fix(changeset): date the pending client entry's closing register — all three 'not addressed here' clauses had been answered (#18765)

Fixes #18652

`.changeset/client-get-session-envelope-and-refresh-read.md` is
**pending release
input**, not a note: `changeset version` copies it verbatim into
`packages/client/CHANGELOG.md`, which is in `@objectstack/client`'s
published
`files[]` and ships in the npm tarball. AGENTS.md's release-owned table
states the
deadline in its own words — *"Your PR's input is its **changeset**, on a
hard,
unwatched deadline: the release that consumes it deletes that input and
publishes
the sentence."*

## ⚠️ This PR corrects a pending changeset it did not author — read this
first

`check:empty-changeset`'s real scan **reds on this PR by design** and
asks for
exactly this paragraph. Its own text names the two classes and their
opposite
remedies; this is the **DELIBERATE CORRECTION** class, and the gate says
*"this
gate stays red either way, and staying red is what puts the decision in
front of a
person instead of routing around it."*

- **The note:**
`.changeset/client-get-session-envelope-and-refresh-read.md`,
introduced by `5de93728a` (PR #17237, **merged** 2026-09-09T21:43:50Z —
confirmed
  at the tree, so there is no open author to defer to).
- **What changed under it:** its closing paragraph was an **undated,
present-tense**
register of what that change left undone, and all three of its clauses
had since
been falsified on `origin/main` by other cards' landings (readings
below).
- **⛔ Not the collision class.** No changeset filename was drawn or
overwritten
here; this PR adds none. ⛔ Do not restore the file from the merge base —
that
  republishes three false sentences.
- ⛔ **I have deliberately NOT applied `skip-changeset`.** It would
exempt the
`changeset-check` job **wholesale** (`lint.yml`'s own note on the
self-test split
says so), which is the one label that would silence this refusal.
Suppressing it
is the routing-around the gate forbids, so the label decision is left to
a human.

## Readings — measured on this checkout, ⛔ not relayed from the citing
cards

Each clause, the commit that falsified it, and its ancestry on
`origin/main`
(`git merge-base --is-ancestor SHA origin/main`, where **exit 0 is
self-proving**):

| clause, as it read | measured now | falsified by | ancestor? |
|:--|:--|:--|:--|
| the anonymous `null` "would need the published return annotation to
widen" | the **producer** moved instead; the annotation is untouched |
`374d9d3afa` (#17881), 2026-09-12 | exit 0 |
| `SessionUser.image` "declared `z.string().optional()`" | `image:
z.string().nullish().describe('Avatar URL')` —
`packages/spec/src/api/auth.zod.ts` | `0e51278f3` (#18501, for #17235),
2026-09-16 | exit 0 |
| `auth.login` / `auth.register` "normalize into `data` but set no
`success`" | both run `normalizeSessionResponse`, which returns `{
success: true, ...body, data }` | `01388fe81` (#17791, for #17234),
2026-09-12 | exit 0 |

**Clause 1, at the definition rather than a call site.**
`packages/plugins/plugin-auth/src/anonymous-session-refusal.ts:124`
`refuseAnonymousSession` keys on the **answer shape** and on nothing
else — three
early returns: `isGetSessionPath(endpointPath)`, `response.status !==
200`, and
`body.trim() !== ANONYMOUS_BODY` where `ANONYMOUS_BODY = 'null'`
(compared as
**text**, so `'0'` / `'""'` / `'false'` are left alone). It never reads
how the
caller became anonymous, so *never signed in*, *unknown cookie* and
*revoked
session* all convert alike. `ANONYMOUS_SESSION_REFUSAL_STATUS = 401`,
and it is
wired at `auth-manager.ts:5679`, the one seam every vendor route passes
through.
⇒ the anonymous answer is not a `null` outside the declared type; it is
a
**rejection**, which a `Promise` of `SessionResponse` annotation already
permits.

## What the diff does, and the one design choice in it

Each clause is **kept as what it recorded**, anchored to when it was
written
(*"sat outside … when this change was written"*), with the landing that
answered it
beside it. Two reasons that shape rather than a straight fact-swap:

1. A fact-swap re-arms the same defect — the census's own header names
the cause:
*"prose does not re-measure itself"*. A dated clause cannot be falsified
again.
2. It is the same rule the card's own fence applies one paragraph up:
the
better-auth **1.7.2** transcript is correct **because** it is dated and
   attributed. ⛔ That transcript is untouched — verified by needle
(`better-auth 1.7.2` and `(anonymous) -> 200 null` both still present at
lines
   10 and 14).

## Scope note — clauses 2 and 3 were outside the card's fence

Card #18652 fenced scope to clause 1 and marked clause 2 **not
re-measured, in
either direction**; clause 3 it did not mention. I measured both and
they are false
too, so fixing clause 1 alone was not available: the sentence enumerates
(*"**Two** answers stay outside the declared type"*), so a clause-1-only
repair
would have had to **newly author** the surviving false clauses into
release input —
strictly worse than what was there. Same defect class, same sentence,
same file, no
new verification surface. ⇒ folded in, declared here, and reported
separately to
the dispatching seat. Strip the last two rows if the seat disagrees; the
diff is
one paragraph.

