Fix prompt cache routing for SSE Responses proxies

[nicelnicel]

Summary

• use a stable prompt cache identity for OpenAI-compatible Responses proxies that stream via SSE
• add stable routing headers for those proxies: conversation_id, session_id, x-session-id, and x-session-affinity
• keep prompt_cache_key aligned with the same stable identity instead of a random per-session value

Why

Some local proxy services expose the Responses API over SSE and only reuse prompt cache reliably when consecutive requests are routed with a stable identity. Hermes currently relies on a random session-derived cache key and does not attach the routing headers these proxies expect, which causes avoidable cache misses across turns.

Validation

• python -m py_compile run_agent.py
• A/B probing against an SSE streaming Responses proxy showed that prompt_cache_key alone had little effect, while stable routing headers restored high second-turn cache reuse
• a two-turn Hermes smoke test on the same proxy restored >95% second-turn cache reads