Lexical caret harvest rows need range-edge contracts

Problem

Lexical caret tests are useful, but most of their surface is not portable to Slate. The valuable part is not the caret object API; it is the behavior around traversal, ordering, splitting, and destructive range edges.

Symptoms

• The source file contains caret constructors, flips, splices, node-key checks, decorator rows, token/segmented modes, traversal ordering, common ancestor logic, range deletion, and splitting helpers in one test suite.
• Slate already has strong Editor.before, Editor.after, positions, Point, Path, split operation, and text-unit owners.
• The missing useful coverage can be small: expanded range deletion should trim both edges and collapse at the anchor.

What Didn't Work

• Copying the Lexical file directly would encode Lexical's caret classes and node-key model.
• Treating every caret test as a query test would miss destructive edit behavior.
• Treating token/segmented text modes as generic Slate core behavior would create fake parity work.

Solution

Split caret harvest files by behavior family before writing Slate tests:

• traversal and ordering -> query contracts;
• text units -> text-unit contracts;
• splitting -> operation or transform contracts;
• destructive range edges -> delete contracts;
• node-key, caret class, decorator, token/segmented rows -> reject or route to a separate explicit owner.

For the LexicalCaret pass, the useful Slate proof landed in .tmp/slate-v2/packages/slate/test/delete-contract.ts:

• expanded range deletion trims both sibling text-leaf edges;
• expanded range deletion trims cross-block edges into the anchor block;
• selection collapses at the anchor after the destructive edit.

Why This Works

The Slate contract should prove behavior a user or API caller can observe. Lexical's caret classes are implementation machinery, but range deletion semantics are editor behavior. Keeping the proof at the range edge avoids a giant copied matrix while still covering the bug-prone part.

Prevention

• Do not port upstream caret helper APIs unless Slate exposes the same API.
• For harvested caret tests, first ask which Slate owner already proves the row: query, text units, delete, split operations, or browser navigation.
• Add one compact behavior proof for the missing edge instead of copying every upstream permutation.
• Keep browser/native caret claims out of package tests.

Related Issues

• Adjacent block void delete must path-delete one target
• Delete selection normalization must distinguish inline spacers from inline targets
• Adjacent mixed inline range delete should share edge delete helper