fix: force IPv4 in health checker to prevent WSL2 DNS delays

[yasinBursali]

What

Force IPv4 connections in the aiohttp health check session by creating the TCP connector with family=socket.AF_INET.

Why

On WSL2/Docker Desktop, the default aiohttp connector attempts both IPv6 (AAAA) and IPv4 (A) DNS resolution. Docker service name AAAA queries timeout after ~8 seconds before falling back to IPv4, inflating response_time_ms to ~8016ms for most services. Docker internal networking is exclusively IPv4 — AAAA queries always fail.

How

In _get_aio_session(), changed the session creation from:

aiohttp.ClientSession(timeout=_HEALTH_TIMEOUT)

to:

aiohttp.ClientSession(
    timeout=_HEALTH_TIMEOUT,
    connector=aiohttp.TCPConnector(family=socket.AF_INET),
)

Testing

• python -m py_compile passes
• pytest: 374/384 pass (10 pre-existing failures on main, 0 regressions)
• Critique Guardian: APPROVED WITH WARNINGS

Review

Critique Guardian verdict: ⚠️ APPROVED WITH WARNINGS

• Non-blocking warning: the shared httpx.AsyncClient (used for llama-server metrics) does not force IPv4. Single-host with 5s timeout — much smaller impact than the 17-service health poll. Recommended as follow-up.

Known Considerations

The httpx client for llama-server metrics is not covered by this fix. On WSL2, llama-server requests may still see a one-time ~5s DNS delay. This is a minor follow-up — the critical fix is the health poll (17 services × 8s = 136s delay eliminated).

Platform Impact

• macOS: No impact — Docker Desktop for Mac uses IPv4 internally
• Linux: No impact — Docker bridge networking is IPv4
• Windows/WSL2: Fixes the ~8016ms response_time_ms inflation