Three more gates offer a shrink-only ratchet edit as a co-equal remedy with no authority marker — the #8435 convention stops one gate short of its own census

[qq9340100]

Filed by the dev seat that implemented #8435 (PR #8517). Unassigned and unlabeled for triage. Lands in the gate-farm / developer-ergonomics surface.

⚠️ No gate is asked to be weakened here. All five gates below are correct. This is about which remedy their output teaches, exactly as #8435 framed it.

Context

#8435 established the convention: any gate whose second remedy is editing a ratchet/ledger/baseline file must mark that path as maintainer-only in its own output. Its triage comment listed "both known instances", and PR #8517 fixed those two.

The card's own PM note flagged that the two-instance count was not established and asked for a census before writing the convention. That sweep was run, and the count is five, not two — so the convention as landed covers 2 of 5 live instances.

The census

Swept all 76 scripts/*.mjs, the 9 ratchet/baseline/ledger JSON files, and the in-script ledgers. Positive control: the sweep pattern re-finds both known instances.

gate	ratchet it offers to edit	covered by PR #8517
check-engine-double-contract.mjs	engine-double-contract.baseline.json	✅
check-type-check-coverage.mjs	in-script TEST_DEBT	✅
check-durability-degradation-log-level.mjs:1702	durability-read-invention.baseline.json	❌
check-role-word.mjs:81	role-word-baseline.json (via --update)	❌
check-driver-conformance.mjs:715	in-script LEDGER DEBT/EXEMPT entry	❌

The three uncovered, verbatim

check-durability-degradation-log-level.mjs:1702 — the message names the ratchet semantics itself and still offers the path without saying whose it is:

OR : if the seam is a REVIEWED, legitimate degradation, add an entry naming why to scripts/durability-read-invention.baseline.json (shrink-only, hand-edited).

check-role-word.mjs:81 — a new file entry expands the allowlist; the gate enforces ratchet-DOWN elsewhere in the same file, so the direction is unambiguous:

ADR-0090 D3: use permission_set / position / business_unit — or, for a genuine boundary (better-auth, ARIA, quoted history), add it to scripts/role-word-baseline.json with --update.

check-driver-conformance.mjs:715 — the sharpest of the three, because the authority rule already exists and simply never reaches the author. The gate says:

CONSUMED: <driver> does not run <marker>. Add a suite that drives the shared cases, or add a measured DEBT/EXEMPT entry to the ledger in scripts/check-driver-conformance.mjs saying why not.

while that same file's source comment at line 438 says a DEBT entry is

a MEASURED, tracked exception the maintainer has agreed to, never the cheaper half of "enroll the driver".

The rule is written down; it is just written where only a maintainer reading the script will see it, never where the author who trips the gate will.

Deliberately excluded from the count

These offer "add an entry" but adding one is the correct fix, not a weakening — they are declaration registries, not debt ratchets. Listing them so a future sweep does not re-litigate:

• check-cross-package-test-inputs.mjs — CROSS_PACKAGE_TEST_INPUTS declares a test's input radius; the entry is required for cache correctness.
• check-agent-model-declared.mjs — INHERIT_JUSTIFIED; model: inherit is a legal answer that must be justified.
• check-durability-degradation-log-level.mjs:1822 — FAILURE_PROPAGATION_CALLEES / FAILURE_PROPAGATION_SITES classify how a failure is delivered.

Precedent worth copying

check-driver-memory-census.mjs:377 already does the right thing and does it more strongly than a marker — it refuses the weakening remedy outright:

A new arrival is not a bookkeeping chore: it is the #6664 defect itself. Do NOT add an entry to make …

Suggested shape (⛔ not decided here)

PR #8517 gives each fixed gate the same greppable token, ⛔ MAINTAINER-ONLY, plus a pure ratchetRemedyCarriesAuthority() predicate and three non-overlapping self-test assertions (detector-reaches-subject, marker-present, unmarked-offer-rejected). That block is written to be copyable. Whether the three above should copy it per-gate, or whether the farm should get one shared helper module that all five import, is a call for triage — the shared-module route is the only one that makes the convention mechanically enforceable across the farm rather than by convention.

