Does Refs #N + post-merge go entirely, or only for the operator-owned remainder? — the scope call split out of #488

[dan-claude-bot]

Split out of discussion #488 at @danmt's direction (2026-08-28T17:55:27Z — "those two probably should have their own discussions so we can close this thread"). This is the second of the two residuals #488 carried, and it is the one whose cost has moved since it was first stated — see §3.

1. The question

Ruling D on #488 was delivered with its rationale attached (2026-08-28T12:37:00Z):

The Ref thing and keeping issues open post-merge with huge walls of text is just retarded.

Triage read that as the reason for D and scoped the build accordingly: #525 abolishes Refs #N + post-merge for the operator-owned remainder — those criteria relocate into their own issue before the merge, and the PR closes its issue normally. It leaves the mechanism standing for a criterion that is post-merge but builder-owned: a tag that must publish before it can be observed, a consumer's docs-sync --check that only reds after the pin bump, a drill record that needs the cut it records.

The narrow reading is what shipped. The question is whether the wider one was meant: does Refs #N + post-merge go entirely, for every criterion, leaving no route by which an issue stays open after its PR merges?

2. Why it was not just assumed

Because the two readings produce different repositories, and only one of them is recoverable cheaply.

• Under the narrow reading (shipped), post-merge keeps one job: a criterion whose evidence does not exist until after the merge, produced by a builder-reachable surface. The queue stays small and the "wall of text" @danmt objected to is gone, because the walls were the operator-owned remainders.
• Under the wider reading, every such criterion has to go somewhere else. There are only three places it can go: into a successor issue minted at merge time (which is D applied to builder-owned criteria — more issues, no walls), into the reviewer's judgement at the merge door (which is the merge gate D removed), or into nothing at all, which means the repo stops asserting the things post-merge exists to assert.

That is a materially bigger change than D, it touches TRIAGE.md, BUILDER.md, REVIEWER.md and LABELS.md, and it would obsolete a queue state the sweep implements, nudges and reconciles. It is not a re-word of #525; it is a different issue, and triage mints it on @danmt's word.

3. The cost, re-measured — and the earlier figure was wrong in a way worth naming

On #488 at 15:39Z triage wrote that "two issues on this board were live under that mechanism when I wrote it; both have since closed, so the cost of the wider reading is lower today".

That was measured on this board alone, and this board is not where the mechanism lives. Measured across the nine governed boards on 2026-08-28T19:0xZ:

board	open post-merge issues
heavy-duty/crew	7
heavy-duty/incubator	5
heavy-duty/box	4
heavy-duty/lafamilia-site	2
heavy-duty/ceremony	0
total	18

So the wider reading retires a queue state that 18 open issues are sitting in right now, each of which would need a disposition — not the near-zero cost the earlier sentence implied. The narrow reading touches none of them.

4. What is being asked

One word decides it. "Narrow" (or silence) confirms what shipped and this thread closes with nothing minted. "Wider" mints a fresh issue that retires Refs #N + post-merge for builder-owned criteria too, states where those criteria go instead, and disposes of the 18 live instances — with the disposition being the expensive part, not the doctrine edit.

Nothing is blocked either way. #525 is claimed and in flight; its tenth criterion holds the build to leaving the builder-owned mechanism untouched, so it ships unchanged under both answers and a "wider" ruling mints successor work rather than invalidating it. Silence here has no timer and triage promises no pick: this changes published governance for every governed repository, which is the operator's call by construction.

[danmt]

My idea was to go with the wider reading. post-merge makes the board keep issues open pointing to merged PRs and its a bit confusing. I'd prefer that it would be divided into a separate issue. I'd go further and say that it doesn't need to be at merge time but during mint, if we need to do something for either a builder or an operator after an issue is closed it should be its own issue. I think it makes it more deterministic in exchange of more issues. @dan-claude-bot

[dan-claude-bot]

Ruled: wider, and further than the question asked. Minted as #538 with two children; this thread closes.

@danmt Taken as written, both halves.

The half #536 asked about. "Wider" — Refs #N + post-merge goes for every acceptance criterion, builder-owned and operator-owned alike, and no issue stays open after its own PR merges.

The half you added, which is the bigger one. "It doesn't need to be at merge time but during mint." #536 framed the wider reading as a merge-door change and priced it that way; you moved it to the mint. That is a stronger rule than #525's route D rather than a wider one — under it the mint-time split is the route and the merge-door relocation is the repair, which is what #525's own D2 already calls it. Nothing in #525 is reopened or invalidated by that; it is generalized.

What is minted

#	what it does	queue at mint
#538	the epic — thirteen decisions, the census, the scope boundaries	epic
#539	the doctrine: TRIAGE.md, BUILDER.md, REVIEWER.md — split at the mint, Closes #N always	blocked behind #537
#540	the queue state: LABELS.md, RELEASES.md, both reconcilers, the taxonomy row	blocked behind #539 and a census

Two children rather than one or four, because the state between them is consistent: #539 stops new post-merge criteria being minted while the sweep goes on deriving the state for the ones that already exist, and #540 removes the state once the last one closes. Splitting further is not legal — test/labels-reconcile.test.sh reads LABELS.md's own table as the machine's expected values, so the taxonomy row and its documentation move in one commit or the suite reds between them.

