I think you do not understand "selfish mining": it's a technical term for a miner trying to take the lead over the existing blockchain, but only in some very specific cases outside normal mining. It's predicted by game theory - google for it, it's interesting!
The rest of your analysis is correct: as rewards decrease, TX fees may increase. You can also see that due to competition for the limited tps, as seen in 2017.
But this is also exactly why the scenario discussed in the paper is implausible: miners get to pick the transactions with the highest fees for inclusion in the next block. If your transaction is not urgent, you pay the minimum and wait. Eventually, it will be processed - if the expected delay (given the mempool) is not to your liking, you can update the fee with RBF.
But again, it means there will be little variance - just a smooth adjustment, meaning the required condition this whole paper is based on is dead wrong.
About the economic model breaking down, you make 2 mistakes: 1) you fail to account for difficulty adjustments made just for this situation, but even people who do often forgot 2) when revenue no longer covers costs (as say for a factory), the company doesn't immediately give up and fire everybody.
It's industrial organization 101, and due to the difference between short term and long term.
Of course, there is less friction and fewer rigidities with software, and mining equipment could be deployed differently if it was still CPUs or GPUs.
But the genius of ASICs having no alternative use means it can't happen (except maybe switching to another coin with the same algorithm)
Overall, regardless of the situation and what you throw at it, there's no situation I can see where the economic model break - except maybe if miners are tracked and executed on sight by the army? But even then, all it would do is move the mining to another country - or lead to bribes!