Description

On the pyarrow backend, clip (and clip_lower / clip_upper) replaced a null value in the data with a clip bound instead of preserving it:

import narwhals as nw, pyarrow as pa
s = nw.from_native(pa.chunked_array([[1, None, 3]]), series_only=True)
s.clip(0, 2).to_list()
# main:   [1, 2, 3]     <- the null became the upper bound
# polars: [1, None, 2]  <- null preserved

The cause is that pyarrow's min_element_wise / max_element_wise skip nulls by default, so min_element_wise([null], 2) returns 2. Passing skip_nulls=False keeps a null as null, matching Polars and pandas.

This is distinct from #2889, which was about None bounds (unbounded clip); here the null is in the data. The existing clip tests use no nulls in the data, so this was uncovered.

What type of PR is this? (check all applicable)

• 🐛 Bug Fix

Related issues

• None

AI assistance

• AI tools were used.

I used Claude Code to help locate the divergence (a differential comparison of pyarrow vs pandas/polars) and draft the fix and test. I reviewed every line, reproduced the bug and the fix locally, and take full responsibility for the change.

Checklist

• Fix applied to clip, clip_lower and clip_upper (the whole class) in the pyarrow backend
• Regression test test_clip_preserves_null added; fails on main for pyarrow, passes after; full clip_test.py green
• ruff check and ruff format clean