docs(spec): correct the published 17.4.0 entry that restated the retired M9.2 promise - #18569

…ed 17.0.0 and 17.0.0-rc.1 entries (objectstack-ai#18852)

Part of objectstack-ai#18740

> ⚠️ **Changed from `Fixes` to `Part of` by the dispatching
`domain:devx` seat.** Eight of the card's nine sites are corrected here.
The ninth — `content/docs/releases/v17/17-0.mdx:1734` — is still a live
carrier, and the delivering agent correctly declined it: its standing
operating rules prohibit editing `content/docs/releases/**`
unconditionally, and **a permission in AGENTS.md does not lift a
prohibition**. A closing keyword here would silently close a card that
still has an open half, so it is withdrawn. The ninth site is routed as
its own card; this PR stays a clean two-file CHANGELOG amendment.

A **dedicated docs-only PR**, which `AGENTS.md`'s Documentation
Guardrails row for `packages/*/CHANGELOG.md` requires: a factual error
in a released entry is amended *in that entry*, ⛔ never as an erratum in
a later entry and ⛔ never as a rider on code changes — "the reader greps
the tombstoned symbol and lands on the old entry, so a correction
anywhere else is one it never reaches".

The diff is exactly two files. No source, no schema, no export, no test,
and **no changeset** (see below).

## What was false

The 2026-09-07 ruling found that no reader of a published surface can
configure `batch.maxBatchSize`. The cap is **embedder policy**:
`RestServerConfig.batch.maxBatchSize` is the argument a host passes when
it constructs the server, through the one door `createRestApiPlugin({
api })`. Neither shipped boot path passes it — `os serve` forwards
exactly two keys out of the stack config's `api:` block
(`api.enableProjectScoping`, `api.projectResolution`), and the dev
plugin calls `createRestApiPlugin()` with no config at all. A
CLI-started deployment therefore always gets the 200 default, and no
flag, config file or CLI option moves it.

The `789ad63` entry carries that falsified claim in **two wordings**,
and it is duplicated under two published version headings, so there are
**eight** carriers, not four.

## The eight sites corrected

Located **by content, across lines** on this tree at `631dcbd4b` — ⛔ not
by the line numbers carried on the card, and ⛔ not with a single-line
match. The `Batch size is / deployment policy` sentence is **wrapped
across two lines**: a single-line grep for it returns 0 and reads as a
false absence.

| file | heading | wording |
|:--|:--|:--|
| `packages/rest/CHANGELOG.md` | `## 17.0.0` (L4733) | `should raise`
instruction + `deployment policy` |
| `packages/rest/CHANGELOG.md` | `## 17.0.0-rc.1` (L14992) | both |
| `packages/spec/CHANGELOG.md` | `## 17.0.0` (L15658) | both |
| `packages/spec/CHANGELOG.md` | `## 17.0.0-rc.1` (L67924) | both |

Per copy:

- **The claim.** `Batch size is deployment policy` now reads `Batch size
is embedder policy`, and the counterfactual beside it — `(a deployment
raising the limit to 500 would still have been refused at 200)` — now
says `a host`, since it presupposed the same unreachable knob.
- **The instruction.** The behaviour bullet read `Deployments that were
quietly relying on unbounded batches should raise` `batch.maxBatchSize`
`(up to 1000) rather than discover the cap in production`. That told an
operator to perform an action no shipped boot path can perform, so a
reader who complied had no way to tell whether they had succeeded. It
now states where the cap actually comes from and issues no instruction.

## Where this landed between "faithful to the record" and "no longer
misleading"

This is a CHANGELOG, so the job is to record what happened in that
version — ⛔ not to rewrite history into "this is what we said at the
time". I did not invent a shape for that: this repository has already
settled it, in objectstack-ai#18569 / objectstack-ai#17849, and I copied it. **The published words
are corrected in place, and the old words are kept as a marked quotation
in a dated erratum line closing the entry** — the in-repo tail already
used at `packages/spec/CHANGELOG.md` (three sites),
`packages/lint/CHANGELOG.md` and
`packages/metadata-protocol/CHANGELOG.md`. So the entry no longer
instructs, and what shipped is still readable verbatim.

The factual vocabulary is likewise copied, ⛔ not invented — from
`ec5db7b` (`packages/rest/CHANGELOG.md:351`) and the pending
`.changeset/18739-batch-cap-embedder-only.md`, which are the fourth and
fifth landings of this same correction.

⛔ No new entry at the top, ⛔ no version heading added, ⛔ nothing this
release published is changed: the 1..1000 range, the 200 default and the
enforcement are all untouched.

## ⛔ Deliberately no changeset

A changeset would compile this correction into a **new** release note —
which is precisely the erratum-in-a-later-entry shape `AGENTS.md`
forbids. Nothing published moves here either: the diff is prose inside
already-shipped entries.

## ⚠️ The ninth site is NOT in this PR — a standing-rule conflict I am
not resolving silently

`content/docs/releases/v17/17-0.mdx:1734` carries the same claim (`stay
under` `batch.maxBatchSize` `(default 200, raisable to 1000) or chunk`)
and the dispatch listed it as the ninth site.

I did not touch it. My standing operating rules carry an
**unconditional** prohibition on editing `content/docs/releases/`, and
they state that such a clause wins over the dispatch word. `AGENTS.md`
*permits* a docs-only PR there; it does not *require* one — a permission
does not override the prohibition, so there is no conflict with
`AGENTS.md`, only with the dispatch. Flagging rather than choosing a
side: that line is **still a live carrier** and needs either a separate
actor or an explicit release of the prohibition.

## ⚠️ Report item, ⛔ not fixed here

`packages/spec/CHANGELOG.md:3581` says the phrase "exists verbatim in
the REST server ... at `packages/rest/src/rest-server.ts:2071` ... it is
owed to a follow-up in `packages/rest`". Re-measured on `631dcbd4b`:
`grep -c "deployment policy" packages/rest/src/rest-server.ts` = **0**,
and `packages/rest/CHANGELOG.md:351` records `ec5db7b` retiring exactly
that phrase. **That follow-up is done and the note is now stale** — a
debt recorded as outstanding that has been paid. It is a different error
from the one this card names, so it is reported, ⛔ not ridden. `objectstack-ai#18740`
does not cover it.

## Firing controls

Computed from `git diff -U0` hunk headers so that context lines cannot
contaminate the reading, and taken from the probed files themselves.

- **C1** — `should raise` + `batch.maxBatchSize`: **0 outside the
diff**. Its only four hits are the verbatim quotations inside the new
erratum lines.
- **C2** — the batch-cap `deployment policy` claim, multiline-aware: 4
hits outside the diff, each classified and none a live carrier —
`packages/rest/CHANGELOG.md:351` (records the phrase's retirement),
`packages/spec/CHANGELOG.md:3581` (the stale note above),
`.changeset/18739-batch-cap-embedder-only.md` (the sibling correction),
`docs/qa/platform-checklist/FOLLOW-UPS.md:575` (states the correct
embedder-only fact).
- **C3** — the wrapped `Batch size is` / `deployment policy` sentence:
**0 hits anywhere in the tree**.
- Control bytes: `grep -naP` over both changed files exits 1 (no match).

A bare `deployment policy` string is ⛔ not a usable control — it matches
unrelated scheduling prose in about 30 files.

---
_Generated by [Claude
Code](https://claude.ai/code/session_017ef78bLdybu3AffehKkhfk)_

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

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