Release 0.4.1 — fix sklearn 1.8 ImportError

[adrian-prior]

Hotfix on top of #285 (v0.4.0). Caught by a TestPyPI smoke install before publishing 0.4.0 to real PyPI — v0.4.0 was never published to PyPI, going straight to v0.4.1.

What broke

scikit-learn 1.8.0 (released recently, the latest resolved version on a fresh install) renamed sklearn.utils.validation._is_pandas_df → is_pandas_df (no underscore). src/tabpfn_extensions/misc/sklearn_compat.py imports the old name at module top-level, so any from tabpfn_extensions import ... on sklearn ≥ 1.8 immediately raises ImportError. That's the entire public API broken for anyone resolving sklearn 1.8 — which is the default on a clean pip install tabpfn-extensions today.

Fix

sklearn_compat.py now tries _is_pandas_df, falls back to is_pandas_df, and finally to the same pure-Python implementation already used in the older-sklearn branch. Three-line defensive shim.

Verification

Reproduced the original bug with a clean venv: pip install -i https://test.pypi.org/simple/ tabpfn-extensions==0.4.0 resolves sklearn 1.8 and crashes on import. After this patch (installed locally) the same import chain succeeds on sklearn 1.8.

Smoke results post-patch:

Check	Result
from tabpfn_extensions import * on sklearn 1.8	✅
warn_if_no_kv_cache silent on non-TabPFN estimator	✅
SurvivalTabPFN import	❌ expected (GPL opt-in, scikit-survival not installed)
TabEBM import	❌ expected (pre-existing RES-1541)

Release flow after merge

1. git tag v0.4.1 <merge-commit> && git push origin v0.4.1
2. TestPyPI dry-run: python -m build && twine upload --repository testpypi dist/* → install in clean venv to re-verify.
3. Real PyPI: twine upload dist/*.
4. GitHub Release for the tag — body = the new [0.4.1] section of CHANGELOG.md.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the project version to 0.4.1 and addresses an import failure caused by the renaming of _is_pandas_df in scikit-learn 1.8+. The fix includes a compatibility shim with a pure-Python fallback. Feedback highlights a missing sys import in the fallback function, a version mismatch in the uv.lock file, and the accidental removal of platform-specific wheels, suggesting a regeneration of the lock file for all architectures.

I am having trouble creating individual review comments. Click here to see my feedback.

src/tabpfn_extensions/misc/sklearn_compat.py (273-279)

The fallback implementation of _is_pandas_df uses sys.modules, but sys is not imported in this scope. This will cause a NameError if the fallback path is ever executed. Adding a local import ensures the fallback works correctly.

            def _is_pandas_df(X):
                """Return True if X is a pandas DataFrame."""
                import sys
                try:
                    pd = sys.modules["pandas"]
                except KeyError:
                    return False
                return isinstance(X, pd.DataFrame)

uv.lock (3356)

There is a version mismatch in the lock file. pyproject.toml is being updated to 0.4.1, but uv.lock is only being updated to 0.4.0. Please run uv lock to ensure the lock file correctly reflects the project version.

uv.lock (299-300)

Multiple platform-specific wheels (e.g., ppc64le, s390x) are being removed from the lock file. This often happens when the lock file is updated in an environment that doesn't resolve all supported architectures. If these platforms are still supported, please regenerate the lock file using uv lock --all-platforms.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[LeoGrin]

lgtm! You updated sklearn_compat to latest version right?

[adrian-prior]

@LeoGrin yes, indeeed