blocked is invisible to decide_state — a frozen PR reaches state:needs-human and tells the maintainer to merge it

[dan-claude-bot]

state:needs-human means exactly one thing — LABELS.md says it "requires zero blockers", and merge-next is cleared the moment a PR is "no longer the thing a human should merge next". Right now, during the freeze, it is lying on two PRs at once.

What the board is currently saying

Both of these carry blocked, applied by triage at 08:19 under the ceremony#111 directive, and both carry state:needs-human with @danmt as the requested reviewer:

• rig#126 — reconciler set state:needs-human at 08:50:49, 31 minutes after triage set blocked.
• rig#128 — reviewer hand-set it at 08:50:33; the reconciler validated the write and left it standing.

In both cases the review round genuinely finished with a full panel approving the current head. The state label is a correct statement about the round and a false statement about the merge, which is the one thing the doctrine promises it is about.

Why

decide_state() is the only place the two axes join, and it demotes to state:addressing when blockers() returns anything (L288–L292). But blockers() only ever emits blocker:conflict, blocker:ci-red and blocker:unrequested (L252–L274) — all three derived from branch facts. The cross-cutting blocked label is never consulted on this axis, so it cannot demote anything.

This is not a read problem. The same script already reads the label: has_label blocked gates the staleness skip at L499, and the merge-next clause deliberately declines to set blocked because it is intent. It is legible and correctly hand-owned — it is simply not consulted on the one axis where it decides whether a maintainer is being invited to merge.

Why this is worth a decision now rather than filed as debt

Before the freeze, blocked on a PR was rare enough that the gap never fired. Under the freeze it is the normal state of every open PR in five repos, and every one of them whose round completes will land on state:needs-human — the two above did so within seventeen minutes of each other, without either being touched by a builder. The failure mode is the expensive one: a maintainer merging a PR the directive parked, on a board that told them to.

The decision

The immediate correctness fix is small; what it should say is not obvious, and that is the part I do not own.

1. blockers() emits blocked when the label is present. One line, reuses the existing demotion. Against: blockers() documents itself as branch facts computed from GitHub, and its return value is also the set written back as blocker:* labels — a non-blocker:* member would need special-casing at the write, which is exactly the kind of precedence tangle the two-axis split was minted to remove.

2. decide_state() consults has_label blocked directly, beside its existing blockers() check. Keeps blockers() purely branch-derived; one extra condition on the only function that already joins the axes. This is the cleaner mechanism — but its demotion target is state:addressing, which means the builder owes work, and under a directive hold the builder owes nothing. ceremony#113 is currently writing the doctrine that says exactly that. So this fix stops the lie about merging by telling a smaller one about whose ball it is.

3. Leave the machine; change the doctrine — redefine state:needs-human as "the panel is done" and treat merge-gating as a separate concern a human reads off blocked. Against: it retracts LABELS.md's own one-thing sentence, and needs-human going stale is the precise bug that split the axes in the first place.

4. A fifth state. None of the four says reviewed, correct, and waiting on something that is nobody-here's ball — which is what a frozen PR with a finished round actually is. The freeze made that state common; before it, it barely existed.

My recommendation: 2 now, 4 as the question worth answering properly. The needs-human promise is the one that must not lie, and imprecision about whose ball it is costs less than a wrong merge — but if the answer is 4, doing 2 first means writing the demotion twice.

Decider: @danmt. The freeze is yours and so is the taxonomy; I have not touched the state labels on either PR (states are machine-owned, and hand-stripping one fires unlabeled → the reconciler recomputes it back within seconds). Both PRs now carry a triage comment saying blocked outranks the state label and neither is to be merged, so the hazard is papered over on those two — but it is papered over in prose, on a board whose whole value is that its labels are true without reading the comments.

[dan-claude-bot]

This thread was opened without the block it needs, and that is my miss. The
body above lists four options and names no default — BUILDER.md's ruling
ask
says more than three means the question is not ready, and without a Default:
there is no ladder and this sits forever. Both fixed here. The analysis above is
unchanged; nothing in it is retracted.

Options 1 and 2 in the body are two mechanisms for one decision — demote a
frozen PR to an existing state — so they collapse into A below. Option 4 is B,
option 3 is C.

🧭 needs-ruling — what does the board say about a PR whose round is finished but which `blocked` holds?
Options:  A — decide_state() consults `blocked` and demotes to the existing state:addressing   B — mint a fifth state meaning "reviewed, correct, waiting on nobody here", and demote to that   C — leave the machine; redefine state:needs-human in doctrine as "the panel is done" and let a human read merge-gating off `blocked`
Recommend: A, because state:needs-human's one promise — a human could merge this now — is the one that must not lie, and being imprecise about whose ball it is costs less than a wrong merge. B is the better end state and stays worth asking separately; taking it now means writing the demotion twice, which is cheap.
Blocked:  nothing. No issue is minted for this and no builder is waiting on it — deliberately, because minting one today would hand a builder the choice. What continues broken meanwhile: every frozen PR whose round completes lands on state:needs-human and invites a merge the directive parked. rig#128 is sitting there now, papered over in prose. rig#126 carried both labels and was merged seven minutes after the reconciler wrote the green one; whether the label caused that or not, it is the sequence the two-axis split exists to prevent.
Default:  A at 2026-07-25T09:00Z if no ruling. Affirmatively reversible inside the PR: one condition in decide_state(), one sentence in LABELS.md, no new label, no bootstrap row, no consumer action beyond the pin bump they are already taking for 0.2.0.

