[Candidate / unverified] Solution for cc86dc1b-994c-45e1-9087-571f77c587ff: Keep one long-lived httpx-backed SDK client per process or event loop; do not ca

## Support semantics This is a **candidate / unverified** Solution from the KFA Solution Factory Tier-2 gate. It is source-grounded research material, not an Outcome, and does **not** claim independent reproduction or confirmed field success. ## Cause httpx transitions the client to CLOSED on context exit or explicit close; send() guards on _state == ClientState.CLOSED. ## Steps 1. Search code for with OpenAI(...) as client or await client.close() while later calls reuse the same instance. 2. Refactor to a module-level or app-scoped client created once and reused. 3. If using a client cache, ensure eviction does not close clients still referenced elsewhere. ## Expected result Subsequent API calls use an OPEN client state and no longer raise this RuntimeError. ## Misleading approaches - Retrying the same closed client instance. - Creating a client per request inside a hot loop without reusing connections (performance issue, though not this exact error). ## Limitations - Older openai-python versions may wrap send failures as APIConnectionError instead of surfacing RuntimeError verbatim. - httpx2 message text not independently verified in this pass. ## Unknowns - Exact SDK version mapping when exceptions are wrapped versus propagated. ## Evidence claims - (documented_platform_behavior) httpx raises RuntimeError('Cannot send a request, as the client has been closed.') when send is called on a closed client. — https://raw.githubusercontent.com/encode/httpx/master/httpx/_client.py