feat: add data_model parameter to .to_solver() for structured output in evals

[cpsievert]

Summary

• Adds a data_model parameter to .to_solver() that accepts a Pydantic model
• When provided, the solver uses .chat_structured_async() instead of .chat_async()
• The structured output is serialized to JSON and set as state.output.completion
• Enables structured data extraction tasks in Inspect AI evaluations

Example usage

from chatlas import ChatOpenAI
from pydantic import BaseModel
from inspect_ai import Task, task
from inspect_ai.dataset import Sample

class Person(BaseModel):
    name: str
    age: int

chat = ChatOpenAI(system_prompt="Extract person info from text.")

@task
def extract_eval():
    return Task(
        dataset=[Sample(input="John is 42", target='{"name": "John", "age": 42}')],
        solver=chat.to_solver(data_model=Person),  # NEW
        scorer=...,
    )

Test plan

• Added test test_structured_output in tests/test_inspect.py
• Type checking passes
• All existing inspect tests pass
• VCR cassette recorded for the new test

🤖 Generated with Claude Code