Hotswappable PCIE is really hard to do right, especially if you are using a non-standard connector. You have to guarantee that pins connect and disconnect in the right order or you risk damaging equipment.

And on top of that, while a connector might be perfectly safe with cold swap (i.e. power off), you can damage the connector if the circuit is completed and power starts flowing but you only have partial contact.

So it was probably preferable to just release a coldswap versions and try dabbling with a hotswap capable chassis in the future.

Connecting the ground first has been solved in many ways. Are there other significant issues here for pin order?

it seems like the right approach would be to have a switch that disconnects the wires while the GPU is being connected to ensure everything is connected physically before anything is connected electricly.

It's two 74 pin connectors that's way too many to physically switch. Also signal integrity through that many switches sounds like an absolute nightmare.

Speaking from experience, there is a big performance difference connecting your eGPU over PCIE and connecting over Thunderbolt. Usually people just take the performance hit with Thunderbolt because of convenience and they are usually using overpowered desktop GPUs but there are eGPU PCIE connectors for people who want the most performance.

Because this is already using a mobile chip with lower performance, Framework probably didn't want to tank performance any more by going with the lower bandwidth Thunderbolt option.

The performance hit is because Thunderbolt has at most PCIe x4 bandwidth, while internal GPUs are connected to PCIe x16 lanes. That's a massive loss of bandwidth.

But it also tells you a lot about why Framework didn't make x16 bus hot swappable, it's a significantly bigger engineering challenge.