## Tests / gates — the pin question, answered with three readings

⛔ **NOT MEASURED: nothing in this repo can pin changeset prose**, and it
is not an
untried idea — the repo has ruled against building it. Three
measurements, none of
them my own instrument:

1. **Step 49, `check:pm-changeset-deadline-census`** — its header:
*"REPORT-ONLY:
it fails nothing and gates nothing."* It measures **path presence**,
never
prose; its own blind-spot list says *"`window-open` means the FILE
exists, never
that the card's sentence about it is still correct."* Live run on this
checkout:
   `#18652 → verdict "window-open", assertsPending "pending changeset",
inListing true`, tally `{window-open: 11, consumed: 3}`,
`falsifiedAssertions: []`.
The row is **identical before and after this diff** (I amend, not
delete).
   ⇒ reachable, but structurally unable to fail. Not a pin.
2. **Step 122, `check:changeset-gate-self-tests`** — `lint.yml`'s own
note:
*"The SELF-TEST halves only — the real scans stay in pr-automation.yml's
   `changeset-check`."* `dispatch-gates` scores self-test-only families
*"checker-health only … NOT a PR verdict."* Green here and green without
this
   diff. Not a pin.
3. **Building one is forbidden right now.**
`changeset-deadline-census.mjs`:
*"⛔ This file is deliberately NOT the enforcement half … report-only
first, the
census is the deliverable, expansion only when the census reads zero
including
its blind spot."* The census reads **3 exposed**, not zero.
Independently,
`.changeset/**` is a ruled scan exemption — *"a changeset is that record
before
   it is compiled into a CHANGELOG"*
   (`packages/objectql/src/action-owner-key-single-source.test.ts`,
`NOT_A_STALE_MENTION`). And the file is consumed and deleted at the next
release, so a pin reading that path becomes a phantom check by
construction.

⭐ The one reachable instrument that **does** respond to this diff is a
third the
roster reading did not name — `check-empty-changeset.mjs`'s real scan —
and its
polarity is **inverted**: it is green without this change and red with
it, on the
**act**, not the prose. That is the human-confirmation gate above, not a
pin.

**Derived sweep** — `node scripts/pm/dispatch-gates.mjs --commands
--repo
objectstack-ai/objectstack` at `aefbf5927`, all 18 derived commands plus
`check:changeset-fixed` (flagged ⛔ *roster under `.changeset`*) run
locally:

```
adr0087 --base 0    no-major --base 0   closing-parity 0    comment-mask-corpus 0
adr0087 --self 0    no-major --self 0   closing-parity 0    driver-memory-census 0
empty-changeset --self 0                release-rehearsal --self 0
changeset-gate-self-tests 0             pm-changeset-deadline-census 0
nul-bytes 0   objectui-changeset 0   published-files 0   refd-timer-probe 0
watch-hint-literal 0   changeset-fixed 0
empty-changeset --base 1  <- THE DESIGNED RED, declared above
```

- **NOT MEASURED: `check:rerun-safety-verdict`.** `dispatch-gates`
flagged a stale
tree; the roster delta across it is exactly this one new whole-tree
family
(`95b21b33b`, #18746). It does not exist in this checkout, so `pnpm`
exited 254
(script-not-found) — ⛔ that is not a red gate and not a pass. It is
self-test-only
and reads only its own fixtures, so it cannot judge this diff either
way.
- **No control characters**: `grep -naP
'[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]'` over the
  changed file exits 1 (clean), beyond `check:nul-bytes`.
- **No `pnpm lint` narrowing claimed, and no build/test**: this diff
compiles
nothing and is read by no test. `turbo`'s graph is not consulted because
the path
  is in no package.

## Does this diff owe a changeset of its own? — measured, ⛔ not assumed

**No, and adding one would be a defect.** `@objectstack/client`'s
published
`files[]` is `["dist","README.md","CHANGELOG.md"]`; `.changeset/**` is
not in it, and
this diff moves no `dist` byte. The published text that *does* move is
this very
entry — so the amended changeset **is** the release input. A second
changeset would
emit a second CHANGELOG bullet correcting the first, which is the shape
AGENTS.md
forbids by name: *"Factual error in a released entry → amend that entry
in a
dedicated docs-only PR, ⛔ never an erratum in a later entry."* Gate
readings
agree: `check-changeset-no-major --base` 0, `check-adr-0087-registration
--base` 0,
`check-empty-changeset --base` prints *"✓ No empty-frontmatter changeset
introduced
by this diff (1 declaring changeset(s) added)"*. ⛔ `major` does not
exist in this
launch window and nothing here is breaking.

Clause-②: no

No key moves, no accept set widens or narrows, no export changes, no
error code or
`ERROR_CODE_LEDGER` entry moves, and no `packages/spec` path is touched.
The diff is
prose inside one pending changeset; zero lines of shipped code change.

