pandas 3: silent behaviour changes that break 2.x code
Checked on pandas 3.1.0 with NumPy 2.5. (1) Chained assignment like df['a'][mask] = 0 never modifies df under Copy-on-Write. It only emits a ChainedAssignmentError, which is a warning and not an exception, so code that filters warnings changes nothing and keeps running; use df.loc[mask, 'a'] = 0. (2) String columns now default to dtype 'str', so checks like df[c].dtype == object return False and the branch is skipped; use pd.api.types.is_string_dtype. Missing values in these columns are NaN, not None. select_dtypes(include='object') still matches str columns, with a Pandas4Warning that this will be removed. (3) Series.to_numpy() and .values return read-only views, so arr[0] = 5 or arr += 1 raises ValueError: assignment destination is read-only; pass copy=True if you need to mutate.
Reproduced all three points on pandas 3.1.0 / NumPy 2.5.4 / Python 3.12, against pandas 2.2.3 as the control. (1) df['a'][mask] = 0 leaves df unchanged and only emits ChainedAssignmentError; on 2.2.3 it modified df with a FutureWarning. (2) Series(['x', None]).dtype is 'str', == object is False, the missing value is nan, and select_dtypes(include='object') still matches with Pandas4Warning. (3) to_numpy()[0] = 5 raises 'assignment destination is read-only'. A small precision for .values += 1: the message there is 'output array is read-only', so a grep for the first message will miss it.