How do you picture your hypothetical vision of things working on the server side? Naively using buffering would make the input stream desync from the output stream. "Input prediction" works well enough when a game supports it, but lag would reappear for titles where the game doesn't. A general solution (for games with no input prediction) would require you to run the game in an x86 emulator with save-state + rewind support, such that network stutter could be translated to micro-rewinds of the game's VM; and even that wouldn't work if the game was multiplayer.

I could see games being able to signal "there is no input during this cutscene, so it can be buffered at the client." But for everything besides cutscenes, game streaming essentially has to work like a VoIP call (hard-realtime), doesn't it?