fix: skip __round__ for datetime keys in number_to_string (fixes #550)

[Mukller]

Summary

Fixes #550 — TypeError: type datetime.datetime doesn't define __round__ method when a datetime is used as a dict key with ignore_numeric_type_changes=True (or the other ignore_string_* flags).

Root cause

number_to_string() includes datetimes in the numbers tuple so that datetime objects participate in numeric key comparison. However, none of the special-case branches (Decimal, Fraction, complex) match datetime objects, so execution falls to the else block:

number = round(number=number, ndigits=significant_digits)

datetime objects do not implement __round__, which causes the TypeError.

Fix

Add elif isinstance(number, datetimes): return number before the else branch. Datetime values are already unique/hashable and require no rounding for key-comparison purposes.

# Before
else:
    number = round(number=number, ndigits=significant_digits)

# After  
elif isinstance(number, datetimes):
    return number  # datetimes don't support __round__; return unchanged
else:
    number = round(number=number, ndigits=significant_digits)

Repro (now works after fix)

import datetime
from deepdiff import DeepDiff

d1 = {datetime.datetime(2020, 5, 17, 22, 15): 10.0}
d2 = {datetime.datetime(2020, 5, 17, 22, 15): 10}
diff = DeepDiff(d1, d2, ignore_numeric_type_changes=True)
# Before: TypeError: type datetime.datetime doesn't define __round__ method
# After:  {}

Changes

File	Change
deepdiff/helper.py	Add elif isinstance(number, datetimes): return number in number_to_string()
tests/test_helper.py	Add test_number_to_string_with_datetime_key + test_deepdiff_with_datetime_key_and_ignore_numeric

[Mukller]

Code review

Why datetimes is in numbers

numbers (line 196) is defined as only_numbers + datetimes to allow datetime objects to be compared via number_to_string — for example, to round-off sub-second precision when significant_digits is set. The intent is that datetime_normalize() typically handles the rounding step for actual datetime values, while number_to_string handles the key path.

Why the early return is safe

The fixed branch does return number (the original datetime object), not a string. Callers that receive a non-string return value from number_to_string use it as a dict key (see diff.py around line 684–699), so returning the original object is correct: the key identity is already established by the datetime's natural hash.

Minimal blast radius

The only change is the insertion of one elif branch. All existing paths through number_to_string (int, float, Decimal, Fraction, complex, numpy scalars) are unaffected. The Fraction branch immediately above already follows the same early-convert, fall-through pattern, so this is stylistically consistent.

Test coverage

Two tests added:

1. Unit — test_number_to_string_with_datetime_key: verifies all four datetime types (datetime, date, timedelta, time) pass through unchanged.
2. Integration — test_deepdiff_with_datetime_key_and_ignore_numeric: reproduces the exact scenario from issue #550 end-to-end.

Edge cases

• numpy.datetime64 is included in datetimes (line 86/100) and would also hit the round() error on NumPy < 2.0. The fix covers it as well.
• datetime.timedelta supports addition but not __round__ — also covered.