Audit and remove StreamingProxy subsystem if obsolete

[eanzhao]

Background

The repo contains a complete StreamingProxy subsystem at agents/Aevatar.GAgents.StreamingProxy/ plus agents/Aevatar.GAgents.StreamingProxyParticipant/. It comes with:

• StreamingProxyGAgent, StreamingProxyParticipantGAgent
• Full endpoints (StreamingProxyEndpoints.cs), room session orchestration, SSE writer
• Projection pipeline (StreamingProxyRoomSessionEventProjector, *CurrentStateProjector)
• Coordinator (StreamingProxyNyxParticipantCoordinator.cs)
• ADR-0015 (docs/adr/0015-agui-sse-projection-session-pipeline.md)
• ~25 source + test files

Now that /v1/responses (and the planned /v1/messages Path B) implement streaming directly on top of LlmSessionGAgent + NyxIdLLMProvider, the StreamingProxy may be:

• Obsolete duplicate streaming surface
• Or still load-bearing for multi-party rooms / NyxID participant coordination

Either way it needs an explicit audit because two parallel SSE paths violate CLAUDE.md "单一主干".

Audit questions

1. Who calls StreamingProxyEndpoints? grep for live callers — frontend (tools/Aevatar.Tools.Cli/Frontend/src/runtime/), demos, production?
2. What business capability does it provide that LlmSessionGAgent + /v1/responses SSE cannot?
   • Multi-participant rooms?
   • Persisted streaming history?
   • Cross-client fanout?
3. Is StreamingProxyNyxParticipantCoordinator still on a hot path? It was central to the older NyxID chat flow — check if NyxIdChat agent still depends on it
4. ADR-0015 status — does it accurately describe today's design, or is it stale?

Decision matrix

Audit outcome	Action
No live callers + capability subsumed by /v1/responses	Delete entire subsystem. Supersede ADR-0015. GAgent count: 19 -> 17
Used by frontend playground only	Delete prod path, keep minimal demo or move to demos/
Multi-participant rooms still needed	Keep, but document the boundary with /v1/responses SSE clearly in a new ADR
Coordinator still load-bearing for NyxID chat	Keep that piece, delete the rest, narrow the public surface

Out of scope

• Re-implementing multi-participant features on top of LlmSessionGAgent if audit says they're still needed — separate proposal
• Frontend rewrite — depends on audit outcome

Acceptance

• Written audit report with answers to the 4 questions above
• Either a deletion PR (preferred if audit says obsolete) or an ADR explaining what stays and why
• No parallel SSE paths left undocumented