docs(changeset): @objectstack/runtime carries NO boot-refusal row for the STACK_* codes now — correct the closing sentence of the stack-refusal-envelopes changeset

[claude]

Fixes #16857

.changeset/stack-refusal-envelopes.md (#16342's changeset, still unconsumed) closed by claiming that @objectstack/runtime "carries one classification row per new code in the dispatcher error-code vocabulary". That was true when it was written and is false against main today — the runtime vocabulary carries zero such rows. This corrects that one sentence, before a release compiles it into a published packages/*/CHANGELOG.md that this repo does not amend per-PR.

The reading

Every number below is a grep over packages/runtime/src/dispatcher-error-vocabulary.ts at blob ff83d6506d53cee86c87f1d124e56e1978e65dc5, which git rev-parse shows is byte-identical at this branch's base 12babac13 and at origin/main c930f8597. A reading is a count plus the tree it was taken against, so both are named. Exit codes were captured before any pipe.

reading	pattern	exit	count
subject — the false claim	code: 'STACK_	1	0
control — line 20's claim, still true	verdict: 'boot-refusal'	0	14
control — population sanity	code: '	0	84
positive control for the zero	STACK_	0	1 — the removal comment at line 302

The positive control is the same tool over the same file, so the zero is a reading and not a broken grep.

Where the rows went, measured rather than recalled: abae16adf deleted exactly seven of them — grep -c "^- code: 'STACK_" over its patch for that file returns 7 — in the same landing that registered those seven codes in ERROR_CODE_LEDGER, where they now sit at packages/spec/src/api/error-code-ledger.zod.ts lines 1045-1051 under the @objectstack/spec key. The gate's own header states the consequence: a registered code drops out of its scan, and a stamp site under packages/spec/src/ may be classified foreign-vocabulary or runtime-pinned "and NOTHING ELSE".

Two things in the same file that did NOT change — proven by hash, not by inspection

invariant	sha1 before	sha1 after	cmp
line 20 — "...one-row-per-refusal granularity (14 rows)..."	71192bf158de167ac2774c576a462350600d7fa8	71192bf158de167ac2774c576a462350600d7fa8	exit 0
front matter, lines 1-4 — including "@objectstack/runtime": minor	c3e97c77ad1e4948e11906c233e2d326ffcf4003	c3e97c77ad1e4948e11906c233e2d326ffcf4003	exit 0

Line 20 is the neighbouring sentence about the same fact family, and it is true — 14 is exactly what control 2 returns. Two sentences describe that family; only one went false.

The front-matter key stays on the seat's ruling in issue comment 5585216825: eight other unconsumed changesets already declare @objectstack/runtime: minor, so the released version is inert either way, but the key also decides whether this paragraph compiles into packages/runtime/CHANGELOG.md, which ships in the tarball via files[] — and that history belongs where a runtime consumer greps for it.

Anchor counts across the edit, each with its exit code:

anchor	before	after
carries one classification row per new code (the old sentence)	1 (exit 0)	0 (exit 1)
carried one classification row per new code (the new one)	0 (exit 1)	1 (exit 0)
one-row-per-refusal granularity (14 rows)	1 (exit 0)	1 (exit 0)
"@objectstack/runtime": minor	1 (exit 0)	1 (exit 0)

git diff --stat: 1 file changed, 1 insertion(+), 1 deletion(-).

Register

This matches PR #16856, which corrected line 24 of the same file in the same landing shift: the true half of the sentence is kept, the false half is replaced by what actually happened, and the card that moved the fact is named (#16449, PR #16652).

Gates

node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --ran reconciles the run against this tree's own derivation:

✓ dispatch-gates --ran: 17 derived famil(ies) accounted for — 17 run, 0 NOT-MEASURED.

All 17 exit 0, at 1be4e378c. Three of them (check-closing-keyword-parity twice, check-comment-mask-corpus) first returned exit 3 — PREREQUISITE NOT MET, which is not a red gate; they were re-run green after pnpm install. check-changeset-fixed.mjs is included deliberately: the derivation flags its roster as living under .changeset, so its silence would not have been evidence in either direction — it was run, and it passes.

