The RX pairs are easy to route on the top layer. I wasn't expecting a challenge there.
This is using 200 μm traces with 150 μm gaps: pretty tight coupling but that's necessary to avoid excessively large traces on OSHPark's 200 μm thick FR408HR prepregs. A loosely coupled pair with 50Ω single ended impedance for the legs would be 380 μm trace and space probably of 500μm or larger!
I could plausibly pull the gap all the way down to the fab limit of 125 μm but that felt like it might be pushing things just a bit too far in terms of repeatability. This is already a non-controlled-impedance process where I'm just hoping I get close enough that the SERDES work decently.