fix(codex): preserve streamed text when completed response has empty output

[sunlei]

Summary

Fix empty response failures on the Codex OAuth Responses path.

The Codex backend sometimes returns a response.completed event carrying an SDK Response object with status="completed" but output=[], even though earlier stream events (output_text.delta) already delivered the full assistant text. _build_response unconditionally returned this empty Response, discarding the collected text and triggering a misleading empty response retry loop.

Related:

• Telegram replies can be dropped when _channel_response_sent leaks across turns bub#101
• Bug: Responses API completed-empty payload is retried as empty response after a successful tool call bub#108

Root cause

In OpenAICodexProvider._build_response:

if isinstance(completed_response, Response):
    return completed_response  # unconditional — discards streamed text

The early return did not check whether output was actually populated. When output=[], the already-collected text from stream events was never used to reconstruct the response.

Changes

Narrow the early-return guard in _build_response so reconstruction from streamed data only happens under precise conditions:

1. output is non-empty → return as-is (output is good)
2. status != "completed" → return as-is (preserve error/incomplete metadata)
3. stream collected nothing → return as-is (avoid pointless reconstruction)

Only when all three checks fail (completed + empty output + stream has content) does the method fall through to the existing reconstruction logic via _build_response_output.

Test plan

• Added regression test: response.completed with a real SDK Response(output=[]) while stream collected text — verifies text is preserved
• All 127 existing tests pass unchanged