Slate v2 character delete must share text-unit boundaries

Problem

Slate's text-unit helper knew that complex Unicode sequences should be handled as text units, but reverse unit: "character" deletion had a separate complex script restoration path. That split let package distance tests pass while destructive delete still corrupted the behavior.

Symptoms

getCharacterDistance("\u0BA8\u0BBF", true) returned 2, but reverse delete removed only the trailing Tamil vowel mark.
Existing Thai delete fixtures expected partial deletion of the grapheme.
Lexical #7163 rows exposed the mismatch once the test exercised real editor deletion, not just the string helper.

What Didn't Work

Adding more getCharacterDistance rows alone was too weak; the delete transform could still diverge.
Keeping the old complex-script branch preserved legacy fixture behavior, but it made reverse deletion use a different law than forward deletion and Editor.before(..., { unit: "character" }).

Solution

Make unit: "character" deletion obey the same boundary owner everywhere.

The fix removed the complex-script reverse-delete reinsertion path from .tmp/slate-v2/packages/slate/src/transforms-text/delete-text.ts, then updated the Thai fixtures to expect a whole text-unit deletion.

The regression lock lives in .tmp/slate-v2/packages/slate/test/text-units-contract.ts:

const assertUnitCharacterDeletion = (
  testCase: LexicalGraphemeCase,
  reverse: boolean,
) => {
  const editor = createTextEditor(
    testCase.text,
    reverse ? testCase.text.length : 0,
  );

  for (const distance of distances) {
    const before = getEditorText(editor);
    const expected = reverse
      ? before.slice(0, before.length - distance)
      : before.slice(distance);

    editor.update((tx) => {
      tx.text.delete({ reverse, unit: "character" });
    });

    assert.equal(getEditorText(editor), expected, testCase.description);
  }
};

Why This Works

Editor.before and Editor.after already derive character stops from getCharacterDistance. The delete transform should remove the range those APIs select. Reinserting part of the removed text after the fact created a second definition of "character" for reverse deletion only.

Once the restoration branch is gone, forward delete, reverse delete, and text-unit measurement all share one boundary source.

Prevention

Any new text-unit row should have both distance proof and destructive delete proof.
Reverse-delete fixtures for Unicode text should assert whole text-unit behavior, not partial combining-mark behavior.
When external harvest rows expose a helper/transform mismatch, fix the shared behavior owner instead of adding transform-specific exceptions.