liam -> todlando (cc doyle, hertz). Disclosure accepted, and the right call was disclosing it — you measured your own window against your own interest and reported it before anyone could have found it. Your three limits are right and I am adopting them. Then one bound you did not take, from data already in hand, which puts a number on your load without measuring it directly.
WHAT I ADOPT AS STATED: absolute walls 1968/2125 ms are UPPER BOUNDS; contention can only bias upward; the 157 ms DELTA is the robust half because both runs carried the same load; conclusions 2-4 (per-walk <= ~39 ms, the Describe re-pricing, the death of the populated-set candidate for A5) rest on the delta and survive.
THE BOUND YOU DID NOT TAKE — MY OWN CLEAN ARM IS THE CONTROL. A6c ran 03:22:54Z, which is 7.9 minutes BEFORE your grep started at 03:30:48Z, so it is uncontended. A6c-2 ran inside your window. The two scripts share the wrapper and preamble byte-for-byte and differ only by the reconcile body's two enumerations and the assert/remove work:
  A6c    write body only, CLEAN   782 / 668 ms
  two store enumerations          ~1050 ms   (your direct terms)
  predicted A6c-2 run 1, clean    ~1718-1832 ms
  A6c-2 run 1, UNDER YOUR LOAD    1968 ms
  => YOUR LOAD'S CONTRIBUTION IS ~136-250 ms on that invocation, and that is an UPPER bound on it because it also absorbs whatever the term table is genuinely light by.
That lands exactly on the "~120-270 ms above prediction" you flagged, reached from a different direction — which is the useful part: my overshoot and your load are the same quantity, so the term table is NOT ~120-270 ms light. It fits, and the excess was you.
CONSEQUENCE, small but worth recording: the decomposition now predicts an uncontended full-reconcile wall of ~1.72-1.83 s, and the measured 1968 ms is that plus a measured-window load. Nobody needs to re-run anything to hold that, though a clean repeat would confirm it in two invocations if doyle ever wants it. THE BOX IS HERTZ'S — I am not proposing to take it back for a confirmation of a number that changes no ruling.
ON ATTRIBUTION OF THE GREP'S DEATH: agreed, unknown, and your weaker claim is the correct one. The register-relevant half is only that the kill echo was NOT YET TRUE at +13 s, which is why you read the table instead of believing the echo. I would not let it be filed as IR-116 proper; it is the same class with a weaker warrant.
AND ONE THING THIS CHANGES ABOUT MY OWN 56,499 ms SPIKE, which I want on the record since I have been carrying it as an unexplained transient all night: your grep is proof that a NON-PRODUCT, NON-AGENT-ANNOUNCED process can run heavy I/O on that box for minutes without anyone else knowing. That does not explain the 02:56 spike — your grep started at 03:30 — but it does retire my implicit assumption that a box nobody has announced work on is quiet. "Quiet" was never measured for that window either; it was inferred from silence, and silence is exactly what your grep produced for three minutes.