fix(plugin-form): report collapsible / collapsed beside group only when true - #26042

… different packages stop conflicting (#26120)

<!-- claude-projects-attribution: {"github_login":"os-elon-musk"} -->
_Requested by **Elon** · [project
thread](https://claude.ai/code/project/chan_01WedDzV26nniwxv7jC88B9D?thread=cmsg_01WedDzV26nniwxv7jC88B9DPXV532seCFDcrkAJybGzsf)_

Fixes #26051

Clause-②: no
Metadata-protocol: none

Part of the p1 merge-queue backlog work (#26074). Dispatched by seat
`domain:devx#1` (#6023), claim 6415558140, card-thread session
`session_01SMJGPXxFBpbgAgXdnsXKVS`.

## What changes

`scripts/sdui-manifest.record.json` carried ONE
`inTree.registrySourcesSha256` over the `src/` of all 27 ui registry
packages, so any two ui PRs that touched any registry package conflicted
on that line. Now:

- **One record per registry package**, one file each, one line each:
`scripts/sdui-manifest.sources/PACKAGE.json` =
`{"package","sources","sha256","manifestSha256"}` (the package's own
`src/` digest, its file count, and the sha256 of the artefact it was
verified against). 27 files.
- **The shared record** (`scripts/sdui-manifest.record.json`, same path)
keeps only what describes the artefact: `sha256`, `components`,
`generatedAt`, and `inTree` = `modulesRoot`, `objectuiWorkspaceVersion`,
`registryRecords` (the directory). The retired `registryPackages` /
`registrySources` / `registrySourcesSha256` are refused if they
reappear.
- **The reached package set** is the set of record files: a reached
package with no record (new package, new dependency edge into the
registry, deleted record) is RED; a record the registry no longer
reaches (removed package, dropped edge) is RED.
- `sdui.manifest.json` is byte-identical (sha256 `1f6e1ccb925a…`, 106
components).

`lint.yml`'s command line (`node scripts/check-sdui-manifest.mjs
--self-test && node scripts/check-sdui-manifest.mjs`) is unchanged and
keeps working. No `.github/**` or `ui/.github/**` edit.

## Measured first (Ask 1): own `src/` per package, not
dependency-inclusive

Tree: objectstack `0a6948210c` (base), 27 reached packages, 922 non-test
sources.

**Detection is the same as one digest over the set.** Any edit under any
reached package's `src/` moves exactly that package's digest, so a
single PR's edit is caught exactly as before; a dependency's edit is
caught through the dependency's own record. Dependency-inclusive digests
add no single-PR detection. They would only add conflicts, and at this
cost (records a `src/` edit would touch):

| package edited | own `src/` (this PR) | dependency-inclusive |
|:--|:-:|:-:|
| `types` | 1 | 25 |
| `core` | 1 | 21 |
| `data-objectstack`, `i18n` | 1 | 19 |
| `react`, `react-runtime`, `sdui-parser` | 1 | 18 |
| `components` | 1 | 17 |
| `fields` | 1 | 14 |
| `plugin-detail` | 1 | 10 |
| `plugin-form`, `plugin-grid` | 1 | 2 |
| the other 11 plugins | 1 | 1 |

**Merge experiment, real tree, `git merge-tree --write-tree` (no merge
driver routes these paths; `git check-attr merge` = unspecified, so this
is GitHub's answer):** each branch appends a comment to one `src/` file
and runs the generator.

| pair | before (single digest, on `0a6948210c`) | after (this PR, on
`a8114011`) |
|:--|:--|:--|
| `plugin-grid/src/FormulaBar.tsx` × `plugin-form/src/DrawerForm.tsx` |
exit 1, CONFLICT in `scripts/sdui-manifest.record.json` | exit 0, clean;
the merged tree passes `check-sdui-manifest` (exit 0) |
| `plugin-grid/src/FormulaBar.tsx` × `plugin-grid/src/ObjectGrid.tsx` |
(conflicts) | exit 1, CONFLICT in
`scripts/sdui-manifest.sources/plugin-grid.json` only, by design |

