Wins and losses stay visible.

Only forward-tracked, settled predictions count. Small samples remain labeled as learning data.