It feels like there's some dancing around happening here, where the subtext is that where an algorithm is in a NIST contest process isn't determinative on it's own, and that there are external reasons we tend to trust the lattice stuff (because it's been studied for decades, &c).
I don't have a strong opinion in either direction and wouldn't be qualified to express one. I'm a vulnerability researcher with a little bit of cryptography to me, so I tend always to err on the side of utterly skeeved out. I assume, regardless of what national standards bodies recommend, this stuff will all be run hybrid for the foreseeable future, but who knows?
I don't think anybody can easily shut down the argument that it's too early for lattice PQC to simply replace ECDLP cryptography.
I do worry though that early standardization might lock us into a bad standard, and current PQC systems are too inflexible to allow for many use cases. The latter point means that if you want PQC, you might have to give up some features. E.g. multikey signature schemes, zero knowledge proofs, etc.
This is reasonable, but runs contrary to the stance taken by CNSA 2.0.
Now multiply that by 100,000 live connections and 10,000 per second, and you'll see that there's a big problem here.
A lot of cryptography arguments forget that people don't place unlimited value on cryptography (or even security in a general sense). These algorithms are costly, and if they do nothing, should not be used.
The ephemeral key gets signed with your long-lived private key, but that's all your long-lived key is used for.
At this point the only PQC algorithm I would trust with my data on its own is good old McEliece.
SIKE, sure, all bets are off.