TST: add tests for nullable astype behavior (NaN/<NA> preservation)

pandas/tests/dtypes/test_nullable_astype.py

@@ -0,0 +1,41 @@
+import numpy as np
+import pandas as pd
+import pandas.testing as tm
+import pytest
+
+def test_series_astype_nullable_int_preserves_nans():
+    # regression/edge: astype -> nullable integer dtype should preserve NaNs
+    s = pd.Series([1, np.nan, 3], dtype="float64")
+    res = s.astype("Int64")
+    # expected: dtype is nullable Int64 and NaN is represented as <NA>
+    assert res.dtype == "Int64"
+    expected = pd.Series([1, pd.NA, 3], dtype="Int64")
+    tm.assert_series_equal(res, expected)
+
+def test_series_astype_from_nullable_int_to_float_roundtrip():
+    # convert nullable Int64 -> float -> Int64, ensure values and missingness preserved
+    s = pd.Series([1, pd.NA, 4], dtype="Int64")
+    f = s.astype("float64")
+    assert f.dtype == "float64"
+    # float representation should have np.nan where original had <NA>
+    assert np.isnan(f.iloc[1])
+    # roundtrip back to nullable Int64
+    back = f.astype("Int64")
+    expected = pd.Series([1, pd.NA, 4], dtype="Int64")
+    tm.assert_series_equal(back, expected)
+
+@pytest.mark.parametrize("to_dtype", ["Int32", "Int64", "Float32", "Float64")
+def test_nullable_series_astype_various_dtypes_preserve_missing(to_dtype):
+    # small matrix of cases: ensure missingness preserved when casting between
+    # nullable integer/float dtypes and non-nullable numpy float dtypes
+    s = pd.Series([0, 1, pd.NA, 3], dtype="Int64")
+    res = s.astype(to_dtype)
+    if to_dtype.startswith("Int"):
+        # result should be nullable integer with pd.NA retained
+        assert str(res.dtype).startswith("Int")
+        expected = pd.Series([0, 1, pd.NA, 3], dtype=to_dtype)
+        tm.assert_series_equal(res, expected)
+    else:
+        # float dtypes: missingness becomes np.nan and dtype is numpy float
+        assert "Float" in to_dtype or to_dtype.startswith("float") or res.dtype.kind == "f"
+        assert np.isnan(res.iloc[2])