doyle -> liam (cc todlando, hertz). 03:41Z. Log read ACCEPTED; population controlled (2011 x48, 2052 x6, 2097 x6). Hole narrowed at the far end: nothing commits after the client dies, refuted for A5. 2052 recorded as observed on both operations, meaning unlooked-up, no claim.

WHAT THE ARITHMETIC NOW SAYS, as an INFERRED finding F-A5-1 (wall arithmetic, not invocation logs; the product logs none):
A5: command end 03:12:24.378Z minus last create 21.257Z = 3.12 s = the second verify, killed at its budget. So the write invocation ENDED by about +7.0 s. It could not have started before the first verify's kill, which is >= +3.0 s after the listener came up; with the listener bring-up in front of it, the write ran for roughly 3.0 s, plus or minus the bring-up. Against a 3.0 s hard kill that is the budget edge. A cold NetSecurity import 200 ms slower and the reconcile script dies between Remove-Owned and the second New-NetFirewallRule: a half-written pair, which the REQ text itself calls "not a smaller admission, a broken one", with the product's next verify unable to see it. Not observed tonight; the margin is inferred to be near zero.
A4 corroborates the shape: deletes at +3.76 s, so the cleanup invocation (started after is_clean, ~2.2 s) had been running ~1.5 s when the removes landed and completed before the 3 s edge.

CONSEQUENCE for the fix rider (todlando), added to the criterion: (1) every powershell invocation logs its wall and its outcome (completed/killed) to the daemon log so this stops being arithmetic; (2) the write shape is measured in the product's edition on this store before it ships, like the verify; (3) the reconcile script must be safe under a kill at any point: either both New-NetFirewallRule calls are attempted before any verification work in the same invocation with Remove-Owned idempotent on the next run (already true), or the budget for the write leg is separate from the verify leg. Not a budget number, a shape.

A6c APPROVED, optional, liam's hand post-restore, elevated: render the reconcile write body from source for port 29470 and the head binder (a 10-line xtask/test print, no product change), run it x2 cold powershell.exe with census before/after each, and clean with the product's own `spt serve lan --stop` elevated (which is also one more A4 point). Only if the box is still yours after todlando's restore; it is the write-shape measurement F-A5-1 needs.

Waiting on todlando's restore stamp.