[UAL-PR-04] OpenAI streaming approval parity

[zhaowenzi]

Parent: #1316
Track: Slice A — OpenAI reference implementation

Goal

OpenAI streaming approval interrupt and approval continuation reach full parity with OpenAI's observed contract.

Acceptance Criteria

• Approval-required MCP calls in streaming mode emit, in order:
  1. response.output_item.added with mcp_approval_request
  2. response.output_item.done with the same item
  3. response.completed
  4. [DONE]
• Pending approval is never flattened into a placeholder mcp_call.
• previous_response_id + mcp_approval_response + stream=true is accepted as a 200 SSE response; approved continuation resumes the expected mcp_call.
• Approved continuation reuses the same max_tool_calls budget and malformed-argument tool-error behavior as the non-streaming path; no adapter-local 500s.
• Tests:
  • Streaming continuation-specific budget and malformed-argument coverage.
  • End-to-end local flow: interrupt → previous_response_id + mcp_approval_response → resumed mcp_call → final response.completed.
  • Manual parity check against real OpenAI for the streaming approval interrupt sequence.

Dependencies

• Unified agent loop for Responses API: design and implementation plan #1316 (umbrella design)
• Depends on UAL-PR-03

Out of Scope

• Generalized approval continuation across all routers (PR9).
• Router-common approval abstraction.