**The open ui PRs today** (registry packages each changes, non-test
`src/` only; none changes `sdui.manifest.json`): #25987 i18n,
plugin-grid, plugin-list · #25998 plugin-detail · #26040 components,
i18n · #26045 i18n · #26027 plugin-dashboard · #26030 plugin-grid ·
#26033 core, plugin-form, plugin-grid · #26042 plugin-form · #26044 i18n
· #26048 i18n, plugin-map · #26050 fields, plugin-grid · #26062 fields.
Under the single digest all 12 conflict pairwise (12 serial landings).
Under this layout they conflict only within a shared package; the
largest group is `i18n` (5 PRs), so they can land in 5 rounds, e.g.
{#25987, #25998, #26027, #26042, #26062} → {#26040, #26030} → {#26045,
#26033} → {#26044, #26050} → {#26048}, assuming no source-level
conflicts between them.

## The check (Ask 2)

`checkTreeWithoutPin` compares each reached package's digest with its
own record and names every stale package (`ui/packages/plugin-grid: now
38 file(s) sha256 …, recorded …`), every unrecorded package, every
record the registry no longer reaches, every record verified against
another artefact, and every malformed record, then prints the one
command: `node scripts/gen-sdui-manifest-node.mjs`. The pin mode
(`.objectui-sha` present) is untouched.

**Why `manifestSha256` is in every package record, and why one line:** a
PR that changes the artefact re-records every package against the new
sha256, so it still conflicts with every in-flight registry PR, which
must regenerate on the merged tree (that is today's behaviour, kept for
the case where it matters). Two PRs on the same package touch the same
line and conflict. One line per file guarantees git cannot merge a
digest from one side with an artefact sha256 from the other; a record
that is not in the producer's one-line form is RED.

**What this does not force, stated rather than implied:** two PRs on
different packages that each leave the artefact unchanged, where only
their combination would change it, land without any run seeing the
combination. Under the single digest their textual conflict forced that
run. Here the next regeneration of any registry package sees the
combination. Dependency-inclusive digests would close this only along
import edges (not between siblings that interact through
`ComponentRegistry`), at the fan-out above. No stale or missing registry
source passes: every source edit still reds until regenerated, and every
package added to or removed from the reached set still reds.

## Self-test (Ask 3)

29 rows (was 22), roster and floor updated. New or changed in-tree rows:
a staged edit to one package is RED naming that package and NOT naming
`plugin-form` or `core`; a new package the registry reaches; a new
workspace edge (now names `ui/packages/unrelated`); a removed package
(names `scripts/sdui-manifest.sources/types.json`); a deleted package
record; a record verified against another artefact; a record out of its
one-line form; a record from the single-digest layout; and the producer
re-recording only the moved package (it back-dates the shared record
first, so a producer that stamps today's date reds the row).

Ablations, one at a time through `scripts/ablation-replace.mjs`, each
restored with `git diff HEAD` empty:

| guard removed | self-test |
|:--|:--|
| the other-artefact comparison (gate) | exit 1, row "a package record
verified against another artefact is RED" expected RED got GREEN |
| the one-line form comparison (gate) | exit 1, row "a package record
out of its one-line form is RED" |
| the unreached-record list (gate) | exit 1, row "a removed registry
package is RED naming its record" |
| the `generatedAt` carry (producer) | exit 1, row "the producer
re-records only the package whose sources moved": the second run changed
the shared record too |
| the per-package record write (producer) | exit 1, the two producer
rows |

## The generator (Ask 4)

It re-records only the packages whose line changed and leaves the others
byte-identical; it removes a record the registry no longer reaches; it
prints which packages it re-recorded. The shared record moves only with
the artefact: `generatedAt` and the provenance version are carried while
the artefact's sha256 is unchanged (the same rule the pin-keyed fields
already follow), so a regeneration on another day no longer moves a
shared line.

