fix various RootModel serialization issues

[davidhewitt]

Change Summary

Adds more testing and fixes to two bugs with RootModel:

• field_serializer not respected with serialize_as_any in 2.12
• RootModel no longer serializes incompatible types with a "root" field, e.g. Path

Related issue number

Xref pydantic/pydantic#12379
Fixes pydantic/pydantic#8963

Checklist

• Unit tests for the changes exist
• Documentation reflects the changes where applicable
• Pydantic tests pass with this pydantic-core (except for expected changes)
• My PR is ready to review, please add a comment including the phrase "please review" to assign reviewers

[davidhewitt]

I wonder if it's a bug that the fallback uses the model_extra? (This preserved existing behaviour.)

[Viicos]

Do you have a MRE that hits this code path? I'm having trouble understanding what is this logic for here.

[davidhewitt]

Looks like only field serializers care about the extra.model, I'm going to tidy up and not worry.

[davidhewitt]

(this would be a case that warns about expecting a model and not getting it, and then somehow falling to type inference which calls a field serializer. I'm pretty sure that's impossible.)

[Viicos]

Suggested change

	/// apply any type inference. (`Model` serialization is strongly coupled with its child
	/// serializer, and in the few cases where `serialize_as_any` applies, it is handled here.)
	/// apply any type inference. (`Model` serialization is strongly coupled with its child
	/// serializer, and in the few cases where `serialize_as_any` applies, it is handled here.)

I wondered if it actually made sense to have the 'model' core schema having an inner 'schema' item, that should be of type 'model-fields' although it is typed as any CoreSchema in the TypedDict definitions 🤔

[davidhewitt]

It exists for "wrap" model serializers (and maybe "before"), in both of those cases the inference should not be used there either.

If I was improving this (maybe another day):

• I would make root model schema completely separate, it works so differently.
• model schema would directly take the fields, no model_fields schema
• model schema would contain any model validators / serializers as special keys in the schema

This would also be much better for performance, the critical model pathways could then just produce the model values without having to go through CombinedValidator / CombinedSerializer abstractions

[codspeed-hq]

CodSpeed Performance Report

Merging #1836 will not alter performance

_{Comparing dh/root-model-fixes (309f09a) with main (d90a93b)}

Summary

✅ 163 untouched

[davidhewitt]

The failing integration test is not a concern IMO, the case currently raises an AttributeError and the new logic here has downgraded that to a warning and inference fallback, which is more consistent with how serialization works everywhere else.