It's the sum of difficulty targets, not the individual difficulty scores. Therefore, two blocks at the same blockchain height are always the cumulative weight (unless we're in a longer blockchain fork across a difficulty reset).

If it were individual difficulty scores, this would open the door for all sorts of gaming, e.g. selfish mining after finding a particularly good block because you're sure you can orphan another, or attempting to orphan an already discovered block because it doesn't exceed the difficulty statement by much. In other words, the blockchain convergence would be a lot less stable if it were individual block scores.