Tests: the two-rotation log-tail test allows PostgreSQL 18's repeated lock-wait line

[erikdarlingdata]

What failed

On #4909's CI run (Darling PG tests (2), run 36900036522), PgLogBurstAndRotationLiveTests.TwoRotationsInOneInterval_FinishTheMarkedFile_ReadTheNewestAndCountTheSkippedOne failed at Assert.Equal(1, WaitCount(cycle.Rows, pidA)): expected 1, actual 2. #4909 changes the blocking and deadlock read windows and doesn't touch the server-log tail or this test.

Why

WaitCount counts a backend's "still waiting" lock-wait lines. PostgreSQL 18 logs that line again when the waiting backend wakes on a latch before the lock is granted, so one wait can leave two or more lines under a loaded runner. The same behaviour made the lock-wait live test flaky before (#4895). The test pinned exactly one line per wait.

What changed (tests only)

• The finished marked file (A) and the newest file (C) must each have at least one "still waiting" line for their wait, and no line may be read twice: the lines for that wait must be distinct by timestamp and text. A re-read of a file repeats both; PostgreSQL 18's repeated line has a later timestamp and a longer wait.
• The skipped middle file (B) must still contribute nothing, and the files-skipped-by-rotation measurement must still be 1.

Testing

• Darling.Tests builds with 0 warnings and 0 errors.
• The test needs a store with stderr logging, so it runs on CI's PG shards.

CHANGELOG

None: a test-only change.