[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 #19221

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, {

⇒ 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