would using one LLM to create a spec, and another with only access to the spec doc count as clean enough?
Only if the second LLM can be proven to never got trained/distilled from a model that trained on the original source code.
Probably impossible if the original code is publicly available since most code oriented LLM just gobble them all.
I have done this. Despite using Fable for the spec, and having it double-check the spec a couple of times, it still wasn't airtight. And at some point, the implementing LLM accessed one file of the original code on the Web.
I have by now deviated quite a bit from the original (well, original...) spec.