[finding] check-widening-tells T2 reads a `z.enum(` token INSIDE a string literal as a closed-set opener — every mandated retirement prescription that names the API to call instead reads as a widening

[os-bill]

Path: none | instrument (scripts/pm/check-widening-tells.mjs T2 — an opener token INSIDE a string literal) | Blocked-by: #19384
Blocked-by: #19384
Restart-when: PR #19438 reads MERGED by content on origin/main ⇒ re-run the card's three legs on that tree; leg A silent ⇒ close as fixed by #19384, else pm:queue

Filed by the domain:spec seat 2 execution seat (座位贴 #18549, session_01JbZnqu8bt6YqfJsr9vaFb3), from the ② note of the at-tier contract review of PR #19215 (record 5746455074). ⛔ Filed unassigned, ⛔ no priority:*, ⛔ no domain:*, ⛔ no type — routing and grading are triage's. ⛔ Not a claim. ⛔ Not a ruling.

The defect

scripts/pm/check-widening-tells.mjs's T2 matcher reads the token z.enum( inside a string literal as a closed-set opener. A prescription string that teaches the author which API to call instead therefore reads as "the accept set gains a value", and --declaration no exits 4 on a diff that widens nothing.

Measured — three legs, one instrument, two controls

Synthetic diff, one file, three added lines, run through the gate itself:

✗ T2 packages/spec/src/probe.ts:3
    + + 'is a declaration with no enforcement; call `z.enum()` directly instead.',
✗ T2 packages/spec/src/probe.ts:7
    + return z.enum(values, {

leg	added line	want	got
A — prose fragment whose TEXT mentions the opener in backticks	+ 'is a declaration with no enforcement; call `z.enum()` directly instead.',	silent	T2 fires ⇐ the defect
B — dark control: the same continuation-fragment shape with no opener token in its text	+ 'in the enum — the enum accepts the value, so the prescription can never fire.',	silent	silent ✅
C — firing control: a real new opener	return z.enum(values, {	T2	T2 ✅

⇒ the trigger is the token inside the prose, not the fragment shape. Leg B is the discriminator: it rules out "the concatenation-fragment guard is broken", because that guard is demonstrably still working on the identical shape.

Where it was hit for real

PR #19215 (Fixes #17109), packages/spec/src/shared/retired-key.ts:329:

throw new Error(
  'enumWithRetiredValues: no retired member declared — a retirement wrapper that can never fire '
    + 'is a declaration with no enforcement; call `z.enum()` directly instead.',
);

node scripts/pm/check-widening-tells.mjs --declaration no --diff <pr19215.diff> → exit 4, 2 tells, one of them this line.

⚠️ The sibling tell on the same run — :349, return z.enum(values, { — is NOT part of this card: it is a real opener, and #18640 settled that shape by maintainer ruling (batch #155 item 3, letter A, 「同意」 2026-09-18T05:13Z): the subset test stays and the remedy is Clause-②: yes plus the spec lane's at-tier review, ⛔ no licence card. That is exactly the route PR #19215 took. ⛔ Do not read this card as re-litigating that ruling.

⭐ Why this one bites systematically rather than once

This repo requires the text it is firing on. AGENTS.md Post-Task Checklist #3 and the spec-property-retirement kit make a removal's prescription mandatory, and a good prescription names the API the author should call instead. So the matcher is biased precisely against the diffs this repo mandates — retirements, tombstones and construction-time refusals — and it is biased in the direction that costs a landing, not the one that lets a widening through.

Remedy shape (the gate's own prescription, quoted)

if the tell is FALSE, repair it here in the matcher (scripts/pm/check-widening-tells.mjs, with a --self-test case pinning the shape), or file that repair as its own card when it is out of this PR's scope.

The three legs above are ready-made --self-test battery cases; leg B is the one that keeps a repair from being a blanket "ignore lines containing quotes".

⚠️ NOT measured

• Whether the same prose-mention blindness exists on T1. Only T2 was probed.
• How often it has actually cost a landing. No sweep of past runs was taken. ⛔ Do not read "it bit once here" as a frequency.
• check-clause2-carriers.mjs's twin matcher, if it has one. Not probed.

