Did this not work?

It worked, but as I said above I wasn’t sure about it because it seemed a bit hacky. Since nobody commented, I figured everyone else was also not sure about it.

Possibly I could have labelled it differently to make it clearer I wanted feedback…

Random thoughts - sorry if this somewhat incoherent; I don't have a full overview of what is going on here either

I'm wondering whether we should strive for a more fundamental solution: The quantities xlim, ylim, aspect, box_aspect (explicit or via a fixed position box) over-specify the plot. Three out of four can be specified. The fourth is then induced. set_position already has the which parameter to distinguish between original and induced (active) value. Tracking that on every quantity seems impractical and I'd suspect that we won't be able to cover all combinations of given vs. induced parameters. Would it be more sustainable to track which of the quantities are set and at runtime calculate the real value? Though I'm not clear whether I'm overengineering this.

Fundamentally we have these cases

Legend:

• ~ induced
• x fixed / for limits: (x) fixed either by the user or inferred from data limits, but in both cases fixed as far as "induced" is concerned in the present context.

(*) The interesting part is in here. If data limits become variable, we must adapt one of xlim/ylim. And in theory one of them is enough. The bug is that we adjust both. My wild guess is that we keep one constant and adapt one as an induced effect to placate the above constraints. Then, constrained layout comes along and adjusts position/box_aspect, which results in another limit update, but that update does not know anymore that we intentionally modified one limit and not the other. Depending on the layout situation, it may adjust the other limit so that in the end both are modified.