fix(openai): add strict flag to correctly use structured-output of openai provider

[thdesc]

• Ensure ChatOpenAI.bind_tools forwards its strict flag into _convert_to_openai_response_format(...), so the resulting response_format includes OpenAI’s "strict": true setting when requested.
• With this change, langchain.agents.create_agent can now drive OpenAI’s Structured Output mode purely by providing a Pydantic response_format: the factory already passes strict=True when it selects ProviderStrategy (see here) and that flag now propagates all the way to the OpenAI API.

Code to reproduce:

from pydantic import BaseModel, Field, ConfigDict

from langchain.agents import create_agent


class ProductRating(BaseModel):
    rating: int = Field(..., description="Rating from 1-5", ge=1, le=5)
    comment: str = Field(description="Review comment")
    model_config = ConfigDict(extra="forbid")

agent = create_agent(
    model="gpt-5-mini",
    tools=[],
    response_format=ProductRating,
    system_prompt="You are a helpful assistant that parses product reviews. Do not make any field or value up.",
)

response = agent.invoke({
    "messages": [{"role": "user", "content": "Toto"}]
})

This could raise an error because the structured output is not forced.

With this PR, the structured output is correctly set and no error will be raised with the provided code.

[thdesc]

This PR could solve this issue #34146.

[codspeed-hq]

CodSpeed Performance Report

Merging #34205 will not alter performance

_{Comparing thdesc:fix/strict_flag (5c9ea0f) with master (3ace4e3)}

Summary

✅ 6 untouched
⏩ 28 skipped¹

Footnotes

1. 28 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[ccurme]

Hi, thanks for submitting this!

We can do this without a breaking change (turning on strict mode for users when it was previously disabled) by supporting

ProviderStrategy(schema, strict=True)

as is done here: #34149

Closing in favor of that approach. We'll plan to release this in langchain 1.2.