stream_events: synthesize SSE for locally executed tools

[leseb]

Context

Praxis core #988 and AI #756 establish one logical client-visible Responses SSE lifecycle across IRR inference rounds. Backend-originated events, response identity normalization, terminal-event handling, and loop continuation are therefore outside this issue.

Server-side execution filters such as openai_mcp_dispatch and openai_web_search also create or update tool activity locally in ResponsesState. Those changes do not necessarily originate in the model backend's SSE stream and still need Responses-compatible client-visible events.

Scope

Synthesize SSE only for tool activity generated locally by server-side execution filters:

• MCP call and result progress events.
• Web-search progress events that were not already emitted by the upstream model.
• Shared event injection and ordering needed by those filters.

File-search-specific progress events remain tracked by #313. MCP tool-listing failure events remain tracked by #320.

Ownership

• Event synthesis and logical sequence allocation belong to openai_stream_events (Filter 6: stream_events #25).
• Tool execution remains owned by the corresponding MCP, web-search, or file-search filter.
• openai_agentic_loop (Filter 7: agentic_loop #26) remains the loop controller and does not classify or execute tools.
• Logical multi-round streaming is provided by feat(openai): stream Responses across IRR agentic rounds #756 on top of feat(filter): resume IRR iterations within streaming responses praxis#988.

Acceptance criteria

• Locally generated MCP activity is emitted exactly once using valid Responses event shapes.
• Locally generated web-search activity not already present in the upstream stream is emitted exactly once.
• Synthesized events use the logical response ID and valid output indices and sequence numbers.
• Synthesized events are emitted before events from the resumed model response.
• No duplicate response lifecycle events or duplicate output items are produced across rounds.
• The final response.completed snapshot agrees with the items previously emitted to the client.
• Non-streaming behavior remains unchanged.
• End-to-end tests assert event type, ordering, identity, sequence numbering, and deduplication for model → tool → model flows.

Out of scope

• Backend-originated SSE parsing or pass-through.
• Logical response lifecycle normalization across IRR rounds (feat(openai): stream Responses across IRR agentic rounds #756).
• Stateful buffering and fail-closed stream handling covered by feat(openai): stream Responses across IRR agentic rounds #756.
• File-search-specific progress events (file_search: SSE event emission for progress streaming #313).
• mcp_list_tools.failed emission (stream_events: emit mcp_list_tools.failed SSE event on MCP tool resolution failure #320).

Relationships

• Depends on feat(openai): stream Responses across IRR agentic rounds #756
• Child of Filter 7: agentic_loop #26
• Related: Filter 6: stream_events #25, file_search: SSE event emission for progress streaming #313, stream_events: emit mcp_list_tools.failed SSE event on MCP tool resolution failure #320
• Part of Epic: Responses API and Agentic Loop Orchestration #93