Serialization schema allowing inf and NaN

[aeisenbarth]

I have a marker class (not a subclass of BaseModel) that implements __get_pydantic_core_schema__ to validate/serialize a third-party library's class as a list of numbers (int, float, nan, inf).

Since Pydantic 2.7, it serializes [inf] as '[null]' (apparently it shouldn't have worked before). I found that one has to explicitly enable ConfigDict(ser_json_inf_nan="constants") (#7007 etc.) because the RFC standard does not allow Infinity/NaN. However, I don't have a BaseModel, so I cannot use ConfigDict.

Example:

from typing import Any
from pydantic import BaseModel, GetCoreSchemaHandler
from pydantic_core import CoreSchema, core_schema
from pydantic_core.core_schema import list_schema


class MyPydanticType:
    @classmethod
    def __get_pydantic_core_schema__(
        cls, source_type: Any, handler: GetCoreSchemaHandler
    ) -> CoreSchema:
        # A type that allows any nested lists containing either integers or floats/nan/inf
        python_schema = core_schema.no_info_before_validator_function(
            lambda data: data,  # Actual implementation is more complex!
            list_schema(),
        )
        json_schema = core_schema.chain_schema([list_schema(), python_schema])

        return core_schema.json_or_python_schema(
            json_schema=json_schema,
            python_schema=python_schema,
            serialization=core_schema.plain_serializer_function_ser_schema( #        <———
                lambda data: list(data), # Actual implementation is more complex!
            ),
        )

def test_serialize_my_type():
    class A(BaseModel):
        a: MyPydanticType

    value = [float("-inf"), float("nan"), float("inf")]
    a = A(a=value)
    # a = A(a=[-inf, nan, inf])
    string = a.json()
    # string = '{"a": [null, null, null]}' #                                         <———
    actual = A.model_validate_json(string).a
    # actual = A(a=[None, None, None]) 
    assert actual == value

The provided serialization schema returns Python built-in types (list), and Pydantic afterwards serializes them to a string. I need to specify that this step afterward allows NaN. plain_serializer_function_ser_schema accepts a return_schema argument, and CoreSchema has a property allow_inf_nan, but I could not get this to work and the documentation for core_schema has not enough usage examples yet.

How can I enable ser_json_inf_nan/allow_inf_nan for json_or_python_schema?