One thing you may want to overturn, because it is my call and not yours

The 18 live instances drain rather than convert. Re-measured across all nine governed boards just now: crew 7, incubator 5, box 4, lafamilia-site 2, and zero on ceremony, la-familia-incubator, landing-site, rig and rig-templates. Each was minted under the contract you are repealing and each is honest under it, so triage ticks and closes them as their evidence arrives and splits only whichever are still open when #540 is otherwise ready. That count reaching 0 is #540's gate.

The alternative is 18 splits now, across four boards, on issues whose contracts are already correct. I took the cheaper reading; say the word and they convert instead, and only #540's gate changes.

One thing that is worth knowing before the build starts

The rule you ruled is already written down, in a narrower form. RELEASES.md L204–L208 says triage "splits the remainder: mint a fresh issue carrying the outstanding criteria verbatim, naming its owner and its wake condition and citing the original, then close the original on what it delivered." Today it fires only inside a release window and only when an open declaration names the issue (#329). Your ruling deletes both conditions and moves it to the mint. So this build generalizes a rule the repo already ships and tests, which is why the doctrine child is prose rather than machinery.

Scope boundaries the build is held to, so no PR widens the ruling

• Refs #N survives in exactly one place: BUILDER.md's round cap, where the predecessor PR's Closes #N → Refs #N edit moves the close to the successor. That is a PR-chain mechanism and has nothing to do with acceptance criteria. A build that greps the two words and sweeps them takes it; TRIAGE.md, BUILDER.md and REVIEWER.md — after-close work is split at the mint, and every PR carries Closes #N #539's ninth criterion rejects that build.
• FLEET.md's two uses of the words, and labels-reconcile.sh's POST_MERGE_WORKFLOW dispatch, are temporal — "a PR already merged is a post-merge wait" — and not the queue state. Both children's criteria reject a build that sweeps them.
• operator's quiet nudge is not post-merge's. LABELS.md + TRIAGE.md + BUILDER.md + actions/labels-reconcile — the issue flow can say what state work is in and not whose it is, so an issue only the operator can do reads ready and lies to every builder #491 gave it the same clock and the sweep implements it as "post-merge's quiet twin". The twin stays, and LABELS.md, RELEASES.md and both reconcilers — the post-merge queue state is retired once its last instance closes #540 owes a mutation proving it still fires.
• The label object survives on your boards. Removing the row from the taxonomy stops it being created; it does not delete it from the nine repositories. LABELS.md, RELEASES.md and both reconcilers — the post-merge queue state is retired once its last instance closes #540's changelog fragment says so, because a consumer will otherwise expect it to vanish.

Nothing is blocked on you. #538 is blocked behind #537 under the ordinary collision rule — the whole doctrine set is carried by the open chain #525 → #526 → #532 → #534 → #537 — so it queues rather than racing. It is added to #528's gate under the standing reading of "clear the board before releasing this", and that is disclosed there rather than here.

Closing this thread; your comment above is the answer of record.

[dan-claude-bot]

Correction — the governed roster is eleven boards, not nine, and my table above was two short

@danmt A measurement error of mine, found while writing #541's criteria and repaired everywhere it landed.

I identified the governed boards as those carrying a post-merge label. That test reads out any board whose ceremony pin predates that row — which is exactly the two furthest behind:

board	labels.yml pin	operator
heavy-duty/martin-reyes-barbershop	0.3.0	❌
heavy-duty/cast	0.1.0	❌

Both call heavy-duty/ceremony/.github/workflows/labels.yml, so both are governed. The right test is that uses: line, plus ceremony itself, which runs its own action — eleven.

What changes: operator is present on 3 of 11, not 3 of 9, and eight boards are owed the row rather than six. #541's table, its task rows, two more acceptance criteria and its close census are amended, and its R2a now records the flawed test so the next census is not built the same way. The two additions are also the two largest pin bumps on the list, which is worth knowing before anyone prices them.

What does not change: the post-merge count of 18 — neither board has that row either, so neither can carry an issue wearing it — nor the 8 open ready issues, nor anything in the argument above about crew's filter.

The same correction is applied to #538, #539, #540 and #525, which all carried the phrase.

[andriujoseba]

Follow-up from PR #569 review: test/attention.test.sh:26-27 still passes the retired post-merge-assigned marker as its suppression argument. #567 deliberately forbids that path, so the builder is not changing it in the retirement union. Recording it here at the intent door so a later mint can re-point the stale test vocabulary without widening #567.

[dan-claude-bot]

Discharged — minted as #586. Late by a day, and the delay is worth naming: this thread closed 2026-08-28T19:56:16Z and your record arrived 2026-08-30T19:16:51Z, so it sat outside every open-discussion sweep. It was still the thing that worked. #567's close-out asserted the mint had been made and it had not; the pass that found the gap found it here, at the door you put it in.

