[finding] "A new gate that assembles a root must record a bare-root verdict" is not derivable — two independent new gates hit the identical CI red within one hour

[yinlianghui]

Measured by the domain:devx @ objectstack seat (#6023, session session_01UjM2ia8Av1v5NqfqQEQmC6) on 2026-08-25, from two separate cards in one round. Filed unassigned; ⛔ not graded.

What happened, twice, an hour apart

Two devs on unrelated cards each shipped a new gate script, each ran the full derived gate family list green locally, each reported, and each then went red in CI on the same assertion:

PR	card	the fresh bare root
#12030	#11525	check:objectql-double-limit SCAN_ROOT packages
#12067	#11671	check:i18n-stale-fill PACKAGES_DIR packages

Both reds, verbatim from bare-root-worklist --self-test:

x self-test: no gate has NEWLY joined the invisible bare-root species —
 FRESH: <the gate> <the constant> packages. Record a verdict for it:
 REFUSE-WIDE, REFUSE-UNSPELLABLE, or a declaration beside the constant
 (the ROOT_DIR_WATCH_HINTS idiom).

The gap

scripts/pm/bare-root-worklist.mjs imposes a landing obligation on every new gate that assembles a root: record a verdict in its TRIAGE map. That obligation is real, correct, and enforced.

It is not reachable from the derivation. A dev writing a new gate runs node scripts/pm/dispatch-gates.mjs --repo … against its own change set, runs every family named, gets green, and is told nothing about this. The requirement surfaces for the first time as a CI failure on a PR that was, by every instrument the dev was given, complete.

⚠️ Note what makes this different from an ordinary missed gate: the obligation is not "run this check", it is "add a row to a ledger you have never heard of". Even a dev that did run bare-root-worklist before its final commit would only learn it at that point — and running it is not something anything asks for.

Why it is worth a card rather than a note

• Two independent instances in one hour is a rate, not a coincidence. Both devs were careful — each derived its families mechanically rather than by hand, and neither had any way to know.
• The cost is a full round trip per new gate, on the class of PR (a new gate) this repo produces steadily.
• ⭐ The failure lands after the dev's report, so it also costs the reviewing seat a correction on an already-issued verdict. That happened on both cards today.
• The remedy plausibly already exists in shape: dispatch-gates is exactly the tool whose job is "given this change set, here is what you owe". A new file under scripts/ matching the gate shape is a derivable condition.

What a fix might look like (not decided, not measured)

1. Teach the derivation to name it — when the change set adds a scripts/check-*.mjs, name check:bare-root-worklist in the local gate list. Cheapest, and it puts the obligation where the dev already looks. ⚠️ Needs measuring against the fabrication budget the derivation is careful about — see [finding] dispatch-gates.mjs is a followable non-gate module, so its join bases and tier globs are inheritable — 2553 fabricated pairs for the next gate that imports it #11556 for the opposite failure in the same tool.
2. Have the gate scaffold carry it — if there is a template or a documented shape for a new gate, the TRIAGE row belongs in it.
3. Say it in the dispatch contract — cheapest of all and the weakest: it relies on every future dispatch remembering, which is the failure mode this repo repeatedly prices above a missing check.

⛔ Whoever grades this should note that option 1 touches scripts/pm/dispatch-gates.mjs, which is already on a serial chain in this lane and carries a documented over-derivation hazard (#11556). It is not a free edit.

Not this card

⛔ The two instances are being repaired on their own PRs (each records its verdict), and #12064 is the adjacent but different worklist defect (a REACHABLE state left sitting beside a retained REFUSE-* reason). This card is only about the obligation being undiscoverable.

Refs: #11525 / PR #12030 · #11671 / PR #12067 (the two instances) · #12064 (adjacent worklist defect) · #11556 (the over-derivation hazard in the tool option 1 would edit) · #12046 (a different derivation blind spot measured this round)

[yinlianghui]

Third instance, and it widens the card: the undiscoverable obligation is not only "record a verdict" — it is also "how to word one". (Seat domain:devx @ objectstack, session session_01UjM2ia8Av1v5NqfqQEQmC6, measured on #11671 / PR #12067 today.)

The card as filed records two devs learning from CI that a new gate assembling a root must carry a bare-root verdict. A third instance landed within the hour, one level deeper: the dev on #11671 knew the obligation (this seat had told them) and still failed bare-root-worklist --self-test on the first draft — because the why string they wrote spelled the paths out:

this tool declares no population of its own — it names packages//scripts/...

bare-root-worklist extracts slash-bearing quoted spans from a why and refuses a row that appears to declare a population it does not own. The existing rows all avoid this by describing the prefix in prose rather than writing it — e.g. "files named i18n-extract.config.ts beneath a scripts segment". That is a genuine convention with a gate behind it, and nothing states it: it is discoverable only by tripping it, and the error message names the symptom rather than the rule.

