fix: wrap optimization handler responses in SSE streaming events

[Klausc06]

Summary

Optimization handlers (fast-prefix detection, quota check, title generation skip, suggestion skip, filepath mock) previously returned a plain JSON MessagesResponse directly. But Claude Code always sends stream=True, so the client expects Anthropic SSE event stream format. This mismatch caused Stream completed without receiving message_start errors and forced non-streaming fallbacks on every optimized request.

This PR adds _messages_response_to_sse_stream() which converts a MessagesResponse into a proper SSE event stream (message_start → content_block_start → content_block_delta → content_block_stop → message_delta → message_stop), and wraps optimization results through it.

中文说明

优化处理器（前缀匹配、配额检查、标题生成跳过、建议跳过等）之前直接返回了普通的 JSON MessagesResponse。但 Claude Code 客户端始终发送 stream=True，期望收到 Anthropic SSE 事件流格式。这个不匹配导致了 Stream completed without receiving message_start 错误，并且每次优化命中的请求都会触发非流式降级。

本次 PR 新增了 _messages_response_to_sse_stream() 函数，将 MessagesResponse 转换为标准的 SSE 事件流，并包裹所有优化处理器的返回结果。

Test plan

• uv run pytest tests/api/test_routes_optimizations.py passes
• Optimization requests return text/event-stream; charset=utf-8 content type
• SSE stream parses correctly with message_start / content_block_* / message_delta / message_stop events
• No Stream completed without receiving message_start errors in Claude Code client logs

Files changed

File	Δ
api/services.py	+81/-1
tests/api/test_routes_optimizations.py	+25/-2

🤖 Generated with Claude Code