~anthropic/claude-sonnet-latest → B (6:4)

Side B refines a core algorithm (pair selection for ranking) with clear rationale, careful priority tie-breaking logic, and new targeted tests, improving the actual ranking quality feature of the app. Side A rips out a large, working (if complex) keystroke-autocomplete parser graph and replaces it with a much simpler paste-and-go textarea, which is a legitimate simplification but net removes substantial functionality (live autocomplete, guides, scrolling suggestions) and deletes a dedicated race-condition browser test without replacing that coverage, making it more of a regression/simplification than a strict improvement.