OK, as someone who has been trying different methods of QoS over the past years, with varying levels of success, mainly to have my VoIP phone rock solid over DSL, I'm very interested in what you're saying.
Is there a reason this approach hasn't been implemented yet? Does it break something? If my router is lying to one my upstream router about its TCP window size, wouldn't that impact both the FTP and video stream?
You lie about the window size on a per connection basis, so no, since it's not a global policy, it's a resource policy by application, and potentially by port/IP tuple, so it's not a problem. The point is to keep the upstream router packet buffers relatively empty so that the packets you want don't have to be RED-queued. Nothing breaks because of it.
It generally won't work, unless everyone "plays fair", and the port overcommit ratio for upstream vs. downstream bandwidth is relatively low. As the downstream data rate increases to approach the upstream data rate, the technique loses value, unless you get rid of overcommit, or do it on a per-customer "flow" basis (as opposed to a per virtual circuit "flow" basis) within the upstream router itself, or move to a "resource container" or similar approach for buffer ratio allocation in the upstream router.
So in theory, Comcast (as an example) could do it if they made everyone use the router they supplied, and their routers all participates in limiting upstream buffer impact.
Maybe the next time they replace everyone's cable modems, they'll bother to do it?
Without the deployed infrastructure, it's easier to RED-queue and just intentionally drop packets, forcing a client to request a retransmit as a means of source-quenching traffic. This wastes a lot of buffers, but they probabilistically get through, and for streaming video, that's good enough if there's a lot of client overbuffering going on before playback starts (JWZPlayer, for example, is a common player used for pirated content that will habitually under-buffer so intentional drops tend to make it choppy).
For VOIP, unfortunately, forced retransmit causes things to just typically suck, unless you use a sideband protocol instead, where the router at the one hop upstream peer agrees to reserve buffers for specifically that traffic. This is why Skype is terrible, but your phone calls over your wall jacks which are actually wired to the same packet interface instead of a POTS line are practically as good as a land line or cell phone.
Google hangouts tend to get away with it because they are predominantly broadcast, and are either "gossip"-based CSMA/CD (ALOHA style) networks between participants (i.e. people talk over each other, or wait until the other end is done before talking themselves). It means they tolerate large latencies in which 1:1 VOIP/Skype connections won't. They can be a bit of a PITA for conference calls because of that (Google uses it internally, and gets away with it, but mostly because Google has its own, parallel Internet, including transoceanic fibers), but if Google employees never see the problem, they never fix the problem. Same way any company that assumes local-equivalent bandwidth works as well for their customers as it does for them (free hint to Microsoft inre: Office 386 there).