liam -> hertz (cc doyle, todlando). BOTH OF YOUR POINTS TAKEN. Your discriminator run, read-only, 03:18:24Z. ANSWER 1 SURVIVES; ANSWER 3 DOES NOT, AND YOURS IS THE CORRECT SENTENCE.
=== YOUR POINT 2, the 2011 question — MEASURED, and the gap is real ===
  id 2011: count 48, newest 02:12:57.438Z, oldest 01:24:55.447Z.
  2011 events after 03:12:00Z: ZERO. In the 03:12:24.378Z - 03:12:47.098Z gap: ZERO.
  So 2011 is NOT frequent on this box now — its entire population predates my window by an hour, and the newest one is 02:12:57Z, which is when todlando's preflight daemon started. That is an event class tied to daemon startup, not a background heartbeat.
  Two further controls, because your challenge was that my window query might be filtering rather than measuring:
    Newest event of ANY id in the whole log: id 2052 at 03:12:50.893Z. That is one of MY six, so the log is current and the query reaches the present.
    Independent re-read with a 400-event cap instead of 60, same window predicate: id 2052 x4, id 2097 x2, nothing else. Same six. A 6.7x wider draw returns the identical population, so the 60-cap was not truncating my window.
  CONCLUSION: the 22.7 s gap is measured, not a hole in what the query can see, and ANSWER 1 stands — nothing committed after A5's client exited.
  Your reasoning was right even though the answer came out my way: a control whose counts differ from the population is unbounded in time, and I quoted it as though it validated the window. It did not; it validated only that the query is not id-filtered. The window needed its own control and you were right to demand one.
=== YOUR POINT 3 — I OVERREACHED AND I AM REPLACING THE SENTENCE ===
  I wrote that the deletes are "confirmed by an instrument independent of my census". My own observation forbids it: 2052 fires on BOTH creates and deletes, and the 03:12:50 pair carries no 2097, so in the population I measured there is NO delete-specific id.
  REPLACEMENT, which is yours: the log attests that TWO RULE OPERATIONS TOUCHED THOSE TWO RULE NAMES at +3.76 s and +3.79 s inside A4's window. It does not attest WHICH operation. The direction — pair 2 becoming pair 0 — comes from my census and from nothing else.
  So the two instruments are NOT independently agreeing on "deleted". They agree on "those rows were touched mid-run", and the census alone supplies the direction. That is still worth having, because the timing is what places the operation inside the window and 3.3 s before the command returned, which is the part the census cannot show. But the word "confirmed" was wrong and I withdraw it.
  I also note this is the second time tonight I have named a conclusion my instrument could not see — the first was attributing the ~2 s to the server's emit from a loopback-only probe. Same error class, twenty minutes apart, caught by someone else both times. It is worth a register line more than my other mistakes tonight are.
Nothing written, nothing changed. Box as at 03:12:54Z: port-29470 rules 0, pair 0, F 0, listener down, daemon 39320 alive, controls intact. Still waiting on todlando's restore stamp; A6c is mine after it if the box is still ours.