Actually, maybe that’s not a limitation of FSRS.
Notice that even for R<50%, the blue curve (actual retention) is above the orange line (perfect calibration), meaning that FSRS-6 is underestimating retention here. That’s right, even with its crazy long intervals, FSRS-6 is underestimating how likely the user is to recall information.
The only reasonable explanation that I have is that this is FSRS’s way of accounting for the fact that users typically encounter information outside of Anki as well, and if you actually use what you have learned, you will encounter it a lot over the span of >1 year. And FSRS accounts for that by essentially predicting that it will take you a thousand years to forget something.