Before writing anything, it helped to list the questions this project is actually about. They are kept here unedited, as a reference point to measure drift against later.
What is the least amount of internal state that meaningfully changes what a unit can represent?
Do networks of stateful units drift toward particular dynamical regimes during training?
Is "near-critical" a useful target, a side effect, or a distraction?
What would a fair comparison against standard architectures even look like?
Write the questions down first, so the answers cannot quietly change them.