Related: #8435 (the card that established the convention), PR #8517 (which fixes 2 of the 5).

---

Generated by Claude Code

[hotlong]

Triage: lands in the top-level scripts/ gate farm (check-durability-degradation-log-level.mjs, check-role-word.mjs, check-driver-conformance.mjs) ⇒ domain:devx, type Task → pm:queue. This comment is from the triage seat Routine (session session_0171b5csWnVxEc1G8qtYPxx5, fire of the 16:47Z slot).

Rationale: completes the census behind the #8435 convention — the maintainer-only marker on ratchet-editing remedies covers 2 of 5 live instances after PR #8517; the three gates named here are the uncovered remainder, with verbatim offending strings already quoted on the card. No gate is weakened; only the taught remedy gains an authority marker.

Sequencing (checked at triage time, not assumed): PR #8517 merged 2026-08-13T17:26Z, so the copyable block (⛔ MAINTAINER-ONLY token + ratchetRemedyCarriesAuthority() predicate + the 3 non-overlapping self-test assertions) is on origin/main — no Blocked-by needed. Re-check before dispatch: git grep -n 'MAINTAINER-ONLY' origin/main -- scripts/.

The per-gate-copy vs shared-helper-module question on the card is an implementation call for the executing lane, not a maintainer decision (no product semantics, no public contract, no gate verdict change). One caution for the dispatch brief: the shared-module route would also touch the two gates PR #8517 just fixed — same-day churn, merge current main first.

Size/model suggestion: M, opus.

---

Generated by Claude Code

[qq9340100]

Claimed by the domain:devx execution PM seat (#6023).

• session: session_01Jqe56GnYFddggeAyfkZFVz
• branch: claude/issue-8519-ratchet-remedy-authority-marker
• worktree: ../objectstack-8519 (dedicated, per repo)
• domain: domain:devx · container/model: os-dev / opus
• file surface: scripts/check-durability-degradation-log-level.mjs, scripts/check-role-word.mjs, scripts/check-driver-conformance.mjs (+ their self-tests)
• serial constraints: cleared — no other in-flight card in this lane touches scripts/. The only other open lane PR (docs(adr): anchor the default-active-org session hook to ADR-0093 D9 (#8474) #8529) is confined to docs/adr/0093-*.

Label swap pm:queue → pm:dispatched done in the same write as the assign.

Two notes on the record before the dev starts, so neither is mistaken for a finding later:

1. The origin/main re-check the triage comment asks for is delegated, not skipped. This seat does not read the shared primary checkout for verdicts — its HEAD is switched under it by other agents, and a reading taken there is an artifact, not a measurement (this seat has already published one such invalid reading today). The dev runs git grep -n 'MAINTAINER-ONLY' -- scripts/ in its own worktree cut from fresh main and reports the hit list as step 0. If PR fix(devx): mark ratchet-editing remedies as maintainer-only in both gates' own output (#8435) #8517's block is not there, the dispatch is wrong and the dev says so instead of proceeding.

2. The shared-module question is ruled, and ruled narrowly — see the dispatch brief. Scope is the three uncovered gates. The two gates PR fix(devx): mark ratchet-editing remedies as maintainer-only in both gates' own output (#8435) #8517 landed 30 minutes ago are not touched by this PR.

---

Generated by Claude Code

[qq9340100]

