Why is this not possible? Devs pls this would be a quality of life upgrade to replays

Within Spring a replay is played as a stream of actions, as calculate within an engine. The state of the game, where units are, what actions are ongoing etc., is not stored. Only the starting state (also indirectly, as it is calculated). To have this, 'key frames' need to be included, where the game-state is fully known. You could call them 'save-points', like a save, but lightweight.

To create these frames you need to serialize the entire state, this can cause memory/CPU spikes and freezes.

To have a more fluent user experience the game and its engine-instance should not fully reset, just the state.

Point is that it hasn't been made.

I think the best way to understand replays is to know precisely what they store. A replay stores all the things that players told their units to do. So when you select a Flea and right click, that move order is saved in the replay.

You can't know where a Flea is after being given a move order, unless you know where the Flea started and the positions of everything it might hit along the way. Keep in mind that there isn't a unique position and orientation associated with a Flea that has finished a move order, and move orders can even be interrupted.

If we restrict ourselves to move orders, then all Fleas are in one of two states:

• Has been given a move order in the past.
• Has never been given any order.

The position of Fleas that have never been given an order depends on orders given to the factory that created them, nearby constructors, and the whole economy. So to figure out where a Flea is, you need to work backwards along all of its orders, and everything that could have affected how it carried out those orders. At which point you're just running the whole game forwards.

It's counter-intuitive how difficult this is, I'm sure of it. How so? I watched Valve completely eat sh*t when they tried to put in dota. We're not talkin' normal absentee valve either, this was a crunch time, pushing huge QOL updates left and right about 5-6 years ago Valve, and they got completely mopped by this task. First it crashed the client to desktop, then it hung the client, then it got removed and sat in derelict for years. I don't know if it even works now, i don't expect it to.
It appears the "recording" needs to be fully offloaded into a separate recreation, so that it can be interactable on the spectator plane in this way. I think about a memory write header going in reverse and my brain explodes... it would have to be mimicry, video disguised as gameplay... pure fucking witchery.

Now that I think about it, the windback doesn't have to be too complicated. You can make timestamps every 1 minute, every 5 minutes, etc, and only make them DURING replays so that they don't take up storage space. But I'm an amateur programmer, so take my words with a grain of salt.

Adding replay checkpoints to engine would be good and I think it's "just" a matter of extending save/load to handle 100% of the synced game state.

anyone knows what starcraft II does for their replays? They can backwards

Starcraft II seems to have the system proposed here. Basically, commands and checkpoints. This is just from my recollection of watching replay analyses ages ago. It looked like the the system was loading the game at some point in the recent past, then simulating forwards.

The game world of Starcraft is a lot simpler than in Zero-K, where all the angles of the limbs of robots technically matter, because they might try to shoot a projectile out of some of them. Zero-K is also simulated at 34% higher fidelity (30 Hz vs 22.4 Hz). This makes Starcraft checkpoints smaller, and makes rapidly simulating between them faster. But PCs can be quite fast these days, so it isn't like we couldn't have the same system. Especially, and idk whether Starcraft does this, if the checkpoints were post-processed after the live game, so lag spikes were not an issue. I think my PC could process a 20 minute 1v1 in only a few minutes.

i was just wondering im not asking for it but can a super fast replay mode that does not render anything (no animations) while holding a key be made? for potato pc's that struggle with 5x catchup speed

because i noticed that if i zoom into the ground while catching up i get a better catchup speed vs zoomed to show all animations

quote: You can make timestamps every 1 minute, every 5 minutes, etc, and only make them DURING replays so that they don't take up storage space

Cliver5 FWIW the term for this is "key frame" which MSPR mentioned in the first reply to OP. Keyframes more generally are a concept from animation (when artists use interpolation/tweening) and video encoding (most video we watch has an I-frame at least once every few seconds).

Or "checkpoint" since if the idea is to do these only once every few minutes then it's kinda odd to think of them as keyframes.

But to get back to the point: it's totally possible, but also a surprisingly bigass task for such a "simple" feature as others explained. Not that I've looked into Spring dev myself (apart from once in an aeon past) but thinking of it just brings to mind so many things that can go slightly-but-replay-breakingly wrong until all the details are hammered out. Still... possible.

So I am a little surprised that this sort of thing was never implemented in Spring over the last ten years. It's such a high-quality FOSS game engine and its community have done amazing work, but I guess it somehow never topped the list of priorities. Because players lurv their replays, so it would clearly provide a lot of value.