~x-ai/grok-latest → A (4:1)

A adds lasting core architecture: a settlement worker that batches vote disk writes and ranking recomputes, a cached-score read path (`ranked_items_cached` / `warm_ranking_cache`), and lock downgrades so ranking panels need only read locks—plus removal of the throwaway demo-counter path. B is a real, well-tested UX fix (per-group rank gradients, center-anchored winner highlighting, polarity/slider mapping), but it is localized UI/CSS polish versus A’s foundational concurrency and ranking design.