Yes, as long as context size increase and llm improve at least there's a way out through using AI but once the progress stops...

Huh? Even if progress somehow stopped, current models are already good enough to help -- and the quality of a given vibe-coded throwaway codebase will be higher the more recently it was created.

No, current models and tools are more than capable of dissecting entire codebases.

And even if they somehow weren’t, we’d just do what we used to do with documents to turn them into "chewable bites": chunking, extracting, summarising,…

The worst I would ever get was "here's our Access database - can you rewrite it". That was utterly useless to me.

What I needed to do was sit with a user (not a manager/the person buying my services) and ask them to show me the different things they did with the software. Then I could write a spec for the actual _feature_ and would only need to look at the existing codebase if they needed data transferring across[1]. I don't see why our new LLM-based future would be any different

[1] Of course this meant I would leave out edge-cases and/or weird quirks of the system - often this was actually a bonus as they were either no longer relevant or worked that way because that was the only way they knew how to do it