testing: drop SO_REUSEADDR from the port-availability probe

[chen3feng]

Problem

testing::PickAvailablePort() / PickAvailableEndpoint() probe a candidate port by binding to it. The probe set SO_REUSEADDR before binding 0.0.0.0:port. On macOS/BSD SO_REUSEADDR lets that bind succeed even while another socket already holds 127.0.0.1:port — so the probe reports an occupied port as free. The caller then binds its real server to the same port and fails:

W socket.cc:150] Cannot bind socket to [127.0.0.1:5136]. : Address already in use [48]
F server.cc:226] Check failed: !!fd Cannot create listener.

This only bites under parallel test execution, where sibling tests hold ports concurrently (observed on logging_test running alongside server_test / server_group_test / integration_test). In isolation the test always passes.

Fix

Drop SO_REUSEADDR from the probe. An availability probe must observe the port exactly as a fresh server bind would; with the override it doesn't. Without it, an occupied port is correctly rejected and PickAvailablePort() just retries another random port (no functional loss).

Verification

• Deterministic repro: occupy 127.0.0.1:P (no reuse); probe 0.0.0.0:P with SO_REUSEADDR reports AVAILABLE (false positive), without it reports occupied (correct).
• Stress: ran the four port-grabbing rpc tests in parallel across many rounds — 0 flakes after the fix.

🤖 Generated with Claude Code