openai/gpt-chat-latest → A (3:2)

Side A changes production behavior by grouping each unranked sibling into its own navigation group instead of combining all unranked siblings, aligning the implementation with the documented design, and adds a regression test verifying the new grouping. Side B adds only a randomized test for the ranking algorithm without changing functionality, so while it improves validation, it contributes less lasting project behavior.