GPT-oss streaming: message not captured when output_item.added is missing

[ashwing]

Summary

The streaming ResponseAccumulator relies on response.output_item.added to create in-flight state before processing deltas. GPT-oss (harmony path) doesn't emit this event for the message item after reasoning — it transitions directly:

response.output_item.added  → type: "reasoning"
response.reasoning_text.delta → ...
response.reasoning_text.done
response.output_item.done   → reasoning completed
response.output_text.done   → "HELLO"          ← no preceding output_item.added for message
response.output_item.done   → type: "message"
response.completed          → full output in payload

Qwen3 emits output_item.added for both reasoning and message items, so it works correctly.

Impact

• Streaming accumulation produces [Reasoning(...)] but drops the message text
• The response.completed payload still contains the full output (both items), so non-streaming and the final response are correct
• Only affects GPT-oss (harmony path); Qwen3 (non-harmony) works correctly

Possible fixes

1. Handle output_text.done without active message — auto-create a message item when text arrives with no in-flight message
2. Handle output_item.done with type "message" — extract the completed item from the done event payload
3. Extract output from response.completed — use the completed event's full output array as fallback when streaming accumulation missed items

Option 1 seems most consistent with the existing pattern. Would like to hear thoughts on preferred approach.

Discovered in

PR #59 cassette tests — reasoning-single-openai-gpt-oss-20b-streaming.yaml

cc @maralbahari @franciscojavierarceo