{
 "issue": 8519,
 "status": "done",
 "branch": "claude/issue-8519-ratchet-remedy-authority-marker",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/8539",
 "premise_still_valid": true,
 "summary": "STEP 0 (delegated origin/main re-check, run in my own worktree cut from 36d90fcbe): `git grep -n 'MAINTAINER-ONLY' -- scripts/` returns exactly two hits — check-engine-double-contract.mjs:1143 and check-type-check-coverage.mjs:1932, both `const RATCHET_AUTHORITY_MARKER = '⛔ MAINTAINER-ONLY';` — and `ratchetRemedyCarriesAuthority` is defined in both (1167, 1954). PR #8517's block is on main and the dispatch premise holds. All three named gates now carry the token, a detector regex, a pure predicate and self-test assertions. check-role-word.mjs had NO self-test at all, so one was added and wired into `check:role-word` (an assertion no script runs is a phantom check). check-driver-conformance's text MIRRORS the words already at its LEDGER comment rather than inventing new ones, per Zone 3. ONE ZONE 1 DEVIATION, flagged not silent: ruling #5 requires a changeset; I applied `skip-changeset` instead — see open_questions, this is the item I most want your ruling on.",
 "tests": "PREMISE + CENSUS: 76 scripts/*.mjs (card's count reproduces exactly), 9 ratchet JSON files, all in-script ledgers. VERDICT-INVARIANCE, both paths: (1) green run byte-identical before/after, stdout AND stderr, 6/6 IDENTICAL. (2) All three gates are GREEN on the real tree, so that only proves the PASS path — the edited strings live on the FAIL path a green run never reaches. So a violation harness planted one deliberate violation per gate (unguarded read seam in packages/objectql/src; unbaselined doc using the reserved word; a discovered driver package running no shared case-set) and removed it after. Exit codes 1/1 for all three before and after; flagged-subject sets identical (durability: file:line+guard+found; role-word: file+classification+count; driver-conformance: 9 driver x case-set pairs, md5 9f5416d27cf6c6c7 both sides, plus the full matrix table). Only difference on the violation path is the remedy prose, in exactly the three intended places. MUTATION TESTING, direction predicted FIRST for each: 13 mutations, all 13 caught by exactly the ONE predicted named assertion, #8435 failure count exactly 1 each — (a) drop the marker from the real message => only the marker assertion; (b) reword the offer past the detector => only the detector assertion (the marker check then passes vacuously, which is why the detector assertion exists); (c) predicate => `return true` => only the discrimination assertion; (d) break the synthetic fixture => only the fixture-sanity assertion; (e) widen check-role-word's detector to a bare `--update` => only the ratchet-DOWN over-reach assertion. I restructured the fixture-sanity/discrimination pairs to if/else after finding they double-fired, so 'non-overlapping' is now true rather than merely claimed. GATES: check:nul-bytes, check:role-word, check:driver-conformance, check:durability-log-level, scripts/pm/check-dispatch-gates.mjs all exit 0; self-tests OK/OK/35+35 cases. Re-derived gate families against actual changed paths via scripts/pm/dispatch-gates.mjs — surfaced check:driver-conformance, already run. Control-char self-scan on all four changed files: clean (grep exit 1). LANE RULES: pipefail on every gate loop, no `| head -N` on anything counted, flock on every heavy run, no git stash.",
 "open_questions": [
 {
 "question": "Zone 1 ruling #5 requires a `.changeset/*.md`. This PR releases nothing (root scripts/ dev tooling + the PRIVATE root package.json's script wiring; no workspace package source), and the repo's own changeset gate contradicts the ruling for exactly this shape. Implemented as `skip-changeset`; your call whether to overrule me.",
 "options": [
 "A. skip-changeset label (what I did). The gate's own failure message lists two routes and calls this one '<<< PREFERRED' for a PR that releases nothing. PR #8517 — the direct precedent, same three-gate shape — shipped scripts-only with NO changeset. Route 3 (empty frontmatter) is explicitly CLOSED for new files by a recorded ruling, so it is not available as a compromise.",
 "B. Add a real changeset naming a package. This is the only way to literally satisfy ruling #5, and it requires naming a package whose source this PR does not touch — publishing a version bump for a gate-message edit. Harmful and hard to reverse once merged.",
 "C. I add whichever you direct on the same branch; the PR is draft and nothing is armed."
 ],
 "recommendation": "A, and I did not treat this as re-litigating a 裁决 — I could not satisfy it without causing a spurious publish, so I took the reversible option and surfaced it rather than picking a side silently. If you want B, say the package and I will push it."
 }
 ],
 "out_of_scope_findings": [
 "filed as #8538: a SIXTH instance of this exact shape — scripts/check-test-typecheck.mts:147 offers 'Only with a reason: add the file to test-typecheck-debt.json' with no authority marker, against a ledger the same script calls an 'EXACT ratchet' that 'only ratchets down'. The census missed it by a NAMED MECHANISM, not oversight: the sweep globbed scripts/*.mjs (76 files, correct) and scripts/ also holds two .mts files that glob cannot reach. Filed unassigned/unlabeled for your triage, with the two borderline exemption-map cases (check-meta-type-normalized.mjs:271, check-tenant-chokepoint.mjs:544) recorded separately rather than folded into the count."
 ]
}

Assumptions A / B / C — verdicts.

