Misc. improvements for demos

[tslil-topos]

As expected we'll probably need to iterate on the prompt and occasionally on the inference logic.

In today's episode:

• the examples about how to do batched table editing were not clear enough
• the LLM was surprised by the tolerant behaviour
• pre-emptively avoid uuid typing/copying
• even though we disable batched tool calls (and now we tell the agent as well), it's still possible that we'll have to handle these so do so gracefully
• we were overriding the entire return result with the validation error so the LLM lost important output