Not owed here, and named rather than skipped silently: no package source is touched, so there is no dependency-closure build and no package test/typecheck to run. Repo-wide pnpm lint is CI's run, not this PR's.

skip-changeset

Measured, not assumed: of the 82 workspace package.json files, 70 declare a files[] whitelist (matching check:published-files's own count of 70 publishable packages), and 0 of their files[] entries mention .changeset — positive control on the same scan: 70 entries mention dist. .changeset/** is repo-root release-notes input and ships in no tarball, so nothing published moves.

Landing

⛔ Draft on purpose. Not flipped ready, not enqueued, no auto-merge armed — the dispatching seat does that.

---

Generated by Claude Code

[zhuangjianguo]

ACCEPT — seat verification at 2026-09-08T12:53Z, head 1be4e378c. Flipping ready; enqueue held until every check name closes.

domain:spec execution seat, session session_016N6xmWt5hYm94ffVEwGH8x.

Verified against the tree, ⛔ not accepted on the report

Each of these was re-derived by this seat from the PR head blob (git show refs/pull/16874/head:.changeset/stack-refusal-envelopes.md), not from the dev's tables:

invariant	seat's reading
the false sentence is gone	carries one classification row per new code → 0 (exit 1)
line 20 is untouched	one-row-per-refusal granularity (14 rows) → 1 (exit 0), and the sentence reads intact
the front matter is untouched	"@objectstack/runtime": minor → 1 (exit 0); lines 1-4 read back verbatim
the hunk cannot have touched either	the diff's single hunk is @@ -25,4 +25,4 @@ ⇒ line 20 and lines 1-4 are structurally outside it, which is stronger than an assertion that they were left alone

Tier fuse (PM): the dev's transcript carries 107 harness-stamped "model" values, all claude-opus-5, zero of any other value — matching the tier this dispatch passed explicitly. ⛔ No self-description was accepted.

⭐ The correction is the right shape, and that is the substance of this card. It does not pretend the old sentence was always false — it says the row existed "when this changeset was written" and carries none now, names the landing that changed it (#16449 / PR #16652), and says why the seven cannot come back (a registered code drops out of that gate's scan, and the gate now refuses any verdict but foreign-vocabulary / runtime-pinned for a stamp site under packages/spec/src/**). That matches the register PR #16856 set on line 24 of the same file two hours earlier: keep the true half, replace the false half with what actually happened, name the card that moved the fact.

On the two things the dev was told not to decide

• The front-matter minor key stayed, per this seat's ruling (5585216825). The dev executed it as an instruction and did not re-open it — correct.
• The halt condition from the previous run is discharged: all three corrected controls re-derived at origin/main c930f8597 and all three agree (code: 'STACK_ → 0 · verdict: 'boot-refusal' → 14 · code: ' → 84). ⭐ The 84 is the value this seat got wrong and the previous dev caught; it now reproduces from a third independent run.

Gate posture

Clause-②: no. .changeset/** only ⇒ the enqueue gate's PATH limb is not hit and no at-tier contract review is owed — the same shape as #16671 / PR #16856, which landed at opus this shift on the same basis. skip-changeset is measured rather than assumed: of 82 workspace package.json files, 70 declare a files[] whitelist and 0 of their entries mention .changeset (positive control on the same scan: 70 mention dist).

⛔ Why this is not enqueued yet

28 checks, 4 still running (Type Check · debt ledger / · workspace / · consumer gates, Lint & Repo Gates), 0 non-green. ⛔ "No red yet" is not an enqueue reading — the criterion is every name completed with success/skipped, aggregated by name. Flipping ready now (which re-runs the pull_request gate family), and enqueueing when the full set closes.

⚠️ The deadline on this card is the release, not the merge. The target file is still an unconsumed changeset today; once a release compiles it, this correction becomes a published-CHANGELOG edit needing a dedicated docs-only PR — which is exactly what already happened to #15058 and #15026. #16850 is the mechanism card for the fact that nothing watches that deadline.

---

Generated by Claude Code