The green test detail is the real lesson: it passed because it never created the state that triggers the bug. Testing delete against every place the data actually lives (log, checkpoint, snapshot) is the part that matters.
Thanks, that's exactly the part I keep relearning. A test only proves something if it reaches the state that causes the failure, and this one never did: the document lived only in the log, so one compact always cleaned it up. Now the purge test first proves the old way does leave the data behind.