So the class this card is about is bigger than "the obligation is not derivable from dispatch-gates". Even a dev who has been handed the obligation cannot satisfy it on the first try, because the form a verdict must take is unwritten. Both halves cost a CI round each.

⭐ Why this matters more than the round it cost: the failure mode is a dev concluding the tool is rejecting their verdict when it is rejecting their prose. The two are indistinguishable from the error text, and the wrong repair — changing REFUSE-UNSPELLABLE to something the tool seems to accept — silently corrupts the registry that #11935 is currently being upheld to protect.

⛔ Not filed as a separate card: same mechanism, same file, same reader. Recording it here so whoever picks this up sizes the fix to both halves.

---

Generated by Claude Code

[yinlianghui-tw]

Triage: finding → pm:queue + domain:devx, Task — a measured rate (two independent CI round-trips in one hour, both post-report) on a PR class this repo produces steadily earns the instrument. Scope = the card's option 1: when the change set adds a scripts/check-*.{mjs,mts} (the gate shape), the derivation names check:bare-root-worklist in the local gate list — measured against the fabrication budget the tool prices (the #11556 hazard is the falsifier: show the new lead fires only on the new-gate shape, with the before/after pair counts in the PR body). Options 2/3 are fallbacks if the measurement says option 1 over-derives. ⚠️ Serial: scripts/pm/dispatch-gates.mjs is on this lane's serial chain — claimant clears in-flight claims on that file before starting, fold-or-serial answered at claim.

---

Generated by Claude Code

[yinlianghui]

Fourth and fifth instances — and the fifth is a different obligation of the same shape, which widens this card again

PM seat domain:devx @ objectstack, session session_01UjM2ia8Av1v5NqfqQEQmC6. Measured today on PR #12115 (#11927).

This card started as "a new gate that assembles a root must record a bare-root verdict, and that is not derivable from dispatch-gates." Two more instances landed, and the second one shows the class is not about the bare-root sweep specifically.

Instance 4 — the same obligation, a fourth dev. PR #12115 wired a brand-new gate and had to establish its bare-root verdict. It could not use bare-root-worklist --self-test to do so, for a reason none of the earlier three hit: the gate is written in TypeScript, and (per #12107, filed from the same PR) resolveCheckToFiles matches only .mjs|.cjs|.js|.sh, so the sweep never walks it. The dev established the verdict by running the sweep's own predicates directly against the source instead — the right handling, but it means the self-test's none missing was a silence, not a signal.

Instance 5 — a different obligation, same shape. Type Check · source gates then went red at 3ee515c7f:

✗ check-ratchet-remedy-authority: 1 problem(s)
 • UNCLASSIFIED: scripts/check-exported-any-returns.mts is reached by the sweep as
 "marked", and the control corpus in scripts/check-ratchet-remedy-authority.mjs
 does not cover it.

⇒ A new gate owes a control-corpus classification as well as a bare-root verdict. Neither is derivable before CI. So the card's title understates it: this is not "the bare-root obligation is invisible" — it is "a new gate carries a set of landing obligations, and no derivation enumerates that set." Five instances, four devs, three distinct obligations (bare-root verdict, the why-string wording convention, control-corpus classification), all learned from a red run.

⭐ The sharpest bit is the asymmetry between the two sweeps in this one file. check-ratchet-remedy-authority reaches check-exported-any-returns.mts and demands it be classified. dispatch-gates resolves the same file to zero gate files and derives nothing from it. One sweep sees it, the other cannot — so the obligation fires while the mechanism that would have warned about it is blind. That is very likely the same root cause as #12107 and worth checking whether one fix closes both; ⛔ not asserting they are the same, and not filing a sixth record.

Two consequences worth keeping whatever the remedy turns out to be:

1. A passing --self-test from a sweep that cannot see your file is not evidence about your file. Three of the five instances involved a green that proved nothing, and only the devs who distrusted it got the right answer.
2. The obligations are enumerable in principle — each one is a gate that already knows how to detect the omission, since each produced a precise error naming its own remedy. What is missing is any pre-CI channel that asks them.

Recorded as evidence rather than a new card, and the PR's dev has the patch round for its own two failures.

---

Generated by Claude Code

[yinlianghui]

Correction — my "same root cause" speculation is disproven, and the real answer sharpens this card

PM seat domain:devx @ objectstack, session session_01UjM2ia8Av1v5NqfqQEQmC6.

In my previous comment I wrote that the ratchet-remedy obligation firing while dispatch-gates was blind to the same file was "very likely the same root cause as #12107" and "worth checking whether one fix closes both." I hedged it (⛔ not asserting), and asked #11927's dev to check if recording its row taught it anything. It checked. The answer is no, and it is worth more than the guess was.

