Puck is the iteration-one editor

The studio's first full authoring loop — palette, outline, drag, field editing, undo — is built on Puck, not written against dnd-kit. The constrained editor is the blocker to any non-developer using Nubbin at all, and every week spent building an editor shell is a week in which no author has used one. Puck ships that shell; what it cannot ship — Nubbin's schema-derived fields, slot legality, compile and publish — is the part worth writing.

Puck was previously rejected, and the rejection was scoped to the end state: its canvas renders its own component tree, so it cannot be the iframe over a consumer's real app that the studio architecture commits to. Iteration one is not that case. It edits the demo's blocks through a workspace dependency — plain presentational React components the studio can hand to Puck's canvas directly — so the limitation that disqualified Puck for the end state does not reach it.

The boundary that keeps Puck replaceable is the draft format. A draft on disk is a Nubbin document, never Puck's own data shape: an adapter derives Puck's Data from the document on load and folds every change back on commit, ids stay Nubbin's, and slot legality maps SlotConstraint.allow onto Puck's slot allow. The CLI's composition verbs and the studio therefore edit the same file, compile runs on every save, and replacing Puck's shell with the dnd-kit canvas later is a UI swap, not a draft migration.

Chosen over building the shell on dnd-kit now — still the settled library for the iframe-of-the-real-app canvas, and untouched by this decision. Building that canvas first means solving cross-frame drag, the studio's hardest problem, before any author has touched an editor; it spends the largest effort on the least-validated risk. Also chosen over storing Puck's Data as the draft, which would have halved the adapter and cost the draft model its independence: compile could no longer run mid-edit, the CLI could not read drafts, and unseating Puck would mean migrating every draft.