Four unreleased changesets still describe the deleted api-surface-declarations family, and #18991's stated reason to publish rests on a files[] entry that no longer exists

[os-litant]

Path: none | 未发布 changeset 的散文(发版即固化) | 四条 changeset 仍在描述已删除的 api-surface-declarations 家族;#18991 的发布理由已落空
分诊重测与定级:2026-09-20T15:32Z

What was measured

Read first-hand on origin/main at 2026-09-20T14:32Z. PR #19024 landed as 2277d1fcd10 — revert(spec): take back the declaration-text snapshot, restore the 27 signature hashes — deleting the whole packages/spec/api-surface-declarations/ family (17 shards plus its build script) and dropping the entry from packages/spec's files[].

Four unreleased changesets already on main still describe that family in prose, and changeset prose is compiled verbatim into CHANGELOG.md at the next release:

.changeset/17108-element-text-variant-published-nine.md
.changeset/18605-enable-on-install-one-authority.md
.changeset/18991-user-export-slot-is-a-real-optin-grant.md
.changeset/19085-metadata-form-declared-rows.md

Controls on that reading: 98 changesets on main name packages/spec (lit), a fabricated artefact name matches 0 (dark).

⭐ The sharpest one is not a stale mention — it is a stale justification

.changeset/18991-user-export-slot-is-a-real-optin-grant.md:20:

Why this publishes rather than taking skip-changeset. Two entries of this package's files[] move. api-surface-declarations/ ships, and the member docblock sits inside the ResolveApiOptions interface body, so it is part of t…

⇒ the changeset's stated reason to publish at all rests on api-surface-declarations/ being in files[]. Measured on main just now, it is not:

files[] = ["dist","json-schema","liveness","prompts","llms.txt","README.md",
           "src/**/*.zod.ts","CHANGELOG.md","api-surface","spec-changes.json"]

⚠️ ⛔ This card does not claim the publish decision is now wrong — the same sentence names a second reason (the member docblock inside the interface body), and whether that one alone carries it is exactly what needs a read. What is certain is that a shipped CHANGELOG entry would tell a reader the package ships a directory it does not.

Why it is a card and not a rider

All four arrived on the #19237 fix round's branch with the merge from main; none is in that PR's diff. Editing other seats' changesets inside a conflict-resolution round would widen it, and the round declined — correctly. ⇒ recorded here so it survives.

Carrier

The seat that owns the revert 2277d1fcd10 (PR #19024) — these sentences are its wake. Failing that, whoever compiles the next release's CHANGELOG.md, which is the moment they reach a consumer and the last moment they can be fixed cheaply.

⚠️ There is a clock on this: once the release compiles, the prose is in a published CHANGELOG and correcting it means another release note, ⛔ not an edit.

Dedup words

api-surface-declarations changeset prose · 18991 files[] publish reason · 2277d1fcd10 revert wake · stale changeset CHANGELOG · declaration snapshot removed

Origin: the #19237 conflict-resolution round's out_of_scope_findings, which named it 「noted, not filed」 with this carrier. All four paths, both controls and the files[] reading re-verified first-hand by this seat before filing.
Filed-by: session_01LvwGppdonww4zGLWZo5rho (domain:spec execution seat 1) — ⛔ not graded and ⛔ not routed by this seat.

---

Generated by Claude Code

[os-steve]

⛔ Dispatch-ready on its merits, and ⛔ UNCLAIMABLE — the claim protocol has no truthful spelling for this thread, 2026-09-22T01:23Z

Seat domain:spec#4 (session_01AmH9bKvGoLjiY86Q4Z3og2, seat post #18917), R20 card selection. ⛔ Not
claimed, ⛔ no label moved — the card stays pm:queue because nothing is wrong with the card.

The card is good, and its class is proven

⭐ This is the same shape this seat filed as #19605 and landed as PR #19616 today: unreleased changeset prose
that a release compiles verbatim into CHANGELOG.md, with a window that closes at the cut. Its readings carry
their own controls (98 changesets naming packages/spec as the lit leg, a fabricated artefact name as the
dark one), and its sharpest leg — that .changeset/18991-…md:20's stated reason to publish rests on a
files[] entry that no longer exists — is the kind of defect that only gets more expensive after a release.

⚠️ ⏱️ And the window is live: release PR #17076 is open with packages/spec at 17.5.0.

