fix(pm): check-widening-tells reads the member’s PRIOR schema — a bound inside a previously-z.unknown() bag is not a new key (#19153)

Fixes #19099

Clause-②: no — one file under `scripts/pm/**`. No published schema,
writable key, enum member, error code or exported type moves.

`check-widening-tells` declines a T1 key line when the member is
**bounded inside a bag the same change block re-declared out of a
universal acceptor**. The matcher read the added key TEXT and never the
member's prior schema, so it could not tell a bag-internal bound from a
new key — and on this board the criterion-honest `Clause-②: no
(narrowing)` was therefore the BLOCKED declaration.

---

## ⛔ Body replaced by the seat at 2026-09-20T15:55Z, for head
`95a7e8c0bf` — the FOURTH version, and this one exists because the PR
got SMALLER

Ruling **D′** (`5748934308`, batch #192 item 4) cut this PR down. The
delivering dev writes a body once and ⛔ never patches it, so the rewrite
is the seat's act; ⛔ what each earlier version claimed is preserved as
HISTORY at the end.

**+937 / −43 → +292 / −27.** One file, 319 changed lines.

⭐ **The method is the part worth reviewing, because it is what makes the
reduction checkable.** The file was restored **byte-identical to
`origin/main`** and the `unreadable` half re-applied on top. ⇒ what
remains is *provably* only that half. ⛔ This was **not** a surgical
removal whose leftovers nobody can enumerate — a reviewer does not have
to take an inventory on trust.

### LANDED

`enclosingDelimiters` answers the whole shown stack beside an
`unreadable` flag, with the two sound discriminations: a `*/` outside a
block comment says the hunk **began** inside one, so the comment-pushed
frames are discarded; and a lone `/` is the division it is.

### DROPPED — enumerated, ⛔ not summarised

the bag-internal decline · `samePlace` / `framePlace` / `shownPlace` /
`placeAt` · the place parameters on `replacesUniversalAcceptorKey` · the
removed-run reading on `declaresUniversalAcceptorKey` · the
`removedPlaces` plumbing in `tellsInFile` · the 67-case `#19099`
battery.

⚠️ **The last two of those go beyond a literal reading of 「drop the bag
decline」, and the dev said so rather than letting it pass as scope.**
The seat's answer, on the record: **the dev read it correctly and the
wider reading is the right one.** The removed-run reading would make
#18234's existing decline *quieter* in a class it did not previously
cover — a behaviour change D′ does not authorise, and one that would
have to be judged on its own evidence. ⇒ dropping it is also exactly
what lets the corpus below read **0 differing rows**. If #18234's
overturn condition is wanted, it is a **separate, additive** change on
its own card; ⛔ it does not ride this one.

## The two round-3 defects the ruling named — both discharged

1. **The false header sentence.** The residual paragraph is rewritten
and now states **both** triggers with the rate each occurs at,
re-measured on this tree. ⛔ The earlier sentence called one shape 「the
whole residual」; naming one while another dominates is the precise
failure being corrected.
2. **`!openedBlockComment` was pinned by nothing.** It is now pinned by
one case that goes red when the conjunct is removed — ablation leg 1,
**1 of 525**.

**Ruling item 3** — the shape's disposition is carried in
`REFUSAL_SENTENCE`, so a refused author reads why the gate fires and
that the remedy is the tier. Verified end to end: the live pair exits
**4** and prints it; the same diff declared `Clause-②: yes` exits **0**.

## 🔴 WITHDRAWN by the seat — the residual measurement was a
SHALLOW-CLONE ARTEFACT, and the card's premise was right all along

⛔ **This section previously claimed the re-measured residual 「DISAGREES
with this card's own premise」 and offered a two-populations
reconciliation. Both are withdrawn.** Round-4 at-tier review (record
`5751047556`) falsified them, and the seat wrote the withdrawn text, so
the withdrawal is the seat's act and is recorded here rather than
quietly deleted.

**The artefact.** The corpus replay ran over a clone whose history is
**shallow**. Git diffs a grafted boundary commit **against the empty
tree**, so `ae8edd2c4f` (the boundary in `.git/shallow`, 2026-08-31)
rendered locally as **「1092 files changed, 309028 insertions(+)」** — 692
non-test `.ts` files under `packages/spec/src`. Via the API that commit
is really **2 files, +41/−7**, a docs gloss.

⇒ **3141 of the 3143 flagged rows came from that one phantom diff.**
Every regex-slash trigger sat on an added line of a diff that does not
exist.

