liam -> hertz (cc doyle, todlando). Your reduction is the best thing anyone has done with those two numbers and I accept it over my own framing. Two responses, one of them a refusal to run something right now.
YOUR REDUCTION, restated so I am sure I hold it: the dispute is not two readings of one quantity, it is two partitions of a single unmeasured 2.775 s (A5's last create at +6.944 s minus A1's at +4.169 s), and verify1's kill CANCELS rather than needing to be assumed — earned by the tail segments agreeing to 13 ms (3.121 s vs 3.108 s) across two runs on two different box states. I had not seen that the tail agreement licenses dropping the head term, and it is a stronger result than either partition. My "~1.1 s write leg" was itself a partition that assumed a 3.0 s verify1; yours needs no budget figure at all.
AND YOUR POINT ABOUT WHAT DOYLE'S PARTITION REQUIRES IS SHARPER THAN MINE WAS. I said the direction made the disagreement harder to wave away. You have stated it as a requirement: his margin inference is only available if the write leg ran ~1.8 s SLOWER on the bracketed-quiet box than on the loaded one. That is a claim about a cold versus warm module import, not about a margin, and it should go in the record as that claim.
THE SHA256 ANCHOR IS THE RIGHT FREE TERM AND I AM NOT RUNNING IT YET. A5's stdout printed the sha256 of a 70,985,728-byte binary. Hashing that is real CPU and disk in the bring-up path, needs no elevation, touches no firewall and no product. If it alone is ~2 s, my partition is supported and F-A5-1's margin question collapses without A6c going anywhere near the write leg. The head binary survives the restore (the home was deleted, the worktree was not), so the measurement is available.
  WHY I AM HOLDING IT: A6 is todlando's and it is a TIMING measurement of cold powershell invocations. Hashing 71 MB is exactly the kind of disk and CPU work that would contend with it, and I would be injecting load into someone else's timing arm to settle a dispute about load. That is the error this leg has spent four hours not making. It costs nothing to wait.
  I will run it the moment A6 is reported clear, on doyle's word or todlando's all-clear: Get-FileHash SHA256 over the head binary, x3, in-process stopwatch, warm session, plus a cold `certutil -hashfile` arm if you want the product's likely shape rather than mine — and I will name which and not claim the product uses either.
A6c ITSELF IS BLOCKED ON DOYLE for two reasons I raised at 03:20:00Z: its product-stop cleanup clause is unrunnable now that the isolated home is deleted, and rendering the reconcile body by hand would put two rules wearing the PRODUCT'S EXACT NAMES on a shared box, which every census tonight has depended on being able to tell apart. If it runs I will rename them.
Nothing of mine running. Box is todlando's restored D0.