fix(providers): disable AsyncOpenAI keep-alive reuse to avoid SSL UNEXPECTED_EOF on idle-reaping gateways

[raymondginger2018-sudo]

Summary

This is the remaining, independent half of #198: disable keep-alive connection reuse in OpenAICompatProvider’s AsyncOpenAI client so third-party gateways (e.g. scnet) that reap idle connections don’t hand back a dead connection, which surfaces as SSL: UNEXPECTED_EOF_WHILE_READING.

The GLM-5.2 catalog entry from #198 was already merged upstream as a standalone commit credited to me (329e98f0), so this PR contains only the keep-alive fix.

Changes

• core/providers/openai_compat.py: pass http_client=httpx.AsyncClient(limits=httpx.Limits(max_keepalive_connections=0)) to AsyncOpenAI (+8 lines). New connection per request (~50–100 ms handshake) is negligible for LLM-second-scale calls.

Tests

• python -m py_compile core/providers/openai_compat.py ✓
• module import smoke ✓ (httpx available)
• full CI pending

Notes

• Based on current upstream/main (c0a6a3c); single file, no CLAUDE.md / .gitleaksignore changes.

Related: #198

[Zongwei9888]

Closing for the same reason as the keep-alive half of #198: disabling keep-alive here is hard-coded for every OpenAI-compatible connection (OpenRouter, DeepSeek, Ollama, …), not just the gateway that reaps idle connections, and the custom httpx.AsyncClient is never closed.

The shape that would be accepted is a connection-level option — for example httpKeepAlive: false on ConnectionProfileConfig — read in OpenAICompatProvider.__init__, defaulting to the current behaviour, with the client closed on provider shutdown. A PR in that shape is welcome.