**The same measurement with every shallow boundary excluded** (real
history, same tip):

| | withdrawn reading | corrected reading |
|---|---|---|
| commits / file diffs | 436 / 1841 | **435 / 1149** |
| corpus identity | 14295 = 14295, 0 differing | **484 = 484, 0
differing** (T2 125 · T1 270 · T4 89 — the gate’s read path, confirmed
round 5) |
| T1 rows | 9482 | **270** — ⭐ round 5 closed the 265-vs-270 gap in
favour of 270 |
| `unreadable` | 3128 (**33.0%**) | **2 (0.7%)** |
| regex-slash trigger | 3069 rows / 72 lines | **0** |
| apostrophe trigger | 59 rows / 3 lines | **2** — `approval.zod.ts:884`
`Entra's` · `solution-blueprint.zod.ts:190` `object's`, both JSDoc prose
|

⭐ **The corpus-identity claim SURVIVES** — 0 differing rows on both
populations, and it is the load-bearing one: the reduced diff still
moves no tell anywhere. What died is the *residual rate*, which is the
one thing ruling D′ item 1 ordered corrected.

⚠️ And the 「59 apostrophe rows from 3 lines」 was wrong **in kind**, not
only in size: 2 were apostrophes, and **57 came from a single CODE
line** in the phantom diff (`api/contract.zod.ts`, a nested template
literal) whose first trigger is the type-blind pop, ⛔ not an apostrophe
at all.

⇒ **The stand-down disposition's reading stands** (apostrophe live,
regex 0 of 250). There were never two true populations — there was **one
broken instrument**. The seat repeated the dev's framing to the board
before this review ran, and that was wrong.

⛔ This is precisely the trap `AGENTS.md` routes through
`scripts/pm/git-history.mjs`: **a shallow clone answers at exit 0 with
no warning.** A corpus reading taken without excluding `.git/shallow` is
not a small overcount — it is a reading of a diff that never happened.

~~**Owed on the next head, ⛔ not fixed in this body:** the file's own
header paragraph still carries the withdrawn numbers and tells the next
author 「THE NUMBER THAT MATTERS is the 33%」 and that a declining reader
would refuse on a third of all T1 rows. On real history that is
**0.7%**. That paragraph is the dev's to rewrite (review items 1–3);
this body is the seat's and is corrected here.~~

⇒ ✅ **DISCHARGED at head `0d69cf2e8d`** — the header paragraph was
rewritten in round 5. Detail below.
## ⚠️ The flag has no suppressing reader, and none was manufactured

With the decline out, `enclosingDelimiter` deliberately collapses
`unreadable` to null for its two firing-only callers — an existing,
reasoned, three-rounds-old decision the dev did not touch. **Disclosed
at the definition and in the header** rather than left for the next
author to find.

A reviewer may reasonably read the exported flag as infrastructure
without a consumer. The counter-argument, stated so the review can weigh
it rather than discover it: the flag is load-bearing **inside** the walk
— it is the `!unreadable` guard that keeps the reset from discarding
real frames (ablation leg 3) — and the reset itself does reach a caller
(the `inParameterList` case, ablation leg 4).

## Measured

🔴 **CORPUS — the numbers below are WITHDRAWN; the live values are in the
withdrawal section above.** Round 5’s review (`5751378085`) caught this
section still asserting `436 / 1841 / 14295 = 14295` as a **live
claim**, ⚠️ **positioned ABOVE the withdrawal that retracts it** — so a
reader met the false figures first and the retraction second. That is
the seat’s defect, ⛔ not the dev’s, and it is corrected here rather than
deleted. **The measured values, on the gate’s own read path**: **435**
non-merge commits (one shallow boundary excluded), **1,149** file diffs
of non-test `.ts`, identity **484 = 484 rows, 0 differing commits** (T2
**125** · T1 **270** · T4 **89**), 1,133 of 1,149 blobs resolved. ⭐ The
**claim** this section exists to make is unchanged and survives every
instrument tried: **0 differing rows** — the reduced diff moves no tell
anywhere on the corpus.

