Slate React keydown must import DOM selection before model-owned navigation

Problem

Browser-native vertical movement and Slate/model-owned horizontal movement can fight if Slate does not import the current DOM selection before handling the next keydown.

Symptoms

ArrowDown moved the visible DOM caret from paragraph 1 to paragraph 2.
Slate's model selection remained in paragraph 1.
ArrowRight then used stale model selection and snapped the DOM caret back to paragraph 1.
Handle-based tests stayed green because they set model selection without proving DOM selection.

What Didn't Work

Flushing only the throttled onDOMSelectionChange handler. The actual selectionchange was queued through the debounced scheduler.
Adding more isolated ArrowLeft/ArrowRight rows. They did not combine a browser-native movement followed by a model-owned movement.
Trusting selectRange as browser setup. It is useful, but it is not raw DOM selection proof unless the helper also places the DOM selection.

Solution

Add a central DOM-to-model import before model-owned key handling:

if (!isInteractiveInternalTarget(editor, event.target)) {
  syncEditorSelectionFromDOM({
    editor,
    preferModelSelectionForInputRef,
  })
}

The helper reads the current DOM selection and selects the equivalent Slate range unless the editor is explicitly in model-owned mode:

if (range && (!selection || !Range.equals(selection, range))) {
  Transforms.select(editor, range)
}

Also fix browser helpers that were model-only. If a test claims browser navigation, setup must place DOM selection too, not just call the Slate browser handle.

Why This Works

The browser is allowed to own vertical caret movement. Slate is allowed to own some horizontal or structural navigation. The boundary between those two modes must be explicit.

Before a model-owned key reads Editor.getLiveSelection(editor), Slate must import the current in-editor DOM selection unless a model-owned operation is already the active source of truth.

This mirrors the discipline used by better editor architectures:

ProseMirror flushes DOM observer/selection state before key handling.
Lexical updates DOM selection inside one update lifecycle.
Edix syncs selection at the start of input handling.
VS Code owns cursor movement in a cursor/view-model controller instead of mixing DOM and model selection ad hoc.

Prevention

Add navigation gauntlets that chain browser-native movement and model-owned movement.
Test ArrowDown followed by ArrowRight, not only ArrowLeft/ArrowRight in a single line.
Require handle helpers used in browser tests to place DOM selection too.
Keep internal-control guards before DOM-to-model import so inputs, buttons, selects, textareas, and nested editors stay app/native-owned.