fix(responses): reassemble split SSE event/data frames before streaming

[davidwushi1145]

Summary

Reassemble split event: + data: SSE fragments in the /v1/responses handler before flushing them downstream, so line-oriented upstream executors no longer emit malformed Responses stream events.

Problem

In the current Responses streaming path, some upstream executors can emit SSE line-by-line rather than as complete framed events. When event: and data: arrive as separate chunks, the handler was incorrectly closing each chunk as its own SSE event, producing malformed output like:

event: response.created

data: {...}

Downstream clients such as OpenClaw then fail to parse the stream and report JSON errors like:

Could not parse message into JSON
Unexpected end of JSON input

This was not limited to a single executor implementation. Any line-oriented openai-response stream routed through the Responses handler could hit the same framing bug.

Fix

Added a small stateful responsesSSEFramer in the Responses handler layer.

It now:

• buffers standalone event: lines until the matching data: line arrives
• preserves already complete SSE events unchanged
• preserves valid data-only SSE events
• drops delimiter-only leftovers safely on flush
• handles both \n\n and \r\n\r\n framing
• inserts the missing line break when adjacent SSE fields arrive in separate chunks without a trailing newline

Also added regression coverage for:

• split event: + data: chunks in Responses stream forwarding
• valid full-event chunk passthrough
• terminal error streaming path compatibility
• bootstrap / ExecuteStreamWithAuthManager acceptance of split openai-response SSE event lines from non-Codex executors

Testing

• Verified locally with Go 1.26.1:
  • /tmp/go1.26.1/go/bin/go test ./sdk/api/handlers/openai -count=1
  • /tmp/go1.26.1/go/bin/go test ./sdk/api/handlers -count=1
  • /tmp/go1.26.1/go/bin/go test ./sdk/api/handlers/... -count=1
  • /tmp/go1.26.1/go/bin/go vet ./sdk/api/handlers/...
• Verified with a temporary patched CLIProxyAPI instance on 127.0.0.1:8317:
  • /v1/models returned 200
  • /v1/responses non-stream returned 200
  • /v1/responses stream returned valid combined event: + data: frames
  • confirmed malformed event: response.created\n\ndata: output no longer appeared

Fixes #2484

[gemini-code-assist]

Code Review

This pull request introduces a responsesSSEFramer to correctly reassemble split Server-Sent Events (SSE) in the OpenAI responses handler, ensuring that partial chunks are buffered until a complete frame is formed. It also updates line-ending handling and adds unit tests for split-event scenarios. The review feedback highlights potential issues with eager flushing of incomplete frames, redundant newline injections when chunks already contain line breaks, and performance optimizations for line-prefix checks to avoid unnecessary allocations.