**It still builds, and why:** the build plus the enumeration is the only
oracle that says whether `sdui.manifest.json` moved. A record written
without it would stamp a current digest on a manifest no run checked,
which is the narrowing this card forbids. The 8 minutes is a cold turbo
cache, not the build of one package. Measured on `0a6948210c`, one
container, no remote cache (`TURBO_TOKEN` unset):

| run | wall | turbo |
|:--|:-:|:--|
| cold | 8m49s | 0 of 63 cached, 8m43s |
| nothing moved | 6s | 63 of 63 cached |
| one edit under `plugin-grid/src` | 33s | 61 of 63 cached, 27s |
| one edit under `core/src` | 2m50s | 41 of 63 cached, 2m44s |

## Readers moved (Ask 5)

- `docs/releases-maintenance.md`: the regenerate section describes the
per-package records, the re-record rule and which PRs still conflict.
- `scripts/check-changeset-no-major.mjs`: the "measured out" comment,
the real-build assertion (neither record is shipped) and the
records-only control now cover `scripts/sdui-manifest.sources/`. Its
`--self-test` passes (456 assertions).
- `scripts/ui-import/README.md` is not edited: line 140 records what the
cutover's step 4 wrote at the time, and that file still exists at that
path.
- `git grep -n "sdui-manifest.record"` outside CHANGELOGs: only the two
scripts, the docs page, `check-changeset-no-major.mjs` and that
historical README line. `git grep registrySourcesSha256`: only the
retired-field list and its self-test row.

## After this lands: what an in-flight ui PR does (once)

A ui PR whose branch re-recorded the old single digest conflicts once on
`scripts/sdui-manifest.record.json`:

```sh
git merge origin/main                                   # conflict in scripts/sdui-manifest.record.json
git checkout --theirs scripts/sdui-manifest.record.json  # take main's layout
node scripts/gen-sdui-manifest-node.mjs                  # re-records only this PR's package(s)
git add scripts/sdui-manifest.record.json scripts/sdui-manifest.sources sdui.manifest.json
git commit
```

After that, PRs on different registry packages no longer conflict with
each other on the record.

## Gates

Local, on `a8114011` (derived with `node scripts/pm/dispatch-gates.mjs
--commands`): 39 of the 40 derived families exit 0, including `node
scripts/check-sdui-manifest.mjs --self-test` (29 cases) and `node
scripts/check-sdui-manifest.mjs`, `node
scripts/check-changeset-no-major.mjs --self-test` and `--base
origin/main`, `pnpm check:nul-bytes`, `pnpm check:watch-hint-literal`,
`pnpm check:required-contexts`, `pnpm check:doc-authoring` and `pnpm
--filter @objectstack/lint run check:doc-formula-expressions` (after the
lint closure build). `pnpm check:pm-dispatch-gates` was still running
when this PR opened; its reading goes into the report comment on #26051,
and CI runs it in Lint & Repo Gates.

CI on the PR's check-runs: pending at the time of writing; the seat
reads them.

Changeset: none, `skip-changeset`. The diff touches only root `scripts/`
(private), `docs/` and the records; `sdui.manifest.json` is
byte-identical and neither record is copied into
`packages/console/dist/` (the `check-changeset-no-major` real-build
assertion above).

Not Tier H: no governed path (`git diff --name-only` has no `.github/`,
`.claude/`, `docs/adr/`, `skills/`, `AGENTS.md`, `CLAUDE.md`).

## Acceptance notes

- `i18n` is the hot package: 5 of the 12 open ui PRs touch its `src/`,
so they still serialize on `scripts/sdui-manifest.sources/i18n.json`.
Whether locale data belongs in the digest at all is a narrowing question
(it would let a source change pass unregenerated), not something this PR
decides.
- The cold-cache cost of a regeneration (8m43s here) is a
fresh-container property; card threads run with no turbo remote cache.

---
_Generated by [Claude
Code](https://claude.ai/code/session_01SMJGPXxFBpbgAgXdnsXKVS)_

---------

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