Sorry, I missed that you addressed the request for details in another branch. I was hoping to see a description of what their CPE was actually doing - metadata-based traffic shaping, throttling after a certain amount of usage, too small NAT table like ATT routers were famous for, just underpowered performance, or what. But now I understand you just investigated enough for your own purposes.

FWIW I think the DiffServ RFC explicitly acknowledges that the QoS marks are generally ignored across administrative boundaries. It was an extension of the earlier 4-level QoS priority, so by that time it was clear that transit networks weren't paying them much attention. What would be surprising is finding a consumer ISP that actually did something different when a customer set them.