Description

Current behavior: in a document whose first block is a block decorator and whose last is a list, Cmd+A then Backspace leaves the root holding a ListNode with no children, and the next inline insert throws Expected node ListNode of type list to have a block ElementNode ancestor out of insertNodes. The invariant is uncaught, so it reaches the application's onError.

Expected, and what this produces: root[paragraph], with the next Shift+Enter inserting normally.

Mechanism

$removeTextFromCaretRange removes the emptied focus block, then walks up removing ancestors emptied with it. The walk breaks when removing parent would empty the root:

if (grandparent && $isRootNode(grandparent) && grandparent.getChildrenSize() <= 1) {
  break;
}

That applies unconditionally, so it stops on a container that has declared it cannot be empty. The fix adds && parent.canBeEmpty().

Three things say the guard is stale. Main has no policy of preserving lists, only one about the root's child count: the same keystrokes over [paragraph, list] remove the list today, because the root then has two children and the guard never fires. The sibling branch of the same function does the same climb with the comment // remove empty parent node even if parent node is canBeEmpty and no guard. And the case the guard protects is now handled below it by $restoreEmptyContainerParagraph, which arrived in #9058 two months after the guard arrived in #8751.

This is not specific to lists

TableNode and the playground's LayoutContainerNode also return canBeEmpty(): false and can sit under the root, so [block decorator, table] goes from root[table[]] to root[paragraph[]] too. The table failure is quieter than the list one: it does not throw, because insertNodes catches an element point on a shadow root first, and the content is spliced into the TableNode instead, giving a table holding a paragraph where its rows belong. One rule looks right for both and both are tested, but happy to split the table half out.

This is visible to consumers: any third party ElementNode returning canBeEmpty(): false that can be a direct child of the root now disappears on this path instead of surviving empty.

Test plan

Six tests in LexicalCaret.test.ts. Four cover the changed branch, a list and a table each replaced with an empty paragraph plus the follow up insert for each. Two are controls that pass in either state.

a container that may be empty is kept is the control that bears on the guard: TestElementNode at the same depth reaches the guard and takes the preserved branch, and replacing the new conjunct with false flips it. Nothing else pins that branch, Issue8745Repro.test.ts from #8751 still passes all 14 with the break neutralized, which is part of why the guard reads as stale.

Reverting only LexicalCaretUtils.ts: 4 failed, 179 passed. Restoring it: 183 passed. Neighbouring suites each pass unchanged, including Issue8745Repro 14, LexicalSelection 100, LexicalLinkNode 74 and LexicalTableNode 42. tsc --noEmit, eslint and prettier --check clean.

I did not run the browser project, so Issue8745DeleteLine.test.ts from #8751 is unexercised. Every element registered there returns canBeEmpty() true, so the added conjunct is a no op and the break expression is identical to main, but that is an argument from the node definitions rather than a run.