**ABLATION — 8 legs, ⛔ all red**, through
`scripts/ablation-replace.mjs`, which verifies the mutation on disk and
restores; every restore proven twice (`git diff HEAD` empty AND `git
hash-object` equal to `git rev-parse HEAD:path`):

| leg | mutation | red |
|---|---|---|
| 1 | `!openedBlockComment` removed from the reset guard | 1/525 — the
ruling's second round-3 defect |
| 2 | `!leadingCommentClosed` removed | 1/525 |
| 3 | `!unreadable` removed | 1/525 — the guard protecting real frames |
| 4 | reset condition forced false | 3/525 |
| 5 | lone-slash discrimination widened so every slash flags | 1/525 |
| 6 | unterminated-string branch made to swallow | 2/525 — includes THE
RESIDUAL case |
| 7 | type-blind pop flag deleted | 1/525 |
| 8 | disposition removed from `REFUSAL_SENTENCE` | 1/525 |

**PROBES, each with its `origin/main` control leg.** The live pair
(#19095's shape) exits **4** with 2 T1 rows at the bag members —
identical to main's 2 rows. #18234's sibling decline still exits **0**.
The same diff declared `Clause-②: yes` exits **0** and is never blocked.

⚠️ **One probe did NOT return what the dispatch asked it to confirm, and
that is reported, not reconciled.** The surplus control — a genuinely
new outer key beside a re-declared bag — reads **2 rows on BOTH legs**,
⛔ not the **1 row** the dispatch's ZONE 2 asked for. The reason is
sound: 「1 row」 was a property of the **dropped** decline, so with the
decline gone the reading is main's, and both legs agreeing is the
correct outcome. ⇒ the dispatch's expectation was written for the larger
diff and did not survive the reduction.

## ⛔ Incidents, reported rather than tidied away

⭐ These are in the body on purpose. A repair whose own history is edited
is worth less than one whose history is legible.

1. **Commit `d603731b4f` does not parse.** Writing about the byte the
walker now reads, the author put a **literal block-comment terminator
inside JSDoc**, closing the comment at that word; `--self-test` exited 1
before reaching a single case, and the exit code was read only **after**
the push. Fixed in `e669c1bda7` with the backslash spelling the rest of
the file uses — ⛔ not amended away.
2. **One ablation leg ran while its new pin was still UNCOMMITTED.** The
leg's restore is `git checkout HEAD -- PATH`, so it reverted the edit
and the pin simply vanished — exit 0, clean tree, ⛔ **no warning**.
Re-applied, committed, re-run from the committed state.
3. ⭐ **One leg came back GREEN, and that is how a missing pin was
found** — the `!unreadable` guard was pinned by **nothing**. Recorded
loudly rather than quietly re-run; the case that pins it now fails its
leg.
4. **This round, leg 5 ran VOID on its first spelling and is reported
rather than quietly re-run.** The replacement text was a substring
already present 4 times, so the tool refused it for a rise of 0 and the
self-test never ran. Re-spelled distinguishably, then red.

## Verification

`dispatch-gates --commands` re-derived **in this worktree from the REAL
changed paths** → **30 families, all exit 0**, each code captured into a
file **before any pipe**, reconciled with `--ran` carrying `command ::
exit N` for all 30: **30 derived, 30 run, 0 NOT-MEASURED, 0 UNRUN** ⇒
the tool reports that zero as **DERIVED**, not claimed.

`--self-test` exit **0, 525 cases** (508 is `origin/main`'s count; +15
the new `#19099` battery, +2 the refusal-sentence cases). Battery
floors: new roster entry at **15**; the refusal-sentence battery raised
**11 → 13**; the 67-case bag battery deleted with its decline.

`eslint . --no-inline-config` at this head: **6943 files, 0 errors, 0
warnings** — the WHOLE population, so ⛔ no narrowing is claimed and none
is owed. `check:nul-bytes` exit 0, plus an independent `grep -naP` over
the control-character class on the edited file: exit 1, no match.

⚠️ **One correction, stated rather than buried.** `pnpm
check:pm-dispatch-gates` first returned **124** — the dev's own 300s
then 540s timeout cap, ⛔ **not a red**. Re-run detached to completion:
exit **0**, 1883 cases, 853.4s. Collected by a foreground blocking wait
on the recorded PID, ⛔ never by a watcher.

⭐ **STALE TREE discharged with a ROSTER, not a count.** `origin/main`
was merged into the branch **exactly once** (`8271c8142`, 15:01Z, clean,
0 conflicts). It moved twice more during the round (`a88a9733`,
`3ff6ddf1`) and was ⛔ deliberately **not** re-merged; the 30-family
roster was re-derived against the newer main and diffed against the
first derivation — **byte-identical**, so the upstream move added and
removed no family.

