Either way, there is nothing magic. At the end of the day, the soonest you can see an action is still physically unchanged. Especially since the prediction model that makes the most sense is just assuming no action.
That said, I dunno enough about MOBAs to try and postulate on how much it matters, and I have no idea if LoL is using lockstep or some other approach.
> The second issue with lockstep is that if one player is lagging, all players are lagging
Where are you getting this?
My game Nebulous uses lockstep, and it's a MMOish mobile multiplayer game. Plenty of clients lag, but one client lagging doesn't cause any other to lag. Furthermore, clients do not have full game state, as an optimization and data transfer reduction first, anti-cheat second.
Path of Exile actually lets you choose lockstep or client-predictive. "Serious" and "competitive" ladder players choose lockstep because, naturally, you want your client's representation of the game state to be as close to the server's representation of the game state as possible, UX be damned.
Prediction could mitigate that (and possibly does, even in sc2), but if one client is suspended and the other is running fine, either every thing that the running client does is moot (and will be reverted) or it gets suspended, regardless of lockstep or prediction.
My speculation: this almost exclusively to protect the private IP of one client being exposed to another client. There is no server processing happening beyond connection verification to help mitigate desync attacks. Deliberate desyncs are still possible though, since I've been the victim of it.
Team Fortress 2 is particularly bad where high latency players can teleport around or you can shoot and hit a laggy player but it doesn’t register by the server due to a rollback. Some players purposely exploit this function by artificially increasing their ping values to absurd levels (700ms or more) which makes them very hard to hit. Many community TF2 servers enforce maximum ping limits for this reason.
Most notably high pings are used to exploit when playing as the Spy class due to the fundamental purpose of the class is to get behind you to backstab you. If you have 700 ping can abuse the fact that you you can get behind players due to the game being out of sync and if you can backstab a player the server will roll back and kill the non-laggy player even though you were no where near behind them from their non-laggy perspective.
I make this distinction because I actually think backwards reconciliation is table stakes for any kind of precision game and is usually pretty easy to implement. It can also help solve teleporting: you run backwards reconciliation on movement commands also (not just shooting or other actions), and combine with a little player position extrapolation.
Nothing fixes a 700ms, actually I would guess the ceiling on competitive play is somewhere around 200ms with incredible netcode. But 200ms gets you pretty far: across the US or the Atlantic Ocean for sure (packet loss over those distances is another story though).
also players only teleport around on your screen in tf2 if you set your cl_interp very low
By contrast, the server rolling back the game state is indistinguishable from the server always being 15ms behind and the clients running the simulation ahead of time (which they do anyway for fluency [1]). The "jumpiness" is called rubberbanding [2].
[1] https://www.gabrielgambetta.com/client-side-prediction-serve...
[2] https://www.urbandictionary.com/define.php?term=rubberbandin...
(I have no idea what LoL in particular supports or implements, but that's the general idea.)