue, plus the review feedback from @Samoed and @ayush1298. Happy to change how the existing tasks are handled if you'd prefer something else.

[its-amann]

Thanks @ayush1298 and @Samoed, that makes sense. Since pydantic already rejects anything that isn't in Licenses or a URL, I'll drop _UNCLEAR_LICENSES and the normalization and just check license is None or license == "not specified". I'll trim the tests down to None, "not specified" and one value pydantic rejects, and I'll also undo the formatting change in _metadata.py so the diff stays small. For the 552 existing tasks, I'm happy to move the list into its own file (something like tests/test_tasks/tasks_with_unclear_license.py), or handle it another way if @KennethEnevoldsen or @isaac-chung prefer. The failing test job is Test lowest dependencies, which is failing on main right now as well.

[isaac-chung]

This has 552 tasks, which is too much. Maybe we can move it to separate file, but will wait for others opinion here

Yep a separate file sounds good. Thanks.

[its-amann]

Thanks @isaac-chung. I moved the list of existing tasks into tests/test_tasks/_tasks_with_unclear_license.py and test_metadata.py now imports it from there, so the test file is back to its normal size. The checks could use a workflow approval when someone has a moment.

[isaac-chung]

Any particular reason it's prefixed with _test ?

[its-amann]

Any particular reason it's prefixed with _test ?

no strong reason, just to mark it as a helper and not a test.

[Samoed]

I think you can test this after this check, because this would skip a lot of tasks in _HISTORIC_DATASETS

[Samoed]

This would create a lot new unvalidating metadata which can be a bit unexpected. I think we can remove _validate_license from _validate_metadata

[its-amann]

@Samoed done in 54587db