I can see a threat model already from 2014.

Anyway, I think it’s a tradeoff that it’s hard to assess quantitatively, as risk is always subjective. From where I stand, the average person using native clients and managing their own keys has a much higher chance to be compromised (by far simpler vectors), for example. On the other hand, someone using a clean OS, storing the key on a yubikey and manually vetting the client tool can resist to sophisticated attacks better compared to using web clients.

I just don’t see this as hill to die on either way. In fact, I also argue in my blog post that for the most part, this technical difference doesn’t impact the security sufficiently to make a difference for the average user.

I guess you disagree and that’s fine.