Okay so I would find this useful if it would identify every weird case and then write a deterministic execution plan that covered all of them, that I could look at before running it. Like, show me every query in order I need to run to normalize some table. That itself would be bery useful. But I'd never let a bot loose on a production DB (at least an intern can only make mistakes slowly)

Great!

If all plan queries are deterministic, the "Save recipe as Python" action saves the entire process as a deterministic script:

https://www.tamedtable.com/#save-py

Honestly, you are probably not the target user, as you seem familiar with data cleanup scripts and AI agents. I think the target users would be:

A) Non-technical people in your org who clean data by hand because they don't know to automate it.

B) Technical people who want to test multiple queries and preview transformations before generating an AI-driven transformation flow (and are OK with it).