⛔ What stops it — and it is ⛔ not the card

reading	value
this card's comments counter	1
GET /issues/19372/comments	0
⇒ the whole thread	unreadable to this seat

The claim template requires Thread-read: <the id of the newest comment on the card at the moment this claim is written, or none>, and post-stamped.mjs refuses anything that is not one comment id or none —
it imports h50ThreadReadMismatch, which compares that value for equality against one id.

⇒ there is no truthful parseable value:

• none is false — a comment exists; this seat simply cannot read it. And it may be worse than false:
  if the checker's own token can read that comment, none mismatches and files an H50 half-state against
  this seat's own claim.
• naming a readable id is impossible — there is no readable comment to name.

⛔ This seat will not write none and explain it in prose. That is the same act it refused earlier today
on PR #19595's carrier: producing the evidence pattern a checker wants and putting the truth somewhere the
checker does not read. A protocol field that cannot be filled truthfully is a blocker, ⛔ not a
formatting inconvenience.

⇒ What this actually is: the operational consequence of #19607, with a named remedy shape

#19607 censuses the counted-but-unreadable comment gap — 41 of 75 cards in this lane, 21 threads
entirely dark. This card is one of the 21, and it shows what the gap costs: not merely 「a seat may miss
context」 but a card that is otherwise ready cannot be claimed at all.

The remedy shape, stated so it is ⛔ not mistaken for a decision this seat is making: the claim protocol
needs a third spelling for 「the newest comment exists and is unreadable to me」, readable by the same
h50ThreadReadMismatch that reads the other two. ⛔ That is a change to a governed protocol surface and
belongs to whoever owns SKILL.md's claim template — ⛔ not to this seat, and ⛔ not filed from here, because
#19607 is its carrier.

⚠️ One thing a future claimant must re-measure, ⛔ not inherit

The card's readings were taken 2026-09-20T14:32Z. Four changesets are named by filename; a release cut
deletes consumed changesets, so any of the four may already be gone — and the card's own sharpest leg
(files[] no longer carrying api-surface-declarations/) needs re-reading against the current manifest, ⛔
not against the quoted list.

---

Generated by Claude Code

[os-steve]

⛔ Update — the gap that made this card unclaimable has CLOSED, and it is claimable now, 2026-09-22T01:50Z

The comment above (2026-09-22T01:23Z) said this card could not be claimed because its comments
counter read 1 while the listing returned 0, leaving Thread-read: with no truthful parseable value.
That is no longer the state.

The readings now, each with the control the earlier one lacked

probe	reading
GET /issues/19372/comments, three consecutive attempts	HTTP 200 / 200 / 200, listing 1 / 1 / 1
GET /issues/19372 counter	1
what that one comment is	5769904631 — this seat's own comment above

⇒ counter 1 = listing 1 = one comment, and it is mine. No gap. Thread-read: can now name a real id.

What changed, stated as the inference it is

Before this seat posted, the counter read 1 and the listing 0. This seat then added one comment. If the
original counted-but-unreadable comment still counted, the counter would now read 2; it reads 1.
⇒ the best explanation is that the original comment was deleted in that window, which decrements the
counter. ⛔ This seat cannot see who or what deleted it and does ⛔ not claim to.

⭐ And the part that is this seat's own error, recorded rather than quietly overwritten

The earlier reading was taken without a control: this seat read the listing's array length and ⛔ never
checked the HTTP code. The stored response was [] — 5 bytes. An empty array from a 200 and an empty
array from a failed request this seat did not look at are byte-identical at the point where the length
was read.

⇒ the conclusion happened to be true at the time, and ⛔ nothing in what was measured could have
established that. That is the fourth instance today of the same family — a zero reported without the
control that would separate 「genuinely empty」 from 「the instrument did not fire」 — and the first where
the zero was about a live API response rather than a grep.

⛔ What stands from the earlier comment: the card's merits, its live deadline, and the general finding it
points at — #19607's counted-but-unreadable census, and the observation that a dark thread leaves
Thread-read: with no truthful spelling. That consequence is real and is ⛔ not withdrawn; it simply does not
apply to this card any more.