## NOT MEASURED, declared

objectui's mirror under any matcher · the place limits of
`rewritesExistingOpener` and `respellsExistingClosedSetBinding` — ⛔
probed by nobody in any round, stated as a gap rather than a verdict ·
merge-queue convergence · CI convergence, which is the seat's to read ·
**whether a reader that DECLINES on these frames is viable at all**: on
real history it would refuse on **0.7%** of the T1 rows this gate
reports (2 of 270) — ⛔ not the 33.0% this body claimed before round 4
falsified it as a shallow-clone artefact (see the withdrawal above),
which is a design constraint on whoever takes the parsed-schema
comparison the ruling points at. ⛔ Not a defect of anything that exists
today — nothing suppresses on the flag — so it is not a filing class; it
is written into the header at the definition so it cannot be missed.

**Attribution (prose, because the edit side of a PR-body write always
appends its own footer):** this body was written by the `domain:spec` PM
**seat 3** in session `session_01HnRAeVTLJevtQ5iCPX6JSm`; the change was
implemented by the dispatched dev on branch
`claude/issue-19099-widening-tell-prior-schema`. The three earlier
versions were written by seat session
`session_01AmH9bKvGoLjiY86Q4Z3og2`.

---


### Round 6 — the harness was fixed before the numbers, and that is the
finding

Head **`059f5e9aafa0`**, prose-only again: **0 non-comment changed
lines**, comment-stripped file byte-identical at **197,712 bytes**,
self-test unchanged at **525**. Gates **30/30 exit 0**, roster
byte-identical to round 5's.

⭐ **The root cause, reproduced directly rather than inferred.** `ROOT`
in this module is `fileURLToPath(new URL('../..', import.meta.url))`.
The round-5 rig kept module copies in a scratchpad **outside any
worktree**, so `gitAt` ran `git -C /tmp/… cat-file` — which fails, **is
caught, and answers `null` with nothing raised**. Probed both ways at
this head: the *same* `headBlobSource` call reads ****336,382 code
units** (⚠️ 「bytes」 was loose — it is a UTF-16 `.length`; the round-6
review located the blob: `packages/spec/src/ui/view.zod.ts`, 339,808
bytes on disk)** from a copy inside the worktree and returns **`null`**
from the copy outside it.

⇒ two silent defects, ⛔ not six wrong figures:

| | effect |
|---|---|
| `tellsInFile` alone, without `wideningTells(files)` | drops
`ledgerRowLicences` (#17300) ⇒ every ledger row in
`migrations/registry.ts` tells ⇒ **T2 inflates** by 243 |
| `readSource` answering `null` | drops #18702 file-local factory
resolution ⇒ **T1 deflates** by 5 |

⚠️ And the earlier claim that looked like a control was not one: 「0
unresolved」 was **true about line indices** and said nothing about
blobs. ⇒ **The rig now asserts resolution on a known blob before it
counts anything, and refuses to report if it comes back empty.**

**No disagreement this round, and re-derived rather than transcribed.**
Every reviewer figure came back identical on the dev's own instrument
once it ran the gate's read path: **484 = 484, 0 differing** (T2 **125**
· T1 **270** · T4 **89**), **1,133 of 1,149** head blobs resolved, 22
unresolved on both legs, `unreadable` **2 of 270 (0.7%)**, comment-start
hunks **35 of 270 (13.0%)**, **1,939** non-empty of 2,298 sides. Both
apostrophe sites confirmed a **third** time.

⇒ **270 is the better reading** for the stated reason: the paragraph
claims to describe 「the T1 rows this gate REPORTS」, so it has to be true
of **the gate**, and 265 answered a different question.

⭐ **What the file gained beyond the six corrections:** the horizon
paragraph now carries **the read path as part of the horizon**. A census
must call `wideningTells` with `headBlobSource` live, because
`tellsInFile` alone drops the licences and a null `readSource` drops
factory resolution — and because `ROOT` comes from `import.meta.url`, a
copy imported from outside a worktree fails every `cat-file` silently. ⇒
it now tells a census to **assert its own resolution before it counts**.

⚠️ **One gate incident, disclosed and handled the expensive way.** The
dev's first family pass returned **exit 3 on six families** —
「PREREQUISITE NOT MET — the dependency yaml is not installed」 — because
it recreated the worktree and measured before `pnpm install`. Read as
**NOT MEASURED, ⛔ never as a red**. It then re-ran **all 29** short
families from scratch rather than only the six, and re-ran the long
family post-install (exit 0, 1883 cases, 860.1s) rather than **reasoning
about whether** the pre-install run that had already exited 0 was
degraded. ⭐ That is the right instinct: a reading whose conditions
changed is re-taken, ⛔ not argued about.

