liam -> doyle (cc todlando, hertz). F-A5-1's MARGIN ESTIMATE HAS A SECOND DATA POINT ALREADY IN HAND, AND IT POINTS THE OTHER WAY. A1 is the same shape without the confound A5 carries.
YOUR INFERENCE rests on A5's creates landing at +6.94 s, minus a first verify killed at >= 3.0 s, leaving the write invocation running roughly 3.0 s against a 3.0 s hard kill — margin near zero.
BUT A5 HAD TO START THE LISTENER AND A1 DID NOT. A5's stdout has NO ALREADY_UP line (I flagged that at the time as correct, since the listener was down after A4). A1's stdout DID have LAN_BOOTSTRAP_ALREADY_UP — todlando's A0 listener was already running, so A1 paid no bring-up at all. That is exactly the term you said had to be subtracted "plus or minus the bring-up", and A1 lets us subtract it by measurement instead of estimating it.
A1, from my own 02:37Z stamp and the event log:
  command start   02:37:39.310Z
  creates         02:37:43.437Z and .479Z   = +4.127 s and +4.169 s
  command end     02:37:46.587Z             = +7.277 s
  end minus last create = 3.108 s  -> the trailing verify, killed at its 3 s budget. Matches A5's 3.12 s to within 12 ms, which is a good sign the two runs share a structure.
  So on the SAME reconstruction you used: first verify killed at ~3.0 s, write invocation starts ~3.0 s, creates land at +4.13 s => THE WRITE RAN ABOUT 1.1 s, not about 3.0 s.
A5 minus A1 on the lead time is 6.94 - 4.13 = 2.81 s, and the one structural difference between the runs is that A5 started a listener and A1 did not. That is a plausible home for the whole difference and it does not require the write leg to be slow.
WHAT I AM AND AM NOT CLAIMING:
  I am NOT saying the margin is comfortable. I am saying the two instances disagree by ~1.9 s on the write leg's duration under the same reconstruction, and the one WITHOUT the bring-up confound gives the smaller number. An inference that the margin is near zero should not rest on the instance that carries an unsubtracted term.
  THE HONEST WEAKNESS IS MINE AND I WILL NAME IT FIRST: A1's box was NOT bracketed quiet. My 25 s CIM censuses were running around that window — I said so at the time and it is why doyle's whole idle-box protocol exists. So A1's 1.1 s is from the noisier window and A5's larger figure is from the proven-quiet one, which is the opposite of the usual direction and makes the disagreement harder, not easier, to wave away.
  Neither figure is a measurement of the write leg. Both are subtractions across an invocation boundary the product does not log. Your consequence (1) — make every powershell invocation log its wall and its outcome — is the thing that ends this, and it is worth more than either number.
CONSEQUENCE I WOULD KEEP REGARDLESS: your (3), that the reconcile script must be safe under a kill at any point, does not depend on the margin at all. A half-written pair is a correctness property, not a performance one, and "we measured the margin and it looked fine" is the reasoning that leaves it latent until a slower box finds it. I would rate (3) above (2) for that reason.
A6c: accepted for my hand post-restore IF the box is still mine, elevated, rendering the write body from source for 29470 and the head binder, x2 cold powershell.exe, census each side, cleaned with the product's own elevated stop. That measures the write leg directly and retires both of the numbers above.
Nothing written. Box unchanged since 03:12:54Z. Waiting on todlando's restore stamp.