Deckbuilding games are a great way to train rationality. They're usually composed of high-dimensional cost/benefit analyses under uncertainty: you don't know precisely what future enemies you might encounter, or the particular hand you'll be dealt. All you can do is shift probabilities by adding or removing cards, tweaking complex systems and synergies with emergent characteristics. One of the high points in such games is uncovering the truly broken combination, where through meticulous planning and arrangement you can heal or deal arbitrary amounts of damage.
Video games of this genre which I'd recommend include:
There are, of course, tabletop deckbuilding games - but these usually operate much slower than their video game counterparts, as the players try to remember the rules and manually execute the consequences of their actions. Similarly, there are trading card games like Magic: The Gathering, though preconstructed decks have less uncertainty associated with future card acquisition. Draft versions of Magic, where players build their decks from random boosters, address operating under uncertainty slightly better, though become expensive quite quickly. They also require friends you can meet in person, who are mutually willing to learn the game, with an open schedule, and a similar skill level. Automating enemy gameplay eliminates these problems.
I think it's more useful to play a variety of such games rather than play the same game on a very deep level: top-tier experts in a specific game might have a deep understanding of the specific mechanics of that system, but the targets to train for are the generalisable skills of operating under uncertainty and managing complex systems. There are other skills which deep knowledge of one system trains too, of course, but I would (weakly) argue they prepare you less for novel scenarios than a wider range of games. (Antifragility or whatever.) There's also obvious consequences if you do this poorly - you lose the game - which encourages more accurate calibration for behaviour in the future.
The player usually has an interesting mix of perfect and imperfect knowledge: they might know exactly the cards in their deck, but not what specific hand they'll draw. If they want to know what cards remain in their deck, they'll have to keep track of which have been drawn and played and plan accordingly. This makes them better candidates than typical roguelikes, which sometimes rely on skills of dexterity rather than cognition, and where the cost-benefit analyses are typically more straightforward. There's also less need to "grind for experience", where it's just execution without inducing insight.
Deckbuilding games hit a sweet spot for this purpose: they're sufficiently abstract for the skills to be generalisable, sufficiently simple to learn mechanics quickly, sufficiently complex to allow for synergistic and emergent gameplay, and sufficiently probabilistic that cost/benefit analyses are extremely useful. Just be careful not to get trapped in a skinner box, sinking hundreds of hours into the one game, and you should be golden.