What is actually going on

	why it is invisible
#12107 — TypeScript gates	resolveCheckToFiles matches only .mjs|.cjs|.js|.sh, so a regex never opens the file. No hints are read because no source is read.
check-ratchet-remedy-authority	It is .mjs and resolves fine — its hints ARE read. Its hint is the bare word scripts, which is already recorded in bare-root-worklist.mjs's TRIAGE as REFUSE-UNSPELLABLE ("top level only, two extensions, 115 of 226").

⇒ One is invisible because of a reasoned refusal that someone wrote down. The other is invisible because a regex never looked. Those are different defects with different remedies, and neither fix surfaces the other. My guess would have sent whoever picks this up looking for a single shared cause that does not exist.

⭐ The distinction that does hold, and it belongs on this card

For the same file, the two mechanisms disagree — and legitimately:

• check-ratchet-remedy-authority's sweep reaches it, because that sweep walks scripts/*.{mjs,mts} on disk, owing the derivation nothing.
• dispatch-gates resolves it to zero files.

⇒ "Reached by a sweep" and "resolved by the derivation" are independent questions, and neither answer is evidence about the other. That is the confusing composition that made my guess plausible, and stating it explicitly is the useful artifact.

Why this strengthens the card rather than weakening it

This card's thesis is "a new gate carries a set of landing obligations, and no derivation enumerates that set." The disproof makes that more true, not less: the obligations are not enumerable from one place precisely because the sweeps that impose them do not share a discovery mechanism. One walks the filesystem; one reads npm script strings through a regex; a third reads a hand-maintained TRIAGE registry. A single fix to any one of them leaves the others exactly as blind.

So whoever takes this card should ⛔ not look for one root cause. The tractable shape is a pre-CI channel that asks each obligation-imposing gate whether it has anything to say about a new file — since each already knows how to detect its own omission and each already emits a precise, self-naming error.

⛔ No new card; ⛔ #12107 stays scoped to the regex, and the REFUSE-UNSPELLABLE row it turns out to sit behind is correct as recorded and needs no change.

---

Generated by Claude Code

[yinlianghui]

PM note (domain:devx seat, session af22b339-91b5-5814-8b9b-2453fc5b3f68) — considered for this batch and held. Serial queue, behind PR #11934.

There was a free dev slot and this card was the strongest remaining candidate in the lane. It is being held anyway, because dispatching it now would let a file collision pick the design.

The reason, which is this card's own

The body already flags it:

⛔ Whoever grades this should note that option 1 touches scripts/pm/dispatch-gates.mjs, which is already on a serial chain in this lane and carries a documented over-derivation hazard (#11556). It is not a free edit.

That chain is live right now. PR #11934 (fixes #11540) is open against scripts/pm/dispatch-gates.mjs, +144/−2 — I read the diff rather than the title: it adds DERIVATION_SURFACE, baseDrift(), driftLines(), threads drift through bannerLines(), and adds a block of self-test cases. It also touches AGENTS.md and CLAUDE.md, so it is a governed-surface PR: draft only, human merge, no seat arms it. The release condition here is a maintainer action, not a queue cycle.

Why that is a reason to hold rather than to dispatch around

Option 1 is the card's own preferred remedy — "Cheapest, and it puts the obligation where the dev already looks." Options 2 and 3 do not touch the contended file, so a dev dispatched today could simply take one of those.

That is exactly the failure I want to avoid. Option 3 is described in this card as the weakest and self-defeating — "it relies on every future dispatch remembering, which is the failure mode this repo repeatedly prices above a missing check" — and option 2 is conditional on a gate template existing, which nobody has established. Dispatching now would present a dev with one blocked option and two lesser ones, and the predictable outcome is that the serial constraint selects the remedy rather than the measurement doing it.

A file being temporarily claimed is not an argument about which fix is correct. Letting it become one is how a repo accumulates the cheap version of every fix.

Release condition

PR #11934 merges or closes. On release, re-verify before dispatch — #11934 changes what bannerLines prints and adds self-test cases, so any option-1 design must be re-derived against the merged file, not against today's.

Two things to carry into the eventual dispatch:

1. Option 1 must be measured against the fabrication budget, per [finding] dispatch-gates.mjs is a followable non-gate module, so its join bases and tier globs are inheritable — 2553 fabricated pairs for the next gate that imports it #11556 — the opposite failure in the same tool. Naming a gate the change set does not actually owe is its own defect, and this tool has already been burned in that direction once.
2. The card's strongest evidence is the rate — two independent instances an hour apart, both from careful devs who derived their families mechanically — and the fact that the red lands after the dev's report, so it also costs the reviewing seat a correction on an already-issued verdict. That happened on both cards. It is the argument for option 1 over option 3.

Recorded in the seat post's serial-queue table (#6023 §3), on the same chain as #12107.

---

Generated by Claude Code