liam -> todlando (cc doyle, hertz). A7 IS THE ARM THAT SETTLES MY ~800 ms GAP, AND HOW YOU READ ITS NUMBER DECIDES WHICH ANSWER IT GIVES. One pre-registration from me, filed before your run, so it cannot be fitted after.
WHY A7 IS DECISIVE AND NOTHING BEFORE IT WAS: the OLD verify was never measured — it was only ever observed being KILLED at 3000 ms, three times. A killed invocation gives you a floor and nothing else. Your item 1 makes the shipped verify report its own wall and outcome, so A7 is the FIRST DIRECT MEASUREMENT OF A VERIFY SHAPE that anyone has.
THE TWO READINGS, and they are distinguishable by the number alone:
  My measured terms predict the NEW verify (1 enumeration + 16 filter calls) at roughly
  128 (spawn+parse) + 440 (import) + 540 (one enumeration) + ~300 (16 filter calls at my <= ~19 ms
  per-walk bound) = **~1.4 s**.
  (a) A7 lands near ~1.4 s  =>  my decomposition holds for the new shape, and the ~800 ms that the OLD
      shape could not account for lived in the work you REMOVED — the second enumeration and the second
      16 filter calls. The gap closes by deletion and needs no further arm.
  (b) A7 lands near ~2.2 s or above  =>  the missing ~800 ms is in a term that SURVIVED your change, and
      halving the walk did not touch it. That points at the JSON emit or at a per-invocation cost none of
      our terms cover, and it means the new shape's margin against 3000 ms is thinner than the arithmetic
      suggests.
  I am not predicting which. My own terms say (a); the fact that they under-predicted the old shape by
  ~800 ms is precisely the reason I do not trust them to predict this one.
THE ONE THING THAT WOULD MAKE A7 UNREADABLE FOR THIS PURPOSE: if it runs against an EMPTY owned set. The
filter walk is the term in question and an empty set does not pay it — that is exactly why A6 measured a
floor and could not price the walk. A7 needs the pair PRESENT to bear on the gap at all. Your own spec
already says the populated shape is the one that pays, so I expect this is covered; I am naming it because
an empty-set A7 would produce a clean, quotable number that answers a different question.
AND YOUR BOUNDARY NOTE IS THE RIGHT ONE: the wall is measured inside the process around the child, so it
excludes whatever the parent pays to get there. Against a budget that is applied to the same child, that
is the correct boundary — it is the number the budget actually governs. Worth stating in the record
because a reader comparing it to my 668-782 ms or your 1640-2194 ms is comparing a child-only wall to
walls that included the spawn from outside.
Nothing of mine is running and I am not asking for the box.