Three things the block compresses

Why A's demotion target is imperfect and I recommend it anyway.
state:addressing means the builder owes work, and under a directed hold the
builder owes nothing — #113
is writing that exact doctrine right now. So A stops the lie about merging by
telling a smaller one about whose ball it is. That is the whole trade, stated
plainly: a wrong merge is expensive and irreversible; a misread of whose turn it
is costs a comment.

Why the mechanism inside A does not need your ruling. Body option 1
(blockers() emits blocked) and option 2 (decide_state() reads the label
directly) reach the same board state. 2 is the better build — blockers()'
return value is also the set written back as blocker:* labels, so a
non-blocker:* member needs special-casing at the write. That is an
implementation call the issue will carry; it is not yours.

Lifting the freeze does not retract this. If the ceremony#111 hold
lifts
at 12:00Z, blocked on a PR goes back to being rare and this reverts from a live
hazard to latent debt. The bug is identical either way; only its urgency moves.
That is why the default is 24h rather than same-day.

No needs-ruling label is set anywhere for this, deliberately: it gates
nothing on any board, and on the two PRs that demonstrate it — both in rig — the
label does not exist at all. That gap is
#102, with the
machinery half minted at #105.

If the default fires I mint A, cite this thread for the reasoning, and stay
accountable for the pick. It still lands as a PR you merge, and you can overturn
it there.

[dan-claude-bot]

Premise update, not a new outcome and not a change to the ladder.

The ceremony#111 freeze is lifted as of 12:00Z — my pick, recorded
there,
after the deadline I set at 09:06Z passed without an answer. The escalation
above anticipated exactly this: "If the ceremony#111 hold lifts at 12:00Z,
blocked on a PR goes back to being rare and this reverts from a live hazard to
latent debt. The bug is identical either way; only its urgency moves."

So: the bug is unchanged, and the urgency has dropped. blocked on a PR is
rare again — there is no open PR in ceremony, rig or incubator carrying it right
now — so no board is currently inviting a merge the directive parked. What
remains is the latent defect: decide_state() still never consults blocked,
and the next PR that carries one with a finished round will still land on
state:needs-human and lie about it.

Nothing else moves. The options, the recommendation (A), and the default —
A at 2026-07-25T09:00Z if no ruling — all stand exactly as posted; the
default was already set at 24h rather than same-day for this reason. @danmt.

[dan-claude-bot]

Decision — the default fired. 2026-07-25T09:00Z passed with no ruling; per the ladder's past-24h rung (LABELS.md, #50 D13–D14) triage picks and stays accountable:

Option A — decide_state() consults blocked and demotes a finished round to the existing state:addressing, as an exclusion exactly parallel to needs-ruling, never as a blockers() emission. One condition, one LABELS.md sentence, no new label. B — the fifth state — stays worth asking separately and is deliberately not part of this pick.

One ladder note against myself: the 12:01Z comment above was a premise update at 2.5 hours, not the 12h re-read, and the 21:27Z rung got no comment of its own. The re-read ran now, before firing: at main (24255a9) the defect is unchanged — decide_state() still joins the axes only through blockers(), and the needs-ruling exclusion at L389 is the exact shape A mirrors — no ruling landed anywhere on the thread, and the 12:01Z premise still holds (no open PR in ceremony, rig or incubator carries blocked this morning). The default is not stale; it fires.

Recorded as #180 — minted ready, carrying the full spec, the test constraints, and this thread as its reasoning.

@danmt — you overturn this at the merge gate, not before it: the change lands as an ordinary PR you can decline, and if the answer was C — leave the machine — closing #180 with that one word undoes everything. Resolving this thread; the decision record is this comment plus the ladder above it.

[dan-claude-bot]

Correction to one sentence in the decision above, from the read I owed before writing it. "No open PR in ceremony, rig or incubator carries blocked this morning" is false: rig#145, the 0.3.1 release candidate, carries blocked — held still for the drill record per rig#132 — alongside state:building and blocker:ci-red. The sentence was true when first written at 12:01Z yesterday; #145 was opened that evening, and I repeated the claim without a fresh read.

What survives: no board is displaying the lie today — state:needs-human needs a finished round and #145's has not started, so no merge is being invited. What changes: #145 is exactly the shape #180 exists for — a family PR whose round can finish while a directed hold stands — so this raises #180's urgency rather than lowering it. The pick and the decision are unchanged.