BUG: Fix MultiIndex partial-key lookup when np.datetime64 indexes datetime.date level (GH#55969)

[logiop]

Summary

Fixes #55969.

When a MultiIndex has an object-dtype level whose values are
datetime.date objects, using np.datetime64 as a partial key
(e.g. df.loc[(np.datetime64("2023-11-01"), "A")]) silently ignores all
key components after the date level and returns every row whose date matches —
regardless of what the remaining key levels say.

Root cause

MultiIndex._partial_tup_index guards with lab not in lev to detect keys
absent from a level and then takes a fast short-circuit path that returns
immediately (without processing the remaining key levels).

The hashtable-based __contains__ does not coerce np.datetime64 →
datetime.date, so the membership test always evaluated to True (key
"absent") for np.datetime64 keys against object-dtype levels holding
datetime.date values. The subsequent key components (e.g. "A", "B")
were therefore never applied, producing wrong results.

The same problem exists in _get_loc_single_level_index, which is called
when the key falls into the follow_key portion of MultiIndex.get_loc.

Fix

Before the lab not in lev guard in both _partial_tup_index and
_get_loc_single_level_index, attempt to convert a bare np.datetime64
key to its Python-scalar equivalent when the target level is object-dtype:

• midnight timestamps → datetime.date
• all other timestamps → datetime.datetime

Only substitute the converted value when it is actually present in the level,
so genuine type-mismatch key errors are preserved.

Example

import datetime, numpy as np, pandas as pd

dates = [datetime.date(2023, 11, 1),
         datetime.date(2023, 11, 1),
         datetime.date(2023, 11, 2)]
df = pd.DataFrame({"dates": dates, "t1": ["A","B","C"], "t2": ["C","D","E"],
                   "vals": [0.1, 0.2, 0.3]})
df.set_index(["dates","t1","t2"], inplace=True)

date = np.datetime64("2023-11-01")

# Before fix: both return 2 rows (all rows for 2023-11-01)
# After fix:  each returns exactly 1 row
df.loc[(date, "A")]   # 1 row: (2023-11-01, A, C)
df.loc[(date, "B")]   # 1 row: (2023-11-01, B, D)

Test plan

• Added test_loc_datetime_date_index_with_np_datetime64 in
  pandas/tests/indexes/multi/test_indexing.py which exercises:
  • df.loc[(np_date, second_level)] for two different second-level values
  • equality with the datetime.date-keyed equivalent
  • MultiIndex.slice_locs with np.datetime64 partial tuple keys

🤖 Generated with Claude Code