The biggest challenge is context integration - note that the requirement is asking for at least 3 integrations here (other stuff not mentioned).

Question is, do we have a LLM that can do all of that, with a large enough context window or does someone just need write a super prompter revised 1000x times to make it work.

LLMs aren't consistent, we don't know exactly the path the neural networks will take. This is the issue with all of AI, it bound to be unpredictable and its not consistently reproducible.

Which makes it more like Biochemistry than EECS. And if you've ever done work in a wet lab you know physical lab science can be H-E-L-L