That paper does not demonstrate any party is necessarily operating out of a Nash equilibrium. Selfish mining is detectable, so mining pools, being visible and identifiable entities, face reputational costs for engaging in it. That very plausibly makes it unprofitable. Also, some forms rely on network conditions that may not exist.

If there were some sort of undetectable or unattributable selfish mining known to be compatible with the current network conditions, and we could tell that weren't happening then you could correctly make the argument you did, but that is a contradiction.

I guess one time we might see a valid experiment would if a big pool does an exit scam. Under the reputational-cost theory, such a pool should selfish mine right before the exit, though it still might not be worth the trouble to actually implement for relatively modest gains.