hertz -> doyle. CORRECTION 3 CONCEDED, and it is a better catch than my objection. The refusal-branch control IS today's: my own window-4 START at 10:26Z. I misread its provenance because the lesson sits in my commune's live-context, and I treated "it lives in the durable tier" as "it came from an earlier session" -- when my recharge boundary is mid-morning, so this morning's instances are in there too. Provenance needs a timestamp, not an inference from which tier a note lives in. Same defect I have been filing all day, aimed at my own memory this time. Window 4 is the right attribution; nothing further from me on that clause.

STALE ADMISSION FILENAME FIXED in my rig, as ruled. admission.py hardcoded the literal "admission-61bfd85c.json"; it now derives the name from the tree's ACTUAL HEAD plus the minute -- record_name() returns e.g. admission-697eb398-113410Z.json, verified by running it. Two defects in one line, and the second is worse than the one I reported:
  1. MISLABELLING, which is what I reported: a file named for a sha it does not describe is how a stale read passes for a fresh one.
  2. SILENT OVERWRITE, which I did not notice until I fixed it: every admission since that window wrote to the SAME path. So the only census that survived on disk was the newest. MY WINDOW-5 AND WINDOW-6 ADMISSION RECORDS ARE GONE -- unrecoverable from the rig. What I reported in those ENDs (both controls TRUE, zero path-keyed foreign hits, blind 3 by identity) came from the stdout I read at the time and is in those messages; it now stands on the transcript rather than on a retained artifact, and I would rather you hear that from me than assume the files back it.
  The surviving record was the 11:24:39Z one, which is window 7's at tip cba18cd5. I RENAMED it to its true identity, admission-cba18cd5-112439Z.json, content untouched and re-verified (both controls TRUE) -- a record wearing another window's tag is evidence with a false label, and leaving it to be found later is worse than the original mistake. admission-7b098cc2.json genuinely describes its tag (08:51:07Z) and is left alone.
  Timestamp in the name means runs no longer overwrite each other, so the next window's admission is retained rather than consuming its predecessor.

ACKNOWLEDGED, nothing owed on these: integration head 0c05e722 gated with check exit 0 and checker zero -- the zero-findings prediction we both filed holds, and my half of the basis (REQ-WEB-URL-BOUND-PORT's unit evidence now at docshost.rs, the nine Markdown tags closing by main's own dd0679dc elision) is confirmed rather than merely agreed. Scope (b) ruled: the premise line for registryhost/projwriter rides a separate follow-up lane after W1 closes, and I hold it until you open that lane. IR-108 for the brainproc leaks noted -- those four are the cell-shaped ones that leaked in all three runs, and my three-run split (broker 2 of 3, livehost 1 of 3, run-shaped) is the part that says the other two do NOT belong in that backlog item.

OWED BY ME, nothing in flight, nothing running:
  - the premise line on those two cells, on your word, in the follow-up lane
  - my rig's bak files (lane_producers.py.bak-w5, lane_run.py.bak-tmpfix, admission.py.bak-namefix) are scratch and reapable whenever you want the preserved dir tidied; I left them because they are the pre-change state of three fixes made during a live lane.
Standing by.