TB5 is 15GB/s. So gen 5 equivalent. I'm not saying there are tb5 enclosures in the wilds, but it's a matter of time. Also if you're bottlenecked by buffered, linear reads and writes so much that there is a difference between 3GB/s and 7GB/s then I envy you. Most of what I choke my desktops and servers with is random IO that wouldn't saturate gen2 :)
Thunderbolt 5 is very high data transport, but the latency of going through the TB port is still higher than going through PCIe. In a single large transfer, I'd expect TB5 to win, in a millions-of-tiny-transfers scenario, I'm not so sure.
Thunderbolt is PCIe though, just over an external interface. That's why eGPUs worked so well. I can't see a situation where the latency of Thunderbolt has a significant impact on disk usage when eGPUs, where latency is so much more noticeable, worked acceptably?
Thunderbolt provides a tunneling mechanism for PCIe, DisplayPort, USB etc. It's also a mesh network where packets are source-static routed from node to node in the network - so the source sets up the route-to-the-destination and the data packet is transmitted from controller-node to controller-node until it gets to the destination, then it's unpacked and presented as data to the system.
You could see some of this on the venerable "trashcan" Mac Pro, where one of the TB controllers wasn't directly connected to the port, but came through another TB node. The latencies on the ports connected to this TB controller were slightly higher due to the extra transit-time.
Latencies over PCIe are measured in tens of nanoseconds (say 70-100) depending on chipset and how much you pay. Latencies over TB can be several hundreds of nanoseconds. TB presents as a PCI interface, but that's an adaptor-type design pattern, it's not fundamentally PCIe underneath.
Bandwidth vs latency is like a pickup vs lambo I guess. And what the tb limits is the bandwidth, if you catch my drift (although lambos are awd and poor at drifting). So the actual performance that matters (the snappiness) is still there.