Fix ci failure issue

[mwfj]

Fix flaky ObsConnMetrics: net.connections.accepted monotonic on slow CI runners

Closes #45.

Summary

TestNetAcceptedIsMonotonic in test/observability/observability_connection_metrics_test.h was failing intermittently on GitHub Actions shared runners with delta=4.000000 expected>=5. The test opened 5 TCP connections in a tight loop with immediate close() and a 20 ms inter-connection sleep, then polled the accepted counter for up to 2 s.

On slow / contended runners one of the connect+close cycles completed before the accept dispatcher processed the connection — the TCP RST arrived before (or nearly simultaneously with) the accept callback that increments reactor.net.connections.accepted. The counter only reached 4 and the test failed.

Fix

Replace the bare connect → close loop with connect → send GET /h → drain response → close (Option 2 from the issue). Driving a full HTTP request/response cycle per connection forces the accept callback to complete before the close, making the +1 deterministic instead of timing-dependent.

The post-loop snapshot poll deadline is also bumped from 2 s to 5 s as a backstop for the small dispatcher-thread visibility gap between the response draining on the client side and the Snapshot() picking up the counter bump.

The Connection: close request header keeps each connection short-lived so the test still exercises the accept-counter path under repeated open/close (which was the original intent), without changing the assertion shape (delta >= N).

Changes

• test/observability/observability_connection_metrics_test.h — TestNetAcceptedIsMonotonic now sends a minimal HTTP request and drains the response on each of the N connections; poll deadline raised from 2 s → 5 s; stale "accept-time +1 fires regardless of any later HTTP traffic" comment replaced with the actual race rationale.

No production code changed. Test-only fix, ~15-line diff in a single file.

Test plan

• make -j4 test_runner — clean build.
• ./test_runner obs_connection_metrics × 20 — all 20 runs pass ([PASS] ObsConnMetrics: net.connections.accepted monotonic over N opens).
• Stress under heavy CPU contention (4 background busy-loop threads simulating loaded CI) × 10 — 10/10 passes.
• CI matrix: gcc / clang / ASan+UBSan / TSan-rest (auto-picked up by obs_connection_metrics membership in the existing all-suites jobs).

Notes for reviewer

• This is in the same class of CI timing sensitivity as Flaky test: ObsConnMetrics H1 protocol gauge read races AttachTransportObservability #43 (ObsConnMetrics protocol gauge race).
• The DrainSocket(fd, 500) call is a 500 ms ceiling, not a fixed sleep — it returns as soon as the response bytes arrive, so the per-connection cost on a healthy runner is microseconds, not 500 ms.
• The catch envelope already on TestFramework::RecordTest covers any exception escapes from SendAll / DrainSocket; the explicit send failed short-circuit just gives a more diagnostic failure message than a generic catch.

[gemini-code-assist]

Code Review

This pull request updates the TestNetAcceptedIsMonotonic test in observability_connection_metrics_test.h to prevent race conditions on slow CI runners. Instead of immediately closing the socket after connecting, the test now sends a full HTTP request and drains the response before closing the connection. Additionally, the polling deadline has been increased from 2 seconds to 5 seconds. I have no feedback to provide as there are no review comments.