I really like the idea of turning a fighting game like Street Fighter into something turn-based that’s extremely time-aware. Laying out moves on a timeline sounds potentially very fun & intuitive. I’m not sure I understand the mechanics 100% yet though. (e.g. What constitutes “processing the effects” on the timeline? How are pauses/delays/feints handled? How much of your opponent’s timeline can you see? What happens if two attacks overlap in time?)
Limiting your scope to just two characters is a good idea—balancing different decks/move-sets & characters sounds challenging/time-consuming.
Thinking about your question about spatial representation - I think there's an argument to be made for both. 1D movement feels like an interesting abstraction and keeps things very focused. 2D movement provides a much richer possibility space, at the cost of extra mental load and possibly extra dev time. In 2D space, you can have faster/stronger attacks that are focused directly ahead versus slower/weaker attacks that are sweeping. You could also explore knock-back based on the direction of swinging attacks. (I’m imagining movement and attacks being grid-based, though diagonals might be a little weird.)
I wonder about the viability of Reinforcement Learning to train an AI over the course of the jam—have you done this before? I don't have much experience, but it seems like this often takes longer than expected. 2D space could exacerbate that problem too. It might be good to consider some basic hardcoded patterns / reactions as a fallback.
I'm curious to see how this idea develops!