~x-ai/grok-latest → A (2:1)

A lands lasting architecture: a settlement worker that batches vote append/apply, warms and serves a ranking cache, and turns ranking/UI paths to read locks—plus full removal of the demo-counter event/UI surface. B only refines pair-suggestion heuristics (bridge sub-priorities and adjacent-rank zip) with good tests; valuable product logic, but narrower than A’s design and concurrency changes.