Python: fix(python): preserve CostControl budget failures in hosted Responses

[banibrata-de]

Motivation & Context

Foundry-hosted Responses agents currently lose CostControl budget classification when a streaming model response ends with an error event. The raw response headers identify budget_exceeded, but the OpenAI SDK raises a generic stream APIError, and the hosting adapter emits response.failed with server_error.

Description & Review Guide

• What are the major changes?
  • Preserve raw response headers when wrapping OpenAI streaming and non-streaming failures.
  • Detect allowlisted CostControl headers across the exception chain.
  • Emit response.failed.error.code as budget_exceeded.
  • Attach the allowlisted headers as compact JSON in response.metadata.costControl.
• What is the impact of these changes?
  • Responses clients can distinguish budget blocks from generic server failures for both streaming and non-streaming requests.
  • Other failures retain the existing server_error behavior.
• What do you want reviewers to focus on?
  • The exception-header lifetime across the raw streaming response boundary.
  • The CostControl allowlist and metadata size bound.

Related Issue

Fixes #8487

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue.
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after any language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

Copilot encountered an error: Your billing is not configured or you have Copilot licenses from multiple standalone organizations or enterprises. To use premium requests, select a billing entity via the GitHub site, under Settings > Copilot > Features.

[Evan Mattson (moonbox3)]

Could we preserve headers from the wrapped raw response before entering _open_event_stream? With Azure GenAI tracing enabled, raw_create_response is an AsyncStreamWrapper with no headers; the headers remain on its stream_async_iter, so _chat_client.py:940 records None and a later budget SSE error is still wrapped without x-ms-budget-cause. The hosted adapter then emits server_error instead of budget_exceeded. Could header extraction inspect the same inner raw response that _open_event_stream unwraps, and apply that to the retrieve path at _chat_client.py:880 too?