Correct, it's all about defining terms and understanding what they represent in the right context. A coin flip will get the right answer (in a scenario where the two answers are equally likely), exactly 50% of the time, and it will trend ever closer to that 50% the more times you run the scenario. In this case, a coin flip is defined as 0% predictive. Not because it gets it wrong every time, but because it gets it wrong exactly 50% of the time. If picking the winner via a proposed method is exactly as good as a coin flip would be, the method being compared is 0% better than coin flip - essentially useless. It is 0% predictive. 0% predictive = getting exactly half the guesses right, 100% predictive means getting every single guess right, 50% predictive means getting 75% of the guesses correct.
SR is 70% predictive overall. That translates to picking the correct winner, for games that have a winner, 85% of the time. (divide the 70% by 2, add it to 50%, and you'll see how it becomes 85% ). For the top 100 teams, it's 64% predictive, picking a winner correctly 82% of the time.
But even if something has a very high likelihood of happening, that doesn't mean it's going to happen every time - as is clearly demonstrated by the finals result here!