fix(ai): preserve response reasoning items

[rekram1-node]

Before

• Mixed raw and summary deltas could share incomplete indexing and state.
• Raw reasoning could be mislabeled as summary.
• Multiple summary entries could become separate visible blocks or be flattened during replay.

Raw and summary streaming therefore lost the exact provider structure needed for reliable continuation.

After

• One visible reasoning block is emitted per provider reasoning item.
• Completed summary is preferred for display, followed by completed raw content and then streamed fallback.
• Exact summary and content arrays plus encrypted state remain in opaque provider replay state.
• OpenAI replays the completed item unchanged.
• Shared Open Responses projects the stored item to its formal input schema, preserving ordered summaries and encrypted content while omitting raw content.