Two notes on how #586 rules it, both of which change your reading slightly and neither of which is a correction to your record:

• The case is retargeted, not deleted. LABELS.md, TRIAGE.md, BUILDER.md, RELEASES.md and both reconcilers — the four 0.7.8 gate contracts land in one pass #567's close-out called it a duplicate wearing a retired state's name, which is true of the name and false of the effect. Mutate lib/attention.sh:29 from [ -n "$2" ] to a literal comparison against claimed-unassigned and the :24 case stays green — :26 is the only case in the suite that reds. It is the sole pin on presence-not-identity, so test/attention.test.sh — the second suppression case names a retired state, so the one property it alone pins reads as a duplicate #586 requires that mutation run and recorded, and forbids the deletion.
• Your scope call stands as written. test/attention.test.sh was outside LABELS.md, TRIAGE.md, BUILDER.md, RELEASES.md and both reconcilers — the four 0.7.8 gate contracts land in one pass #567's authorized union, the reviewer did not request it, and holding the line there was correct. Widening the union would have been the worse trade even knowing the mint would be late.

Nothing is owed on this thread and it stays closed; #586 carries it from here.

[dan-claude-bot]

✅ Answered — late, and in place. Your raise is discharged: the retired marker is gone from main, and the thread stays closed

@andriujoseba Your 2026-08-30T19:16:51Z record at the intent door landed on a thread I had closed 44 hours earlier, and it never got a reply here — two days late, and that is my miss, not yours. The work you asked for happened. Nobody told you in the venue you wrote in, which is the failure mode where a raise is answered by a build and stays unanswered in writing.

Converging in place rather than reopening: the substance shipped, so a reopen would signal unconverged work that does not exist.

---

1. What you asked for, and what landed

You wrote: "test/attention.test.sh:26-27 still passes the retired post-merge-assigned marker as its suppression argument … Recording it here at the intent door so a later mint can re-point the stale test vocabulary without widening #567."

step	when
your record, at the intent door	2026-08-30T19:16:51Z
#586 minted — bug, scope:labels	2026-08-31T21:03:53Z
PR #587 merged at head b5e8dac	2026-09-01T07:54:54Z
#586 closed	2026-09-01T07:54:56Z

Measured at origin/main = 5c47768, not read off the merge:

• grep -rn 'post-merge-assigned' . over the whole tree returns nothing — rc=1. It survives in no test, no library, no action.
• The line you flagged now reads attention_comment_decision MALFORMED_UNASSIGNED any-other-precedence at test/attention.test.sh:35.
• Every remaining post-merge string in the tree is one of the classes Retire post-merge — after-close work is decided at the mint and becomes its own issue #538's close-out ring-fenced as not the queue state: the POST_MERGE_WORKFLOW dispatch and its #461 D2/D3 wake comment in labels-reconcile.sh, FLEET.md's temporal "a PR already merged is a post-merge wait", CHANGELOG.md history, the drills/ records, and test/labels-reconcile.test.sh:345 — which is the assertion that the core row is retired. So the vocabulary you were pointing at is closed out, not merely relocated.

---

2. The part of your framing that turned out to be load-bearing

You wrote re-point, not delete. That was the right word, and the build had to discover why the hard way.

The case reads like a duplicate of its neighbour, and deleting it was the obvious repair. It is not a duplicate: attention_comment_decision tests its second argument for presence — lib/attention.sh uses [ -n "$2" ], not a comparison — so this is the only case in the suite pinning that the verdict does not key on which precedence it was handed. Under a literal-comparison mutation of that line it is the sole red; its neighbour stays green. Deleting it would have silently dropped the one property it alone holds.

So it was retargeted to a synthetic literal and given a comment saying why, rather than removed. Your sentence is the one that got that right.

---

3. Attribution, stated honestly rather than generously

I am not going to claim your comment triggered the mint, because the record does not support it. #586 was minted on a hygiene pass that re-derived the residue from the tree, and its body traces the source to #567's close-out. Your record is cited in that close-out's in-place correction — where the same pass found that its own "minted separately" clause was false when written and struck it — as the guard that was standing while the routing did not happen.

What is true is narrower and still worth saying: you filed it at the right door, in the right words, against a disclosure that had already quietly failed, and the only thing that went wrong afterwards is that nobody came back here to tell you.

---

4. One live follow-on, so you can see where the file is now

#594 — documentation, scope:labels, claimed by @cndgrr since 20:58:01Z, in flight at PR #595. This is not your raise and does not reopen it. It is the next layer of the same eight-line comment: that comment locates its predicate by line number (lib/attention.sh:29), the only such pointer in the tree, and nothing reds when the line moves. It came from a declined non-blocking nit on #586's own panel.

Mentioning it only so the file's state is legible from here. Nothing on it is owed to you or waits on you.

---

Nothing is asked of you and nothing is owed back on this thread. It stays closed and stays answered — @danmt's 19:14:26Z comment remains the answer of record for the question #536 asked, and your record is now discharged rather than pending. If you find another instance of retired vocabulary outliving its state, the intent door is exactly where it should go again.

Converged in place by heavy-duty/ceremony triage, 2026-09-01T21:1xZ. No label moves and no attention is set: this comment delivers no assignee's next move, #594's claim is proceeding normally, and a discussion carries no board state.