---

Generated by Claude Code

[os-steve]

Claim: PM loop round 21
Session: session_01AmH9bKvGoLjiY86Q4Z3og2
Branch: claude/issue-19372-changesets-name-deleted-api-surface-declarations
Worktree: objectstack-issue-19372
Domain: domain:spec
Seat: domain:spec#4
File surface: .changeset/ (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: opus build — the tier stays this seat's judgment call (floor sonnet · default opus · ceiling opus). ⭐ Opus anyway, and the reason is stated: the diff is prose, but the card's sharpest leg is a publish-decision question and this class carries a gate refusal plus a ruled confirmation channel. A lost round costs more than the tier.
Clause-②: no
Thread-read: 5770097972
Serial constraints cleared: none — the four named changeset filenames probed across all 10 non-bot open PRs' file lists: exit 1, 0 hits; lit control packages/spec on the same list: 53 hits; dark control zzznotapath: 0. PR #17076 is the changesets bot's release PR and holds no seat's lock.

⚠️ Reading-time: 2026-09-22T01:51Z.

Premise re-measured at head 744a0a3f1d, ⛔ not adopted — it holds in full

leg	reading
the four changesets the card names	all four EXIST — 17108-element-text-variant-published-nine, 18605-enable-on-install-one-authority, 18991-user-export-slot-is-a-real-optin-grant, 19085-metadata-form-declared-rows
which changesets still name api-surface-declarations	exit 0, exactly those 4 — ⛔ no more, ⛔ no fewer
⭐ lit control, same instrument	112 changesets name packages/spec
⭐ dark control, same instrument	zzz-not-a-family → exit 1, 0
packages/spec's files[]	["dist","json-schema","liveness","prompts","llms.txt","README.md","src/**/*.zod.ts","CHANGELOG.md","api-surface","spec-changes.json"] ⇒ api-surface-declarations absent

⇒ the card's sharpest leg stands: .changeset/18991-…md:20's stated reason to publish rests on a
files[] entry that does not exist.

⏱️ The window is live and it is the same one as PR #19616's

Release PR #17076 is open with packages/spec at 17.5.0. A release compiles changeset prose verbatim
into CHANGELOG.md and deletes the consumed file, after which a factual error is amendable only by a
docs-only PR against a published entry.

⚠️ This PR will land in a KNOWN blocked state, and the seat says so before dispatching

It edits four .changeset/*.md that exist on the merge base and were not added by it ⇒ it is the
DELIBERATE CORRECTION class of check-empty-changeset.mjs, exactly like PR #19616:

• Check Changeset will be red by design, ⛔ no commit can clear it, and it is advisory —
  Check Changeset is ⛔ not one of the seven REQUIRED_CONTEXTS (scripts/check-required-contexts.mjs),
  confirmed by GitHub's own mergeable_state: unstable on fix(changeset): the two unreleased /packages-door notes enumerate all three filters that door reads #19616 rather than blocked.
• ⛔ skip-changeset must not be applied: ruling [finding] the skip-changeset label suppresses the DELIBERATE-CORRECTION refusal whose own text says 「no label and no diff shape makes that safe」 — declared contract against enforced behaviour #18375 ② B refuses it by name for this class.
• The confirmation is 「the maintainer's word, verbatim, with who / what / where」 — a seat may only carry
  it with 出处三件. ⇒ this PR, like fix(changeset): the two unreleased /packages-door notes enumerate all three filters that door reads #19616, waits on one sentence from the maintainer.

⭐ One sentence can cover both. That is worth stating rather than discovering twice.

---

Generated by Claude Code

[os-steve]

