Upstream BlockNote issue draft: PM node-view recreation discards React state in custom blocks

Reported from a production OSS editor (Windlass) built on BlockNote 0.55.

Proposed title

Custom block React state silently resets/unmounts when ProseMirror recreates the node view (DOMObserver treats external DOM writes as content changes)

Environment

• @blocknote/core / @blocknote/react / @blocknote/shadcn: 0.55.0
• react / react-dom: 19.3.0
• yjs: 13.6.33, y-prosemirror: 1.3.7
• ProseMirror (resolved): prosemirror-view 1.42.6, prosemirror-state 1.4.4, prosemirror-model 1.25.12, prosemirror-transform 1.12.2
• Radix stack under the block's chrome: @radix-ui/react-dialog 1.1.23, vaul 1.1.2
• Browser: Chromium (Chrome 141-class); reproduced in dev and in jsdom-based replay tests.

Minimal repro

1. Render a BlockNote doc containing a custom block whose React component holds
   useState (e.g. an open/closed flag for an in-block dialog, or any form draft).
2. Open the dialog / focus a form field inside the custom block.
3. Trigger either of the two recreation classes below (easiest: open a Radix modal
   dialog anywhere in the app so react-remove-scroll writes aria-hidden onto
   elements outside the dialog, including the editor's node-view DOM).
4. Observe: within one flush cycle (~25 ms in our traces) the custom block's React
   subtree is destroyed and recreated — every useState resets. An open sheet closes
   itself; a half-filled form resets. The editor content is unchanged; only the block's
   local UI state is lost.

Forensics (what we measured)

• The loop: PM's DOMObserver.flush() → readDOMChange sees DOM mutations inside
  the node view's DOM, fails to attribute them to a known transaction, and the view's
  update cycle ends in ViewTreeUpdater/NodeViewDesc deciding the container
  "fails matchesNode" and recreates the node view. React then unmounts the
  custom block's subtree and mounts a fresh one — all local state gone.
• Trigger class 1 — selection transactions: every selection change moves widget
  decorations between DOM containers. Our block's widget DOM lands in a new parent;
  if the move races the observer flush, the container diff is misread as a content
  change and the block is recreated.
• Trigger class 2 — Radix modal chrome: Radix (and vaul's drawer, built on it)
  writes aria-hidden (via react-remove-scroll) onto elements outside the dialog —
  including BlockNote's node-view DOM inside the editor contenteditable. PM's
  DOMObserver does not ignore those attribute mutations, so the containers get marked
  NODE_DIRTY and the next updateState recreates the affected blocks.
• Both classes share the root cause: PM cannot distinguish "DOM changed by the app
  around my node views" from "content changed", and its fallback is destruction.
  Any React state living inside the node view is collateral damage.

Our fixes (shipped, tested)

1. Stop the modal-chrome trigger: render the block's sheets/dialogs with
   modal={false} (Radix) so react-remove-scroll never writes aria-hidden into
   the editor DOM. This removes one of the two recreation triggers entirely.
2. Survive the remaining trigger (state hoisting): move the block's UI open-state
   (and any state that must survive) OUT of the block's React subtree:
   • a module-level store keyed by block id, read through useSyncExternalStore;
   • the sheet rendered through a STABLE module-level React createRoot on a detached
     host div outside the editor DOM, so a recreation of the owning block cannot
     unmount it;
   • handler indirection ("bridge") so the stable-root-rendered sheet always calls the
     current instance's callbacks after a recreation re-syncs them.
     Regression tests replay the destruction (unmount + fresh mount of the block while
     the sheet is open) and assert the sheet stays open, keeps its draft, and completes
     its flow through the recreated instance.

Fix 2 works but is a lot of ceremony for "a checkbox that survives a repaint"; we'd
rather not keep building it per block.

Proposed upstream directions (pick one, or propose better)

1. Node-view reuse stability guarantee: make DOMObserver ignore attribute-only
   mutations that PM itself did not observe as content (or at least aria-hidden /
   data-* writes by known UI libraries), and make ViewTreeUpdater prefer updating
   an existing node view over recreate when only the container's position moved.
   Concretely: treat container moves caused by widget decoration redraws as non-content
   events.
2. Official external-store escape hatch: a first-class API for custom blocks to
   declare state that must survive node-view recreation (e.g. a useBlockStableState
   hook or a per-block "state root" the host mounts outside the node view), so apps
   don't hand-roll module-level stores + stable roots + handler bridges.
3. Documentation escape: if neither is feasible short-term, document the hazard
   (state inside custom blocks is ephemeral under recreation; modal chrome writes into
   the editor DOM are a trigger) prominently, with the modal={false} recommendation
   for any overlay mounted near editor content.

Reproduction pointers for maintainers

• Our replay pattern is framework-agnostic: mount a custom block with a controlled
  Radix/vaul sheet inside it, open the sheet, dispatch a selection transaction or open
  a second modal elsewhere, and watch the block's React subtree remount (React DevTools
  or a mount counter shows the reset).
• Happy to provide a standalone repro repo if useful.