Initial Checks

• I confirm that I'm using Pydantic V2

Description

pydantic/pydantic-core@37ec6e7 introduced a regression in which objects on fields annotated as datetime, date, and time will be silently handled as if the object was of the incorrect type (falling through to infer_to_python unconditionally). This is most evident when using a union type, where I would expect that if serializing as a datetime fails, Pydantic would try the next available type. The logic seems to have been borrowed from the existing timedelta serialization which was also buggy in a similar way (verified on 2.11.7).

The actual issue seems to be in pydantic-core. The TypeSerializer used to downcast to ensure that the values are the correct type, but it does not do so now, only doing so whenever mode='json' is passed in. Indeed, the code works correctly if I pass mode="json" to BaseModel.model_dump().

This particular bug is insidious because it happens that most of the time, infer_to_python actually does the right thing. But with a specific user-defined type, one can see that we are silently doing the wrong thing.

I have a PR to fix this within pydantic-core.

Example Code

from datetime import date, datetime, time, timedelta

from pydantic import BaseModel
from pydantic_core import core_schema


class SampleType:
    @classmethod
    def __get_pydantic_core_schema__(cls, source, handler) -> core_schema.CoreSchema:
        serialization_schema = core_schema.plain_serializer_function_ser_schema(lambda v: None)
        json_validation_schema = core_schema.no_info_plain_validator_function(
            function=lambda v: v, serialization=serialization_schema
        )
        python_validation_schema = core_schema.any_schema()

        return core_schema.json_or_python_schema(
            json_schema=json_validation_schema,
            python_schema=python_validation_schema,
            serialization=serialization_schema,
        )


class SampleModel(BaseModel):
    d: datetime | SampleType
    dt: datetime | SampleType
    t: time | SampleType
    td: timedelta | SampleType


print(SampleModel(d='z', dt='z', t='z', td='z').model_dump())

# Expected: {'d': None, 'dt': None, 't': None, 'td': None} (from SampleType)
# Actual: {'d': 'z', 'dt': 'z', 't': 'z', 'td': 'z'} (from pydantic-core temporal serializers)

Python, Pydantic & OS Version

pydantic version: 2.13.0a0+dev
        pydantic-core version: 2.41.4
          pydantic-core build: profile=release pgo=false
               python version: 3.11.4 (main, Jul 22 2023, 23:52:07) [Clang 14.0.0 (clang-1400.0.29.202)]
                     platform: macOS-15.7.1-arm64-arm-64bit
             related packages: typing_extensions-4.14.1
                       commit: c8035874e