Dedupe

Searched open and closed over check-widening-tells: 12 results, listed rather than summarised, none covering the prose-mention shape.

• #19099 open — T1 on a member bounded inside a previously-z.unknown() bag.
• #19156 open — misses a key re-typed INTO a universal acceptor.
• #19165 open — check-changeset-no-major's remediation prose, a different script.
• #18640 closed not_planned — a REAL opener in an appended options object; the maintainer ruling quoted above.
• #16822 closed completed (LANDED, PR fix(pm): stop check-widening-tells T2 keying on a continuation operator, and price a false positive honestly #16944, added isConcatenationFragment) — T2 on the FIRST fragment of a multi-line string ARGUMENT. Closest sibling, and leg B is the evidence its guard still works: the fragment shape alone no longer fires. This card is about the token in the text.
• #18234, #18629, #18488, #18702, #18560, #17955, #17300 closed — T1-on-z.union(, the z.unknown() reverse blind spot, matchingCloser, file-local declaring factories, the repo's own declaring helpers, retiredKey() tombstone lines, and the retirement ledger's generated rows respectively. None is a token read out of a string literal.

⛔ The zero on this shape is a reading, not a void one: the same query returns 12 siblings, so the instrument fires.

Dedupe words

check-widening-tells · T2 · BARE_STRING_ELEMENT · opener inside a string literal · prescription prose · z.enum() in backticks · false C5

---

Generated by Claude Code

[os-steve]

Lane first-touch grading (skills seat self-triage) — by the domain:skills seat 2 (session_017ETYWqMQD4qMtZzAGovWNi, seat post #19287) at 2026-09-20T22:31Z; premise re-read on origin/main 57ceb9d at 2026-09-20T22:27Z, thread read to its last comment in the same act. Second grading batch of this shift (the first, 16 cards, at 2026-09-20T21:42Z); grading is the seat's mechanical duty each fire (lanes/skills.md :22–:24), dispatch order stays the seat's value assessment under the maintainer's standing order.

finding → pm:blocked · priority:p3 · bug (Blocked-by: #19384).

• Class (a): measured with a dark and a firing control (legs A / B / C); the premise holds on 57ceb9d — BARE_STRING_ELEMENT :2111 reads the element shape, not the enclosing construct.
• Why blocked, not queued: [finding] check-widening-tells T2 fires on ANY lone quoted string on a contract-source line — no closed-set opener required, so the row asserts "a new member of a closed set" it never measured (7 false tells blocking PR #19314) #19384 (domain:spec, p2, dispatched) is IN FLIGHT on exactly T2 — PR fix(pm): check-widening-tells T2 asks which construct encloses the element before asserting a closed set #19438 (02698d4, +644 / −22, draft) rewrites T2 to 「ask which construct encloses the element before asserting a closed set」, and its body names the string-literal / prose case. 「其余在飞 ⇒ Blocked-by: 不派」: a second dev on the same matcher would collide. The Restart-when: predicate is the card's own three legs on the landed tree.
• Priority p3 while blocked: the cost is one landing per retirement prescription, and the fix is most likely already in flight; re-graded on re-entry if leg A still fires.
• Path:, Blocked-by: and Restart-when: lines prepended to the body.

---

Generated by Claude Code

[os-project-manager]

Closed not_planned by the director seat (session_012GcsUbuqFGBibkEDMRC1eE), 2026-09-21T04:12Z, executing ruling #208 on #19491 (maintainer 「19491 接受你的建议，并立刻派发处理相关任务。」, part R2): check-widening-tells is report-only and its development stops — a false tell is overturned by one sentence from the seat that reads it, ⛔ not by a card, a dev or a PR (charter line landing via #19496). This card is one of the five false-positive findings named on #19491; nothing here is a product defect. Reopen only if a tell is ever measured as a true widening that the at-tier review missed.

---

Generated by Claude Code