Might as well switch to IPsec. Everything is already in place, including hardware acceleration. But most people won't, and we'll live in a world with duck-tape hacks built around a compromised Wireguard-ish layer.
Storytime: 15 years ago I co-founded a startup to build easy-to-use infrastructure management for AWS. At that time, it did not have cross-region VPC peering or routing, so you couldn't easily and safely have apps that communicate between regions.
So I started working on creating an overlay network. My idea was to use IPsec, it even has an RFC that documents its kernel interface. So that when a host wants to send a packet to the secure network, the kernel goes to my userspace daemon, that in turn goes to the central server that provides it the key for the given host pair.
And it turned out that the interface lacked a crucial part - on-demand key negotiation for incoming packets. It had this for _outgoing_ packets, but not incoming. The only sane way to make it work was to create a proactively updated database of all the hosts.
Well, I did that. It also did not work (tm). I found so many issues with broken MTU handling, broken NAT traversal, etc.
I eventually gave up on that idea and started working on a simple TUN/TAP-based overlay. I almost made everything work, but our startup got acquired by AWS, and this line of work was abandoned.
I don't think there's anything necessarily complicated about MLKEM that would make this too difficult.
Switching to IPSEC loses you other WireGuard benefits; the wins don't end at "just one carefully curated set of cryptography primitives", but extend into things like DoS prevention and a design that admits to processing incoming frames without dynamic allocation.
Things like Google’s PSP already achieve that while still being hardware acceleratable: https://github.com/google/psp
Post-quantum negotiation will kinda suck but you don’t have to sacrifice everything.
My implementation relied on the extremely finicky Linux kernel IPsec and would never have worked in userspace macOS or windows. But yes, a Tailscale control plane with a per-node switchable data plane, one of which is IPsec with PQC, is a real option that would work.
By locking down the control plane it becomes incompatible with classic IPsec, but also avoids most of the security gotchas that plagued IPsec.
It could have worked in macOS! It implements RFC-2367 (PF_KEY socket type) and I even made it work on my laptop.
I traced the SA database operations on Windows, and it could have been done. Probably. But I abandoned that to preserve my sanity.