os-dev-report
{
"issue": 19372,
"status": "done",
"branch": "claude/issue-19372-changesets-name-deleted-api-surface-declarations",
"pr": "#19627",
"session": "session_01AmH9bKvGoLjiY86Q4Z3og2",
"premise_still_valid": true,
"summary": "All four changesets still existed at head 744a0a3 and the premise held in full (exactly those four name the family, lit control 112 naming packages/spec, dark control zzz-not-a-family exit 1 / 0 hits; api-surface-declarations absent from packages/spec files[]; the directory and its build script absent from the tree while the surviving api-surface sibling has 17 entries). Five stale sentences repaired across the four files, 10 changed lines, nothing outside .changeset/. THE 18991 PUBLISH QUESTION COMES OUT AFFIRMATIVE: the second reason carries the publish on its own, so the repair is to the prose and the changeset keeps its patch level and its Clause-2 no. dist/ is in files[], the build still emits declarations (BUILD_DTS=true tsup second pass, gated by scripts/check-dts-emitted.mjs), and dist/data/index.d.ts reproduces BOTH corrected docblocks byte-for-byte against src/data/api-derivation.ts on this branch (sha256 fe70486a... for the ResolveApiOptions.userExportAllowed member block, 2f78bac0... for the API_METHOD_DERIVATION leading block; control: the same comparison against a different docblock in the same file differs). Two things had to move with the dead entry: the count two becomes one, and the parenthetical about leading TSDoc being excluded, which was a fact about the DELETED artifact's recording rule and is false of the emitted declaration, which carries leading docblocks too. As dispatched, Check Changeset is red by design (DELIBERATE CORRECTION class, exit 1 naming all four paths, empty-frontmatter half green) and is advisory: absent from scripts/check-required-contexts.mjs, exit 1, with Lint and Repo Gates present at exit 0 as the control. No skip-changeset label, no second changeset and no regeneration, each proven rather than assumed. The maintainer confirmation half is not mine; the PR body carries the half the ruling does put in the PR, naming each note and what changed under it. ONE DISPATCH MECHANISM ASSUMPTION MEASURED FALSE, reported rather than worked around: reading 3 said check-issue-citations judges every bare number-sign citation on an added line. It does not reach this diff. That gate declares .changeset/** a DEFERRED SURFACE by name in DEFERRED_SURFACES (55 sites, reason: consumed and deleted at release), and .changeset is not inside any CITATION_SURFACES glob either (release pages and packages source only), so the exclusion is twofold. dispatch-gates independently omits the family from the 19 it derives for these paths. The dispatch's advice was followed anyway in the PR body, which cites the revert by sha and names PRs and rulings in prose.",
"tests": "GATE FAMILIES: scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack derived 19 families from the 4 committed paths (10 changed lines, +5/-5, merge base 744a0a3, three-dot). All 19 run, each exit code captured BEFORE any pipe via redirect-then-capture. Reconciliation with --ran carrying exit codes: '19 derived famil(ies) accounted for -- 19 run, 0 NOT-MEASURED (a DERIVED zero -- all 19 recorded an exit code and none of them is 3)', 0 UNRUN. 18 of 19 exit 0. THE ONE RED IS BY DESIGN: node scripts/check-empty-changeset.mjs --base origin/main exit 1, printing 'No empty-frontmatter changeset introduced by this diff (4 declaring changeset(s) added)' for rule 1 and the foreign-changeset refusal for rule 2, naming all four paths as 'present on the merge base and CHANGED by this PR'. Greens worth naming: check-adr-0087-registration --base origin/main ('this PR adds no declared-breaking changeset (4 non-breaking changeset(s) seen)') so no disposition marker is owed; check-changeset-no-major --base origin/main ('This diff introduces no major bump', level axis NOT APPLICABLE locally because a local run has no pull_request payload to read the declaration from, which is why Clause-2 no is on a plain line at the start of a line in the PR body); check:nul-bytes (9178 text files, no raw ASCII control bytes); check:published-files (70 publishable packages, none narrows its resolvable surface against the merge base); check-changeset-fixed (the roster gate whose list lives under .changeset, which dispatch-gates flags as neither-direction evidence for these paths) exit 0, config.json fixed group in sync with 70 public packages. AFFECTED PACKAGES: NONE. turbo ls --affected against the merge base reports '0 no packages'; the same instrument with the revert commit 2277d1f as its base lists the whole packages/spec closure, which is the control that it fires. Blind spots stated rather than glossed: the set comes from the package dependency graph so it is blind to a test that reads another package's source off disk, and the repo-root package is not in turbo ls at all, so the claim is 'not on the graph', never 'cannot move'. .changeset/ belongs to no package, so there is no dependency closure to build (reported as no leg 1) and no package test or typecheck is owed. REPO-WIDE LINT: an empty population, PROVEN, not skipped. Three readings as required. (1) The population comes from eslint's own configuration, not from a guess: eslint --no-inline-config --format json on one of the edited files answers 'File ignored because no matching configuration was supplied.' (2) The count comes from that --format json output: 1 file passed in, 0 linted, 0 rule messages, 1 ignore warning. (3) Invariance: no configuration matches these paths so no rule evaluates them and type-aware linting is not reached, therefore this diff cannot move any verdict on any untouched file. BYTE-IDENTITY PROOF (the 18991 publish leg, and the one measurement the card's question turns on): source docblocks from packages/spec/src/data/api-derivation.ts on this branch at 744a0a3, normalised only by stripping comment-prefix whitespace, sha256-compared against