[Bug]: Pydantic aliases gain unwanted internal-name properties #237

Version of the library

0.64.0, current main d34b372556392f9bfb49e8590d2c4f253c103309; Python 3.13.13 / Pydantic 2.14.0 / macOS arm64.

Describe the bug

When using the documented Pydantic-model schema integration with an optional aliased field, schema normalization adds a second property under the internal Python field name. Repairs can consequently return data that the original Pydantic model rejects, even when extra="forbid" is set.

schema_from_input iterates model_fields by Python name when adding defaults, while Pydantic's validation JSON Schema uses aliases. In the example below the schema initially contains only displayName; normalization adds display_name, and repair inserts that unwanted field.

This affects typed structured-output processing: the repaired dict cannot pass the same model's model_validate step. Found by code inspection and reproduced with the actual Pydantic model/API; this is not a report of production usage.

How to reproduce

from pydantic import BaseModel, ConfigDict, Field
from json_repair import loads

class Payload(BaseModel):
    model_config = ConfigDict(extra="forbid")
    display_name: str = Field(default="unknown", alias="displayName")

repaired = loads('{displayName: "Alice"}', schema=Payload)
print(repaired)
Payload.model_validate(repaired)

Actual:

{'displayName': 'Alice', 'display_name': 'unknown'}
pydantic_core.ValidationError: 1 validation error for Payload
display_name
  Extra inputs are not permitted [type=extra_forbidden, input_value='unknown', input_type=str]

The same problem occurs when valid JSON needs scalar coercion, e.g. '{"displayName": 42}'.

Expected behavior

Return {'displayName': 'Alice'}, which validates against Payload, without adding a Python-name property absent from the original JSON Schema. Default/default-factory augmentation should target the property's validation alias when present, preserving Pydantic's schema names rather than widening extra="forbid".

I intend to fix the field-to-property mapping and add a focused regression covering repaired output and downstream model validation. Codex assisted with investigation and local reproduction.