Version of Awkward Array

2.8.7

Description and code to reproduce

It looks that ak.to_arrow and ak.to_arrow_table make an error when processing 64-bit date types into arrows 32-bit date type, causing data to be changed in the conversion.

(EDIT: much simpler example)

ak.to_arrow(np.array(['2011-01-27', '2011-01-28', '2011-01-29', '2011-01-30'], dtype='datetime64[D]'), extensionarray=True)

<awkward._connect.pyarrow.extn_types.AwkwardArrowArray object at 0x0000023BAFC4E2C0>
[
  2011-01-27,
  1970-01-01,
  2011-01-28,
  1970-01-01
]

Original more complex demo.
I'm using polars here purely for nice visual and comparison:

data = np.array([(0, '2025-03-07'), (1, '2025-03-10')], dtype=[('index', '<i8'), ('date', '<M8[D]')])
df1 = pl.from_numpy(data)
df2 = pl.from_arrow(ak.to_arrow_table(ak.to_packed(ak.from_numpy(data)), extensionarray=False))
pltst.assert_frame_equal(df1, df2)

• I'm expecting a round-trip fidelity test of dates via numpy via awkard via arrow to be 100% faithful
• ak.to_arrow_table(ak.to_packed(ak.from_numpy(data))) gives a 1970-01-01 when it should be 2025-03-10.

pyarrow.Table
index: extension<awkward<AwkwardArrowType>> not null
date: extension<awkward<AwkwardArrowType>> not null
----
index: [[0,1]]
date: [[2025-03-07,1970-01-01]]

Awkward array is awesome - thank you for such a useful and powerful package.