Plugin conformance tests are flaky when CloudWatch logs are not yet visible

[zhongkechen]

Summary

The conformance (plugin) job can fail because expected plugin log records are not returned by CloudWatch, even though the durable execution and history validation complete successfully.

Evidence

• Previous successful run: 33907972521 at commit 4ce54bcf.
• Failed run: 33922277509 at commit 676c75a.
• Attempt 1, job 101185043031: test 10-11 (plugin_nested_parent_linkage) failed because both expected operation-end records matched 0 logs: the inner step record and the child-context record.
• Retry attempt 2, job 101188790974: test 10-11 passed, but unrelated test 10-7 (plugin_terminal_failure) failed because its invocation-start and invocation-end records matched 0 logs.

The failure moving between tests on the same commit strongly suggests CloudWatch ingestion/filter timing or log-stream lookup flakiness rather than a deterministic SDK regression. The diff from the previous successful commit only changes OTel release metadata and does not affect the regular core plugin suite.

Suggested investigation

• Increase or make configurable the CloudWatch log propagation/polling window in the conformance runner.
• Log the resolved log group, execution ARN filter, event count, and final retrieved records when ExpectedLogs fails.
• Pin the conformance runner to a known commit in CI instead of installing @main.
• Re-run the plugin suite repeatedly against the same commit to confirm the failure rate.

Relevant workflow: .github/workflows/conformance-tests.yml, step Run conformance suite.