The cost of an incidental CPFP is tiny compared to the amount they save on total fees.
It would actually be fairly trivial to write a script which places their transactions at 0.1 vMB from the tip of the mempool, and then constantly watches the mempool. As soon as the transaction falls below 0.7 vMB from the tip (example numbers), then they broadcast a CPFP which puts it back at 0.1 vMB from the tip. They would still be guaranteed the same confirmation times that they currently get by overpaying their fees by 10-100x, they would save a bunch of money, they would stop bloating the mempool so much, and they could pass huge savings on to their users.
But, as you say, they have no incentive to do that, since their users seem quite happy to be utterly fleeced with their ridiculous withdrawal fees, and it plays in to their narrative of promoting their centralized shitcoin.
Having a fraction of the blocks being occupied by a few high fee-rate transactions isn't a problem for the rest of the network, as long as the bulk of the transactions are still paying a normal fee rate.
That's kind of the problem though - the bulk
aren't paying a "normal" fee rate. Every transaction in the last 12 hours could have paid 1 sat/vbyte and taken at most 2 blocks to be confirmed, whereas we only have around 10-20% of transactions paying this amount, meaning 80-90% of transactions are overpaying to some degree. If you want to over pay 2-5 sats/vbyte because you desperately want confirmed in 1 block and not 2, sure, I can appreciate that. But paying 10-500 sats/vbyte is just plain stupid.