This does look like a more accurate formula.
I saw the implementation in cheesecake method · Luc-Mcgrady/anki@ff5b104 · GitHub . This uses sum(R)_end - sum(R)_start in both the Memorized and Ratio graphs. I think it would be better to use the difference only in the Ratio graph. The Memorized graph should use sum(R)_end. Also, because sum(R)_start is the same for all values of DR, it can be calculated separately and used when plotting the Ratio graph.
To be more informative, we can add a dashed horizontal line to the Memorized graph representing the sum(R)_start.
Also, as some comments in original PR on GitHub suggest, the time in the ratio can be replaced by number of reviews. If both of these changes are made, the Ratio graph can be called “Knowledge gain per review”. This is actually “knowledge gain over simulation span per (review/day)”. So, the unit will be cards*days/review. The correct term is complicated and the above suggested term is potentially misleading as, for example, you can’t have a knowledge gain of 3 cards by doing just a single review. The solution is to use total reviews in the denominator instead of average reviews. The only problem with this approach is that the values of the ratio obtained can be quite low (of the order 10^-2). The usual solution for such a problem is to use a multiplier, e.g. “Knowledge gain per 100 reviews”. But, I am not sure if people would like it.
Regarding optimal DR, I too think that the calculated value of optimal DR might be below 70%. But, with such DR, the first intervals of the cards would be in years or decades, which makes no sense. This is actually a limitation of FSRS. FSRS is not very accurate in the low R range because most data used for training FSRS is in the high R range. See Expertium’s explanation below.
This means that Anki will practically never show you a card again. This defeats the purpose of using Anki in the first place. So, it is important to set a reasonable lower limit of DR (i.e. 70%).