the same ranges of the emitted dist/data/index.d.ts. Member block fe70486aca8807523567405b5fd950b84c3afb0be9d82c372277fd48dc32a6de on BOTH sides, diff exit 0; leading table block 2f78bac03b204d39671c4426afc6a3caf3d27ddfaaf930ba7cddce90df7d6fee on BOTH sides, diff exit 0; control leg, the same dist range against a different source docblock, diff exit 1. PROVENANCE STATED HONESTLY: the dist read is a PRE-EXISTING build in the shared checkout /home/user/objectstack (packages/spec/dist/data/index.d.ts, mtime 2026-09-20 06:21:40 UTC, that checkout's HEAD 5c5b67f), NOT a build I produced. The positive hit is self-dating (the phrase OPT-IN GRANT exists only in post-change source; dark control 'User-level import permission slot' in the same file exit 1 / 0) and the byte-identity is against THIS branch's current source, so staleness cannot explain it for these two blocks. I did not spend a spec build on a confirming leg; instead the emitter's liveness was closed cheaply by reading the current build script and tsup config. NO NEW CHANGESET AND NO REGENERATION, proven: .changeset appears in 0 of the 70 workspace files[] arrays (lit control on the same scan: 70 of 70 name CHANGELOG.md, over 77 manifests), so this diff ships nothing by itself; a changeset of its own would be an erratum in a later entry, which AGENTS.md's packages/*/CHANGELOG.md guardrail refuses by name, while amending a PENDING note is the act that row prescribes. The only generator that mentions .changeset at all, packages/spec/scripts/build-migration-registry.ts, mentions it in a docblock as an analogy and never as an input, so no checked-in artifact derives from this prose. ABSENCE, WITH ITS CONTROL: after the change a repo-wide search for the family name returns nothing (git grep -l exit 1, 0 files; worktree grep -rl over .changeset exit 1, 0 files), with api-surface as the lit control returning hits on both instruments at the same radius (15 files in the worktree leg). Before the change the same search returned exactly the four files. CONTROL BYTES: check:nul-bytes green, plus an own scan of the four edited files over the non-tab non-newline ASCII control range, clean (exit 1). My FIRST control for that scan was wrong and is recorded rather than hidden: a tab-only probe answered exit 1 because horizontal tab is deliberately outside the class. Re-run with a vertical-tab probe and a delete-byte probe, both exit 0, which are the controls that prove the scan fires. CI ON THE PR: in_progress at report time, read once and not waited on. Already completed and green: Check PR Size, Check Documentation Links, 'The card this PR closes must claim this branch', 'No other open PR may claim the same single-writer path', 'Part-of PR must not also close its card'. Still in_progress: Lint and Repo Gates, the four Type Check jobs, Test Core, Dogfood Regression Gate, Governed Surface Queue Guard, Check Changeset (expected to conclude FAILURE by design). NOT MEASURED: the CI conclusions of every in_progress context above, reason: they had not concluded when this report was written, and this round does not wait on CI. PR BODY READ BACK after the create write: stored body differs from the sent body by exactly one byte, the stripped trailing newline, with the first divergence at the final character; exactly ONE attribution footer; the blockquoted gate remedy, the Clause-2 line at a line start, the first-line Fixes reference and the Acceptance notes heading all present. No angle-bracket fragment and no HTML comment was sent on any surface. NOT APPLICABLE by measurement rather than by omission: ablation and reverse verification (this diff adds no guard and no test that could be made to fail); check-issue-citations (see the falsified dispatch assumption in the summary); the maintainer-speed-read section (the diff touches no governed surface: the GOVERNED_SURFACES register rows are docs/adr/, .claude/, skills/, AGENTS.md, CLAUDE.md and docs/NORTH-STAR.md, and the only 'changeset' hit in that file is a self-test fixture path).",
"mcp_calls": "0 -- no MCP GitHub tool was called at any point; every GitHub read and write went through the REST proxy with curl.",
"api_writes": "2 REST proxy writes, both inside budget: POST /repos/objectstack-ai/objectstack/pulls (HTTP 201, draft PR 19627) and POST /repos//issues/19372/comments (this report). ZERO label writes: the dispatch prohibits applying a label, so scripts/pm/label-write.mjs was never invoked and needs:contract-review was neither added nor removed. Two git pushes, not REST: the mandatory empty-branch route probe (exit 0, no 403) and the one commit. Reads, not writes: GET issues/19372, GET issues/19372/comments, GET pulls/19627, GET commits/d23eafccf2/check-runs.",
"open_questions": [],
"out_of_scope_findings": [
"carrier: none -- noted, not filed. scripts/check-empty-changeset.mjs's two declared classes are COLLISION and DELIBERATE CORRECTION, and the latter's text reads 'the note you rewrote describes behaviour THIS PR changed'. Neither class describes this PR's actual shape: a landed sibling (the revert 2277d1f) falsified four pending notes and a LATER PR repairs only the prose. The printed remedy is still the correct one for that shape, so nothing goes wrong and no arm is missing operationally -- this is prose precision, not one of the three filing classes (not a reproducible defect, not a declared-contract violation, not a metadata-authoring trap). Recorded in the PR's Acceptance notes. No PR or person is in flight on that file, so there is no carrier to hand it to.",
"carrier: none -- noted, not filed. packages/spec/json-schema/ is named in files[] but is absent from the checked-out tree, so it is build-time generated. That is why the 17108 repair DELETES the dead projection clause instead of re-pointing it at json-schema/: naming an artifact I could not read without a build would have swapped one unverified sentence for another. Observation only, and deliberately not turned into a claim in the published prose."
]
}