## HISTORY — the two justifications this body used to carry, both
REFUTED

1. **Round 1 — fact 2 as a NAME reading**, with 「the same BLOCK removed
that key carrying a universal acceptor」 and no requirement that the bag
sit where the removed key sat. Refuted by three git-emitted diffs.
2. **Round 1 — 「this one leaves no 'but this member widens' case
open」.** False as written; the open case was a bag at another path, and
it was unbounded.
3. **Round 2 — the place as opener + head TEXT.** Refuted: two keyless
parents spelling their opener identically read as one place, so the same
unbounded silence returned for `discriminatedUnion` arms, tuple members
and `.or()`/`.and()` arguments.
4. **Round 2 — 「a division operator raises it too」** as the flag's
disclosed trigger. Measured at **0 of 21** raisings; the real one, ` */`
at a hunk that begins inside a comment, was **21 of 21** and
undisclosed.
5. **Round 2 — 「byte-identical」** for the firing-only readers. Textually
false (`enclosingDelimiter`'s body differs by two lines); the true and
now-stated claim is **answer-identical**, over 431,750 side-lines.
### Round 5 — the correction landed, and it carries TWO unresolved
disagreements rather than a reconciliation

Head **`0d69cf2e8d`**, **prose-only and proven two ways**: no changed
line is a non-comment line, and the comment-stripped file is
**byte-identical** to `95a7e8c0` at **197,712 bytes on both sides**.
Self-test stays **525/525**; the 30-family roster diffs byte-identical
against round 4's; all 30 exit 0.

⭐ **F1 was confirmed on the dev's OWN instrument, ⛔ not adopted from the
review.** `cat .git/shallow` names exactly one boundary, `ae8edd2c4f`;
locally it renders against the empty tree at **1,092 files / 309,028
insertions** scoped to `packages/spec/src` — byte-identical to the
reviewer's figure — while the API reports **2 files, +41/−7**. Two
instruments, one finding, agreeing on the load-bearing facts: the
boundary sha, its phantom render, 1,149 file diffs, **0 differing
commits**, `unreadable` = **2**, regex-slash = **0**, and **both
apostrophe sites file for file**.

⚠️ **Two numbers still disagree, and they are printed rather than
smoothed** — the instruction was to report a disagreement, ⛔ never to
reconcile to the reviewer:

| | dev (round 5) | reviewer (round 4) |
|---|---|---|
| corpus identity rows | **722 = 722** (T2 368 + T1 265 + T4 89 —
*every* tell kind) | **484 = 484** |
| T1 denominator | **265** | **270** |
| `unreadable` rate | 2 of 265 = **0.8%** | 2 of 270 = **0.7%** |

The numerator is **2 on both sides** and both name the same two
JSDoc-prose apostrophes, so ⛔ nothing load-bearing turns on this. The
dev checked the obvious gap on its own side — **all 265 rows resolved an
index, 0 unresolved** — so the residual **5 rows are unexplained from
here**, and it declined to quote a figure it had not measured. The file
carries **265** with its population spelled out so the number is
checkable. ⇒ round 5's verification owns closing it.

### 🧾 Two prose figures moved that the review did not name — and the
seat's answer is that moving them was RIGHT

The dev flagged this itself against a bound that read 「nothing else in
the code moves」. ⛔ Nothing executable moved for either.

1. 「` */` fired on **24%** of readable T1 stacks」 — rested on the same
contaminated corpus as the four phrases the review ordered deleted. ⭐
The dev did **not** restate round 1's number, because it could not
re-measure round 1's behaviour; it replaced an **unverifiable rate**
with **the population that rate is about**, measured on the clean
corpus: a hunk that begins inside a comment is **34 of 265 T1 rows
(12.8%)**.
2. 「431,750 side-lines (1,881 sides of 250 real commit diffs)」 —
re-measured as **84,924 side-lines over 2,298 sides** of the 1,149 file
diffs, **0 disagreements** for each firing-only reader.

**Seat's ruling, on the record:** ⛔ leaving a knowingly-false rate in
the file **during the round whose entire purpose is removing false
rates** would be incoherent, and the bound's word was *code*, which did
not move. ⭐ And (1) is the better instinct: when a rate cannot be
re-measured, the honest replacement is the **measured population**, ⛔
not a deleted sentence and ⛔ not a carried-over number. Both stand.

⚠️ One sentence the file now makes explicit, worth naming because it is
the trap this whole arc was: the regex-slash **zero** must ⛔ **NOT** be
read as 「regex literals are rare here」. They are common. The flag is
deliberately conservative, and a zero on a real corpus is a fact about
**this window**, ⛔ never about the shape.

---
_Generated by [Claude Code](https://claude.ai/code)_

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-steve