fix: retry server_is_overloaded for OpenAI Responses stream failures

[amitksingh1490]

Summary

Make OpenAI Responses stream overload failures retryable by preserving structured upstream error codes and classifying server_is_overloaded as retryable.

Context

OpenAI Responses streaming failures were being surfaced as formatted strings (Upstream response failed: ...), which dropped structured provider error metadata. Because of that, retry classification could not reliably identify the server_is_overloaded code from stream failures and the request was not retried.

Changes

• Preserved response.failed stream errors as structured OpenAI DTO errors instead of debug-only strings.
• Added OpenAI overloaded error classification in retry mapping for server_is_overloaded.
• Kept transport-code retry matching centralized with TRANSPORT_ERROR_CODES.
• Added regression tests for both stream error-code preservation and retryability behavior.

Key Implementation Details

• Added into_response_failed_error(...) to convert ResponseFailedEvent into forge_app::dto::openai::Error::Response, retaining code and message fields from upstream payloads.
• Updated stream event handling to use structured conversion for ResponseFailed events.
• Extended retry logic with a typed RetryableApiErrorCode classifier and recursive code lookup through nested OpenAI error payloads.
• Added explicit overloaded-code constant (server_is_overloaded) and transport constant usage (TRANSPORT_ERROR_CODES).

Use Cases

• If OpenAI Responses stream emits:
  • code: "server_is_overloaded"
  • message: "Our servers are currently overloaded. Please try again later."
    then Forge now marks the failure as retryable.
• Existing transport retry behavior for ERR_STREAM_PREMATURE_CLOSE, ECONNRESET, and ETIMEDOUT remains intact.

Testing

cargo test -p forge_repo test_stream_with_response_failed_preserves_error_code
cargo test -p forge_repo test_openai_server_overloaded_error_is_retryable
cargo test -p forge_repo test_into_retry_with_transport_errors

Links

• Related issues: N/A