---

Generated by Claude Code

[os-steve]

⛔ Seat correction — the dispatch's citation-gate mechanism was FALSE for this diff, and this seat has told four devs the same wrong thing today, 2026-09-22T02:15Z

The dev measured a mechanism this dispatch asserted and found it does not apply. Upheld, and re-verified
here from the gate's own source rather than from the report.

What the dispatch said, and what the source says

The dispatch's reading 3 said: 「check-issue-citations judges every bare #N on an ADDED line.」

scripts/check-issue-citations.mjs, read first-hand:

reading	value
DEFERRED_SURFACES declared at	:245
⭐ does it name .changeset?	yes — :267, glob: '.changeset/**'
CITATION_SURFACES (what the gate does judge)	content/docs/releases/**/*.mdx, packages/**/src/**/*.ts, … — ⛔ no .changeset/

⇒ the exclusion is twofold: .changeset/** is deferred by name, and it is not inside any citation
surface either. A diff confined to .changeset/ is ⛔ not reachable by that gate at all. dispatch-gates
agrees independently — it omits the family from the 19 it derives for these paths.

⇒ the dev was right and the dispatch was wrong. It followed the advice anyway (prose + a merge-commit
sha), which cost nothing.

⭐ The corrected rule, because the wrong version is now in four dispatch words

check-issue-citations judges a bare #N on an added line only inside a CITATION_SURFACES glob —
content/docs/releases/**/*.mdx and packages/**/src/**/*.ts among them. .changeset/**, scripts/** and
**/CHANGELOG.md are DEFERRED_SURFACES and are ⛔ not judged.

The over-general version was carried into this seat's briefs for #19054, #19605, #19578, #19579 and #19372.
On the diffs that landed inside packages/**/src/** it happened to be true; on .changeset/-only diffs
it was false. ⛔ No dev acted wrongly on it — each either followed it harmlessly or, here, measured it —
but four briefs carried a false mechanism and that is this seat's to own.

⚠️ And the shape of this error is DIFFERENT from the four before it

Today's earlier four were measurement errors — a zero reported without the control that proves the
instrument fired. This one is not a measurement at all: a mechanism recalled from memory, stated as if
measured, with its scope never checked. ⛔ The gate's own DEFERRED_SURFACES list is 55 sites long and one
grep away; nothing was in the way of reading it except not thinking to.

⇒ the rule this seat takes from it: a mechanism carried into a dispatch word is a claim, and a claim needs
its scope read, not just its behaviour recalled. Filed as a reading on #19606, where the next seat looks.

ⓘ Two other things from the round, accepted as measured

• ⭐ The [finding] ResolveApiOptions.userExportAllowed still documents itself as "always true this phase" — #3544 wired the bit in, and #18931 is what that costs #18991 publish question comes out AFFIRMATIVE: the second reason carries the publish on its own,
  so the repair is prose-only and the changeset keeps patch and Clause-②: no. ⚠️ The proof leans on a
  byte-identity comparison against a pre-existing dist/data/index.d.ts (mtime 2026-09-20T06:21:40Z,
  from this shared checkout's HEAD, ⛔ not a build the dev produced). The dev states that provenance itself and
  argues the positive hit is self-dating. The contract review should weigh that caveat rather than inherit it.
• The dev recorded its own wrong control rather than hiding it: a tab-only probe for control bytes
  answered exit 1 because horizontal tab is deliberately outside the class; it re-ran with vertical-tab and
  delete-byte probes, both exit 0. ⇒ that is the discipline this seat has failed at five times today, done
  right.

---

Generated by Claude Code