## ⛔ Reported, not touched: the open Version Packages PR

**#17076 `chore: version packages` is OPEN and already renders this
entry** — at its
head `1c0ce1713` the changeset is **absent** (contents API `404`,
against a `200`
positive control on `.changeset/config.json`) and the paragraph is
compiled into
`packages/client/CHANGELOG.md:506`. ⛔ I did not touch that PR, ran no
release, and
merged nothing.

⭐ **But an open Version Packages PR does not mean the window has closed,
and the
dispatch order's reading that it does is falsified by the repo's own
workflow.**
That branch is a **derived artefact, regenerated from scratch**:

- `release.yml`'s `version-pr` job is `if: github.event_name ==
'schedule' || (… workflow_dispatch && inputs.refresh_version_pr)` — ⛔
**not** `push`; the file's
own note: *"this job regenerates the PR from scratch, so the newest
run's result
is the one that was wanted anyway"*, and *"renders changesets that are
already
  committed on main, so lateness costs nothing."*
- The mechanism, quoted in that file from changesets/action v1: `git
reset --hard SHA` → `pnpm run version` → `git push … --force`.
- Corroborated at the object: #17076 was created 2026-09-09 and carries
**one**
commit, `1c0ce1713`, authored by `github-actions[bot]` at
2026-09-17T18:14:46Z,
whose parent `e77a23f02` is an **ancestor of `origin/main`** (exit 0)
and only 7
commits behind it. Created eight days before the commit it holds ⇒
force-rebuilt.

⇒ The window closes when a **release consumes** the changeset
(AGENTS.md's wording),
i.e. when #17076 is merged and published — a human-only act
(`release.yml`:
*"TWO LANES, ONE INVARIANT: ONLY A HUMAN PUBLISHES."*). Until then this
entry is
still amendable at one paragraph, and the next 6-hourly tick re-renders
#17076 from
the corrected text with no action on that PR. The deadline is real and
this PR is
inside it.

---
🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01DvvamiacK328idtBYJBxV3

---
_Generated by [Claude
Code](https://claude.ai/code/session_01DvvamiacK328idtBYJBxV3)_
---

### Landing note added by the dispatching seat (#6024) — clearing a red
that was the seat's, not this PR's

The earlier red on `The card this PR closes must claim this branch` was
caused by the **claim comment's shape**, not by anything in this diff:
the seat's claim on #18652 opened `**CLAIM · …**`, and
`scripts/check-closing-target-claim.mjs` selects a comment by a line
beginning `Claim:` plus a **separate** `Branch:` directive line. Fixed
on the card (`5720097499`), measured before and after —
`check-clause2-carriers --pair 18765` went from `EXIT=4 /
claim.selected: none` to `EXIT=0 / 1 comment(s) in the pool`, both
carriers reading `no`.

That guard is body- and comment-scoped and re-fires only on a
`pull_request` or `merge_group` event, so **this edit is the event that
re-judges it** — ⛔ not an empty commit, which the deliverer correctly
refused, and ⛔ not a push, since the seat has ruled that all three
clauses stay and no further code change is owed.

⚠️ The two remaining reds are **by design** and are not touched by this
edit: `check-empty-changeset` asks in its own words for a human to
confirm the DELIBERATE CORRECTION class, and `pr-automation.yml`'s
`changeset-check` counts only **added** changesets (this PR amends one
and adds zero). The ruling request is on #18652 (`5720153848`). ⛔
`skip-changeset` is deliberately NOT applied — it exempts that job
wholesale and would silence the refusal that exists to put this decision
in front of a person.


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

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

Author: os-support-ai

.changeset/client-get-session-envelope-and-refresh-read.md

@@ -57,8 +57,21 @@ inert as it was.
 `token.signature` form interchangeably, so a client that held the signed form
 stays signed in across the call.
 
-Two answers stay outside the declared type and are **not** addressed here: the
-anonymous `null`, which would need the published return annotation to widen, and
-`SessionUser.image`, declared `z.string().optional()` against a route that
-serves `null` (#17235). The sibling `auth.login` / `auth.register`, which
-normalize into `data` but set no `success`, are #17234.
+Three answers sat outside the declared type when this change was written and
+are **not** addressed by it. Each has since been answered on its own card, so a
+caller reading this entry does not have to code around any of them:
+
+- the **anonymous** `/get-session` answer, recorded above as `200 null`. It no
+  longer needs the published return annotation to widen, because the producer
+  moved instead: since #17881 `plugin-auth`'s `refuseAnonymousSession` converts
+  better-auth's `200` plus the literal JSON `null` into the declared ADR-0112
+  refusal — HTTP `401` with `code: UNAUTHENTICATED` — before it leaves the
+  process. The SDK's shared `fetch` wrapper throws on any non-2xx, so an
+  anonymous `auth.me()` **rejects** rather than resolving outside its own type.
+  Ruled by #17238: the producer moved and `SessionResponseSchema` is untouched.
+- `SessionUser.image`, then declared `z.string().optional()` against a route
+  that serves `null` (#17235). It is now declared `z.string().nullish()`, so
+  the `"image": null` every `/auth/*` session body carries parses.
+- the sibling `auth.login` / `auth.register`, which then normalized into `data`
+  but set no `success` (#17234). They now run this entry's own lift, which
+  fills `success` as well as `data`.