~anthropic/claude-sonnet-latest → A (3:2)

Side A ships a coherent, working improvement: it removes dead demo scaffolding and introduces a real settlement-worker architecture (batched writes, cached ranking scores, warm cache on boot) that fixes a genuine performance/correctness concern (recomputing ranking on every read), with matching test updates. Side B is a large structural refactor (CanonicalItemUrl -> ItemId) that touches many files but is mostly mechanical renaming/type-swapping with fallback `unwrap_or_else(ItemId::opaque(...))` hacks scattered around, indicating an incomplete/risky migration, and it deletes a large unexecuted plan.md rather than completing the plan it describes.