[yangjj-iso]

Summary

• Fix the intermittent assertion in TestDiag_ReusedConnBlackHoleReproducesClientTimeout. On unchanged main 97031ea8, 4 of 30 local runs failed because the returned timeout lacked the literal Client.Timeout exceeded while awaiting headers suffix.
• The diagnostic wraps the transport in traceTransport. In Go 1.20, net/http can return the request-context deadline error before its separate client-timeout timer marks the error for that suffix. Both outcomes are timeouts; the exact wording is not the behavior being tested.
• Assert net.Error.Timeout() instead, and explicitly verify that the second request reused the black-holed connection. Keep the real HTTP server, client timeout, and all test execution; do not skip or shorten the diagnostic. Production code is unchanged.

Testing

• Go 1.20.14, unchanged baseline: focused diagnostic -count=30 failed 4 times on the original string assertion.
• Final patch: the same focused diagnostic -count=30 passed (27.059s).
• go test ./... -count=1 passed (11.840s), with no skipped diagnostic.
• go test -race ./... -count=1 passed (13.050s); go vet ./... and git diff --check passed.
• Live sandbox e2e not run; this is a local HTTP transport diagnostic, not a production runtime change.

Breaking Changes

• None

Checklist

• Clearly described reproduction and motivation
• Test assertions updated without dropping timeout/reuse coverage
• Security impact considered: no production or access-control changes
• Backward compatibility considered
• No documentation change needed for this test-only correction.

Related failure reports appeared in #2452 and #2468. This PR fixes that independent test-reliability problem, not those SDK defects.