Expanded delete across inline boundaries must remove crossed inline ancestors

Problem

Lexical regression 1083 exposed an expanded-delete boundary that Slate did not prove: deleting a range that starts at inline link content and continues outside that inline should remove the inline element, not leave an empty shell.

Symptoms

• A full selection whose visible content starts with a link could leave an invalid start point under the removed inline.
• A partial selection like Say[link:Hello]World, selecting HelloWorld, could delete text but leave an empty link plus empty trailing text.
• A broad fix broke existing fixtures where deletion happens wholly inside an inline and the empty inline shell is supposed to remain.

What Didn't Work

• Treating every fully-selected inline as removable. That broke selections whose entire range lives inside the inline.
• Copying Lexical's exact browser toolbar setup. The portable invariant is delete behavior across inline boundaries, not link creation UI.

Solution

Add package proof for both accepted rows in .tmp/slate-v2/packages/slate/test/delete-contract.ts:

• delete a full selection where the visible content starts with an inline link
• delete a partial expanded range that starts at an inline link and ends after it

Then make delete-text.ts collect fully-selected inline ancestors only when the expanded range crosses outside that inline. That keeps legacy inside-inline delete behavior intact while removing empty inline shells for boundary-crossing deletes.

Why This Works

Slate normalizes an empty spacer before leading inline content. The full selection proof must include that spacer, while the partial proof starts inside the link. In both accepted rows, the selected inline is not merely emptied; the selection crosses out of it, so the inline element is part of the deleted structure.

Selections wholly inside an inline are different. Those should delete inline contents and preserve the inline shell, which the existing delete fixtures lock.

Prevention

• When changing delete behavior around inlines, run the dedicated delete contract and the transforms/delete fixture suite.
• Split "selection inside inline" from "selection crosses inline boundary" in tests. They look similar but require opposite structural outcomes.
• Keep browser product setup out of the first proof unless the bug is genuinely browser transport.

Related Issues

• Lexical history harvest rows need stack-law contracts