~x-ai/grok-latest → A (2:1)

A introduces a real settlement worker (batched disk append + vote apply + ranking recompute) and a cached read path so HTTP ranking no longer needs write locks or on-the-fly score iteration, while deleting the whole demo-counter feature surface. B is solid parser/UX work (deterministic typed block masks, prose item-ref tokenization, braced bodies, garden linkify for URLs/dash refs), but it extends presentation/DSL edges rather than core vote/ranking durability.