Bug when row-filtering with null values?

[yohplala]

    """Test categorical data with nulls and read with filters"""
    fn = os.path.join(str(tempdir), 'test.parquet')
    # Create DataFrame with categorical and nullable columns
    df = pd.DataFrame({
        'cat_col': ['A', 'B', None, 'C'] * 2,
        'filter_col': list(range(8)),
        'nullable_int': pd.array([1, None, 3, 4] * 2, dtype="Int64")
    })
    df['cat_col'] = df['cat_col'].astype('category')
    print("df")
    print(df)
    # Write DataFrame
    write(fn, df, file_scheme='hive', row_group_offsets=[0, 4])
    # Test reading with row_filter and value filter
    pf = ParquetFile(fn)
    # Test with row_filter=True and filter_col > 6
    df_filtered = pf.to_pandas(filters=[('filter_col', '>', 6)], row_filter=True)
    expected = df[df['filter_col'] > 6].reset_index(drop=True)
    print("df_filtered")
    print(df_filtered)
    print("expected")
    print(expected)
    assert_frame_equal(df_filtered, expected)

shows:

testing.pyx:173: AssertionError
---------------------------- Captured stdout call -----------------------------
df
  cat_col  filter_col  nullable_int
0       A           0             1
1       B           1          <NA>
2     NaN           2             3
3       C           3             4
4       A           4             1
5       B           5          <NA>
6     NaN           6             3
7       C           7             4

df_filtered
  cat_col  filter_col   nullable_int
0     NaN           7  2267141176816

expected
  cat_col  filter_col  nullable_int
0       C           7             4
=========================== short test summary info ===========================
FAILED fastparquet/test/test_output.py::test_categorical_with_nulls_and_filters - AssertionError: DataFrame.iloc[:, 0] (column name="cat_col") are different

DataFrame.iloc[:, 0] (column name="cat_col") values are different (100.0 %)
[index]: [0]
[left]:  [NaN]
Categories (3, object): ['A', 'B', 'C']
[right]: ['C']
Categories (3, object): ['A', 'B', 'C']
At positional index 0, first diff: nan != C
============================== 1 failed in 8.17s ==============================

Describe the issue:

This test would tend to show a limitation when combining row filtering with nulls within the same dataframe.
I have pushed this test case in PR #956
and will try to investigate.

Environment:

• Python version: 3.13
• Operating System: Windows
• Install method (conda, pip, source): mix of source + unzipping windows wheels to retrieve compiled code.