fix(apicompat): surface reasoning-only chat streams

[Fool0ntheHill]

Summary

Fix Chat Completions -> Responses streaming compatibility for DeepSeek-backed models when the stream has reasoning output but no visible content.

In this case the bridge used to close a Responses reasoning item but never create a visible message / response.output_text.* item, so Codex clients could receive a successful stream that still looked blank.

This patch synthesizes a visible fallback message from collected reasoning text only when all of these are true at stream finalization:

• no message item exists
• no visible text has been emitted
• reasoning text is non-blank
• no tool calls are present

Normal content streams and tool-call streams are unchanged.

Real sub2api SSE sample

Captured from a real sub2api /v1/responses streaming request routed to a DeepSeek-backed model, redacted and minimized.

Request shape:

{
  "model": "deepseek-v4-flash",
  "input": "只回复 pong",
  "stream": true,
  "max_output_tokens": 8,
  "reasoning": { "effort": "medium" }
}

Observed summary after this patch:
responses_output_text_delta=2
responses_reasoning_items=2
responses_message_items=2
responses_completed=1

Relevant SSE excerpt:
event: response.output_item.added
data: {"item":{"id":"<redacted>","status":"in_progress","summary":[],"type":"reasoning"},"output_index":0,"type":"response.output_item.added"}

event: response.reasoning_summary_text.delta
data: {"delta":"<reasoning delta>","item_id":"<redacted>","output_index":0,"summary_index":0,"type":"response.reasoning_summary_text.delta"}

event: response.output_item.done
data: {"item":{"id":"<redacted>","status":"completed","summary":[{"text":"<reasoning text>","type":"summary_text"}],"type":"reasoning"},"output_index":0,"type":"response.output_item.done"}

event: response.output_item.added
data: {"item":{"content":[{"text":"<fallback visible text>","type":"output_text"}],"id":"<redacted>","role":"assistant","status":"in_progress","type":"message"},"output_index":1,"type":"response.output_item.added"}

event: response.output_text.delta
data: {"content_index":0,"delta":"<fallback visible text delta>","item_id":"<redacted>","output_index":1,"type":"response.output_text.delta"}

event: response.output_text.done
data: {"content_index":0,"item_id":"<redacted>","output_index":1,"text":"<fallback visible text>","type":"response.output_text.done"}

event: response.completed
data: {"type":"response.completed","response":{"id":"<redacted>","object":"response","model":"deepseek-v4-flash","status":"completed","output":[{"type":"reasoning","id":"<redacted>","summary":[{"type":"summary_text","text":"<reasoning text>"}]},{"type":"message","id":"<redacted>","role":"assistant","content":[{"type":"output_text","text":"<fallback visible text>"}],"status":"completed"}]}}

Tests
Added lifecycle coverage for:
reasoning-only stream synthesizes visible text
blank reasoning does not synthesize visible text
normal content after reasoning does not duplicate fallback text
tool-call streams do not synthesize fallback visible text

Verified:
go test ./internal/pkg/apicompat -count=1
Note: I kept this PR scoped to the Chat Completions -> Responses bridge. On my local Windows environment, full go test ./... still exposes an unrelated existing internal/service WebSocket large-frame test issue, so I did not mix that fix into this PR.

[github-actions]

All contributors have signed the CLA. ✅
_{Posted by the CLA Assistant Lite bot.}

[Fool0ntheHill]

I have read the CLA Document and I hereby sign the CLA

[iuhl]

The failing backend-security job is caused by govulncheck running on Go 1.26.3. The reported stdlib vulnerabilities GO-2026-5039 and GO-2026-5037 are both fixed in Go 1.26.4. PR #3001 already applies the required Go 1.26.4 bump and its Security Scan passed. Once #3001 is merged or this branch rebases onto it, the backend-security failure should clear.