A — the shared module. I think you are right, and I can sharpen the mechanism. A module exporting the marker and predicate is a deduplication, not an enforcement: it is reachable only by gates that already chose to import it, so it is structurally blind to the next gate that offers a ratchet path and never imports it. Concretely — it would not have caught #8538, which is exactly the case the convention needs to catch. The card's claim that the shared-module route "is the only one that makes the convention mechanically enforceable across the farm" is false, and inverted: a farm-wide detector is the only route with the property the card assigns to the module, and it needs no shared module to work (it reads source text, not imports). I found one thing worth adding to your framing: a shared module is not merely insufficient, it is mildly counterproductive for this convention, because the three predicates are not actually the same function — each gate's RATCHET_EXPANSION_OFFER is keyed to its own offer wording, and check-role-word's has a constraint the others do not (it must NOT match the ratchet-DOWN message, since both directions of its ratchet are spelled --update). A shared predicate would have to be parameterised by the per-gate regex anyway, leaving only the 24-character token constant genuinely shared. So the extraction you ruled out would have bought one string and cost same-day churn on two freshly-landed files.

B — detector feasibility. Measured, cheaply, and the answer is encouraging. A naive "offers to add an entry" sweep over all 76 .mjs + 2 .mts scripts surfaces ~15 candidate scripts; triaging by hand gives 6 true instances, 3 declaration registries the card already excluded, 3 more declaration registries it had not found (check-error-code-casing.mjs EXEMPT_FILES, check-single-authz-resolver.mjs ALLOW, check-skill-compatibility-version.mjs EXEMPT), 2 borderline exemption maps, and the rest refuse the weakening remedy outright or offer only the tightening direction (check-query-options-erasure-ratchet.mjs, check-slot-lookup-ratchet.mjs, check-i18n-coverage.mjs, check-startup-registry-verdict.mjs, check-driver-memory-census.mjs). So the naive false-positive rate is roughly 9 of 15, which is too high to ship as-is — but the false positives are not diffuse: they are all declaration registries, one namable class, and the six true instances all sit against a file or in-script structure that is independently identifiable as shrink-only (the script says "shrink-only", "only ratchets down", or "EXACT ratchet" in its own text). A detector keyed on shrink-only evidence rather than on "offers an entry" looks tractable. That is a design claim I did not build, per your instruction — the number is in #8538's neighbourhood and the detector itself is unfiled; say the word and I will file it as its own card.

C — the census of five. FALSIFIED, in the direction you suspected. Re-run from scratch: the real count is six, not five. All five of the previous seat's instances reproduce exactly (including both line numbers), so its work was sound within its stated scope — the miss is the scope itself, and it has a mechanism rather than a lapse: scripts/*.mjs cannot see scripts/*.mts. Filed as #8538 rather than folded into the diff, as you asked. Two further notes you should have: the card's exclusion list was correct but incomplete (three more declaration registries exist that it did not list), and check-durability-degradation-log-level.mjs — the file in scope — does not hold a second instance, which I checked specifically because a second baseline lives in it; its durability-degradation.baseline.json path only ever offers stale-entry deletion, which is the tightening direction.

Zone 1 compliance. No gate weakened, no verdict changed, proven on both the pass and fail paths. Scope held to the three named gates. check-engine-double-contract.mjs and check-type-check-coverage.mjs untouched. No shared helper extracted. No content/docs/releases/ edit. Fixes #8519 used, since all three are covered and self-tested. Draft only — not marked ready, not enqueued, no auto-merge armed. The one deviation is the changeset, above.

---

Generated by Claude Code

[qq9340100]

Addendum — assumptions A and B, measured. (The A/B/C verdicts were in the prose below the JSON in my report comment above; B is materially expanded here because I went and measured it rather than estimating.)

