[Madraka]

RFC-0045 rollout step 6, first of two PRs: the blocks and the entry editor. Step 6b (the builder, the delete dialog and the publish warnings) follows once this merges.

Both types stay refused on every production write path until step 7, so no production collection can define them and Studio looks unchanged there.

Blocks (packages/ui/src/blocks, all words as props)

Block	What it does
FieldControl	One typed control, extracted from EntryFieldsForm (444 → ~240 lines); EntryFieldsForm and repeater items share it. A reference is a fieldset around the host's renderReference.
RepeaterField (+ repeater-state.ts)	Items as fieldsets with an ICU legend ("Item 2: Ada Lovelace"), folded in Collapsibles; "Add item" + "3 of 20"; up/down icon buttons from sm and a ⋯ menu (move, duplicate, remove). Focus goes to the new item's first control after add, to the next item's toggle (else "Add item") after remove; onRemoved hands the host an undo. Moves and removals are announced in a polite live region. A new error map opens the items it names; collapsed items with problems are listed above the items.
ReferenceList	InsetPanel rows: the title linked, collection · language, "Draft" / "In review" / "Archived" with "Will not show until published", "Not found" with the id in mono; up/down from sm and ⋯ (open, move, remove).
EntryPicker	A LayerDialog opened by its trigger, with focus returning to it. A labelled search over a combobox listbox (aria-activedescendant, arrows, Home, End, Enter) and a polite result count. Collection and status badges. One pick closes the dialog; several are ticked up to the remaining count. "Show more" or "Keep searching" follows the cursor after a truncated read. It takes a search(q, cursor) prop and makes no API calls itself.
UsedByList	A two-column Table (the entry, where the value sits, who else shows it; the status), so it fits 390 px.

Studio

• FieldReference composes ReferenceList and EntryPicker around GET …/entries/pick. The entry's own group and targets already chosen can't be picked.
• Repeaters use FieldImage and FieldReference for their sub-fields. Removing an item shows an info toast with "Undo" for 10 s.
• Save issues map by path: fields/speakers/2/person shows on that control.
• On a translation:
  • shared references and repeaters are read-only, with "From the original" and a link;
  • empty localized ones say "The original: …";
  • a shared list says "Not translated: this list is shared by every language." on multilingual sites.
• "Used by" (UsedByCard) is a flush SectionCard from GET …/entries/:id/references, published referrers first, with "Show more". It shows only when some field of the site can reference the collection, so no production site shows it yet.
• EntryFieldCards lazy-loads both cards; ArticleEditorPage lost its inline lazy (446 → 442 lines in the baseline).
• Bundle: every new module is in LAZY_ONLY. The initial load is unchanged at 391.9 kB gzip.

API (one deviation)

New read-only GET /api/sites/:siteId/entries/groups?ids=a,b (content:read; scopes table and playground allow-list). It answers each group's members in the site, any status, as the picker's cards, with [] for a group with none left. It takes 1 to 100 well-formed ids.

The picker finds new targets only, and references lists referrers, so the editor otherwise had no way to name a stored id or say "Not found". A shared-suite case runs on SQLite and D1.

Deviations (also in the RFC, "Amended in 6a")

• Legends read "Item 2: Ada Lovelace", not "Speaker 2: …". The contract has no singular label for a repeater. The name comes from itemLabel, else the first text, select or reference sub-field.
• The member the editor shows for a target is the one in the entry's language, else the default language's, else the first, whatever its status. This is not pickMember; the locale-fallback cases are left to the publish warnings in 6b.
• EntryPicker's several-choice marks are check icons inside each option, not Kumo Checkboxes, so the listbox nests no controls. A group found in several languages is listed once.
• No capability check in 6a. The editor can't meet the types in production before step 7. 6b gates the builder on the server's capability.

i18n

New words go in the lazy fields namespace (fields.repeater.*, fields.reference.*, fields.usedBy.*, three editor keys and two issue codes). en is hand-written. The other 18 languages are AI drafts awaiting review, imported with pnpm i18n:import. No language is marked reviewed.

Checks

• pnpm design:lint (rtl-logical and rtl-icon included) and pnpm size:check pass.
• pnpm design:shots at 1440 and 390 px passes: no sideways scroll, no console errors. It covered the event editor, a referenced place ("Used by") and the open picker, on a local scratch site seeded through a store with fieldTypes.
• New tests:
  • packages/ui/test/repeater-field.test.tsx: add focus, move focus and announcement, remove focus and undo, last-item removal to "Add item", duplicate, errors opening items and the collapsed list;
  • entry-picker.test.tsx: search through the prop, polite count, arrows and Enter, focus back to the trigger, several up to the remaining count, keep searching from the cursor, plus ReferenceList and UsedByList;
  • entry-fields-form.test.tsx: slots and paths, errors by path, translation fallbacks, undo hand-off;
  • apps/studio/test/fields-references.test.tsx: picker over the pick route, Not found, the undo toast, item issues, Used by.
• pnpm check locally. Lint, size, generated-file checks and typecheck pass. The test stage has three caveats:
  • Run with turbo run test --concurrency=1 --continue, 18 of 19 packages pass, including Studio (84 files) and UI (20). The default --concurrency=3 overloads this machine: unrelated Studio tests time out at 20 s.
  • I raised the timeout of the one test that kept timing out. i18n-messages.test.ts "every catalog file is valid ICU" now has 60 s instead of the 20 s default; it compiles every catalog synchronously and takes about 5 s alone.
  • core-ts: every SQLite variant passes. Only the Miniflare D1 and Durable Object variants of store.test.ts and list-values-query.test.ts fail, all with EADDRNOTAVAIL. That is ephemeral-port exhaustion, with another project's workerd processes running. core-ts's change here is type-only, and CI covers those variants.

🤖 Generated with Claude Code