fix: allow deleting predictions imported without model_version

[kcw2034]

Summary

Resolves #9717. Predictions imported without a model_version (e.g., from cloud storage or JSON files lacking the field) could not be deleted because the DELETE endpoint rejected falsy values and the import layer stored the literal string 'undefined' instead of NULL.

This is an alternative to #9735 that fixes the root cause (import-side) rather than masking it in the API.

Why not #9735

#9735 patches the DELETE endpoint to treat the literal string 'undefined' and 'null' as null-version. That works, but:

1. The 'undefined' string originates from three BE import sites (data_import/api.py ×2, tasks/serializers.py) defaulting to 'undefined' when model_version is missing. The frontend just echoes it back.
2. Magic-string handling at the API layer creates a foot-gun: a user with a legitimately-named 'undefined' model would lose unrelated predictions.
3. The mixed data state (NULL vs 'undefined' vs '') is left unresolved.
4. No regression tests.

What this PR does

Backend

• Import sites stop defaulting to 'undefined' — three call sites now use item.get('model_version') or None, so future imports without a version land as NULL.
• DELETE endpoint accepts JSON null / \"\" / 'undefined' as "no model version" via a single normalization step, then routes through an isnull-aware delete.
• Project.delete_predictions(None) sweeps NULL ∪ '' ∪ 'undefined' so callers don't need to know about the pre-migration data layout. This is robust during rolling deployments where some workers may still write the old value.
• Migration tasks/0062 backfills existing 'undefined' rows to NULL. Reverse is intentionally a no-op (can't distinguish migrated rows from genuinely-NULL ones).

Frontend

• PredictionsList.jsx removes the === \"undefined\" magic-string comparison. Falsy + legacy 'undefined' now both display "No model version" with an explanatory tooltip.

Tests

• projects/tests/test_api.py::TestDeletePredictionsAPI covers six scenarios: specific version, JSON null, empty string, legacy 'undefined' string, mixed null-ish sweep, and missing-key 400.

Test plan

• pytest label_studio/projects/tests/test_api.py::TestDeletePredictionsAPI — 6/6 pass
• pytest label_studio/projects/tests/test_api.py::TestProjectModelVersionsAPI — 3/3 pass (no regression)
• Migration detected by Django: showmigrations tasks lists 0062_backfill_undefined_model_version
• Manual: import predictions JSON without model_version, verify they delete cleanly

Files changed (7, +149 / -18)

File	Purpose
data_import/api.py (×2)	Stop storing 'undefined' on missing field
tasks/serializers.py	Same
projects/api.py	DELETE endpoint normalization
projects/models.py	delete_predictions(None) sweeps null-ish
tasks/migrations/0062_*.py	Backfill legacy data
projects/tests/test_api.py	Regression coverage
web/apps/labelstudio/.../PredictionsList.jsx	FE displays "No model version"

🤖 Generated with Claude Code

[netlify]

👷 Deploy request for heartex-docs pending review.

Visit the deploys page to approve it

Name	Link
🔨 Latest commit	4c79132

[netlify]

👷 Deploy request for label-studio-docs-new-theme pending review.

Visit the deploys page to approve it

Name	Link
🔨 Latest commit	4c79132

[netlify]

✅ Deploy Preview for label-studio-playground canceled.

Name	Link
🔨 Latest commit	4c79132
🔍 Latest deploy log	https://app.netlify.com/projects/label-studio-playground/deploys/6a0173ce5aa0da0008d9cde4

[netlify]

✅ Deploy Preview for label-studio-storybook canceled.

Name	Link
🔨 Latest commit	4c79132
🔍 Latest deploy log	https://app.netlify.com/projects/label-studio-storybook/deploys/6a0173ce8338be000815fe96

[kcw2034]

Addressed review feedback:

#1 delete_predictions caller audit — verified. The Python codebase has exactly one caller of Project.delete_predictions(model_version=...): the DELETE endpoint in this PR (projects/api.py:896). The other 3 grep hits are ml_models.ModelRun.delete_predictions() (different class, no args) and unrelated UI label mappings. No caller relies on the previous "delete-all" semantics.

#2 Title prefix — changed fix(predictions): → fix: to match repo convention.

#3 Migration batching — tasks/0062 now runs the UPDATE in 1000-row batches with atomic = False. The single long transaction is gone; each batch commits independently.

#5 Import-site integration test — added TestImportPredictionsModelVersionDefault (3 cases) covering missing / empty / explicit model_version through the /api/projects/<pk>/import/predictions endpoint. Verifies the DB row lands as NULL (or the given value), closing the loop between the import-side fix and the delete-side fix.

#4 Prediction.model_version field cleanup — intentionally left out. The current null=True, default='' combination is awkward, but a schema migration on the Prediction table on large installs is a separate risk class and deserves its own PR with focused review.

Final state: 12/12 new tests pass (6 delete + 3 import + 3 existing model-versions). Branch is 2 commits ahead of develop.