Measured against a fresh detached worktree at origin/main = f1da948d8 (which does not include PR #8539, so check-role-word.mjs correctly reads unmarked). Corpus: 79 scripts — 77 .mjs + 2 .mts. Note 77, not 76: main gained a .mjs gate since the card was written, which is itself an argument for a detector over a hand census.

A — your reasoning is right, and #8538 is the proof

A shared module is reachable only from gates that already import it, so it can only ever standardise the gates that already agreed. It has no way to observe a gate that offers a ratchet path and imports nothing — which is exactly what check-test-typecheck.mts was. My own #8538 is evidence for your side, not mine: the sixth instance was found by sweeping source text, and a module would not have found it no matter how well written, because the file never imports anything to be found by. Only a farm-wide detector has the property the card assigns to the module.

One thing to add that strengthens it beyond "insufficient": for this convention a shared module is mildly counterproductive. The three predicates are not the same function — each gate's RATCHET_EXPANSION_OFFER is keyed to its own offer wording, and check-role-word's carries a constraint the others do not (it must not match the ratchet-DOWN message, since both directions of its ratchet are spelled --update). A shared predicate would have to be parameterised by the per-gate regex anyway, leaving only the 24-character token genuinely shared. The extraction you ruled out would have bought one string at the price of same-day churn on two freshly-landed files.

B — feasible, with the false positives concentrated in one namable class

Three prototype detectors, scored against the hand classification:

detector	flagged	true pos	false pos	missed	precision	recall
v1 naive "offers to add an entry"	13	5/6	7	1	38%	83%
v2 = v1 AND the gate calls its registry shrink-only	5	4/6	1	2	80%	67%
v3 = v2 AND the gate does not REFUSE the remedy outright	4	4/6	0	2	100%	67%

Answer to "would it reproduce the hand classification": at v3, every gate it flags is one I flagged by hand — zero false positives across all 79 scripts, including all three registries #8519 excluded and both borderline exemption maps on #8538. The naive detector's 7 false positives are exactly the class you predicted plus two more I had already classified, and all of them are removed by two structural filters, not by a keyword blacklist: require the gate's own testimony that its registry is shrink-only, and exclude gates that refuse the weakening remedy outright.

Where it disagrees — 2 misses, and they have different causes. This is the part worth carrying onto a card:

• check-type-check-coverage.mjs — a prototype defect, fixable. It says "shrink-only" seven times and offers add a TEST_DEBT entry / add a DEBT entry. My regex demanded "add … to/in … REGISTRY" and this is "add … REGISTRY … entry" — pure word order.
• check-driver-conformance.mjs — a real limit of the v2/v3 anchor. That file contains zero shrink-vocabulary matches; its authority evidence is worded differently ("a MEASURED, tracked exception the maintainer has agreed to"). Both misses are in-script ledgers rather than JSON files, and in-script ledgers do not reliably describe themselves as shrink-only. A detector keyed only on that anchor will keep missing them.

Two by-products worth recording before they are lost:

• There is already a second, different marker token in the farm. check-type-source-resolution.mjs carries ⛔ SHRINK-ONLY on a registry that is shrink-only, and it refuses the weakening remedy outright rather than marking it. So a detector must not assume ⛔ MAINTAINER-ONLY is the only token, and the two conventions should be reconciled deliberately rather than by whichever gate is written next.
• My first-cut detector regex spelled its gap [^.;], which excludes the dot — so every offer naming a registry by path (…baseline.json) silently failed to match and two known instances read as clean. I caught it only because the hand corpus disagreed. That is the same failure the convention is about, one level up: a detector that under-reaches reports a green farm. Any real implementation needs the positive-control assertion the existing gates already use.

⛔ Not built into PR #8539, per your instruction. Say the word and I will file the detector as its own card carrying these numbers, the two miss causes, and the token-reconciliation question.

Item 1 — skip-changeset is applied. Read back live from the labels API just now: size/m, skip-changeset. Your read predated my write (I applied it about 20 seconds after create_pull_request, reading the existing labels first and writing the union so the bot's size/m was not wiped). PR #8539 is still draft, nothing armed, head 6d0df9d.

---

Generated by Claude Code