Can a test be run to see if top human corr players (playing as centaurs) can defeat unaided top engines running on some 'regular' hardware?"
Well the 'test' inmho is still simply the ICCF competition system itself; so
why running a simple test about something which (still) is obvious for more advanced chess players (at least for me (*))? Maybe SFNNue looks like approaching perfect chess but there still are areas where it's not perfect, e,g, complicated positions (look at the latest thread on video review of some deep analysis programs), practical endgames (maybe in such positions we would need another Nnue than for the opening/middlegame (just and idea), etc
For the 'centaur' it depends on his/her (advanced) chess knowledge, eg in practical endgames it's a matter of strategy how to win positions with double bishops (the famous bishop pair) vs bishop knight,SF- Nnue doesn't seem to give a large advantage for the bishop pair; indeed it's often difficult or impossible to win just based on such an advantage, while it depends on lots of other positional elements (usually pawns, king's position, and sometimes also other pieces).
This also shows that correspondence chess still has a 'purpose', namely to enhance your understanding of chess, especially openings (although Nnue is pretty good, i expect new/updated opening trees will emerge as result of the new engines), but also middlegame/endgame positions, and 'practical' endgames.
Sure we are approaching at top level draw equality, but i don't think that running an SF engine for an hour without human assistance/understanding will give top level correspondence chess and will guarantee a draw. Especially not when some rules would be modified, to decrease the (increasing) draw score at *top* level correspondence chess (as was discussed in the 'Kaufman' thread about correspondence chess in the Nnue area).
If you think otherwise, try it, you may get to Iccf 'expert' level, and in the meantime also increase your chess understanding (although there may be more time effective ways to achieve this, i agree), but i still think you won't get
beyond 'expert' level without more in-depth understanding & experience.
Or accept following my challenge, I start with White (**), you can replay with your engine move (let it run for one or two hours, as Uri B suggested should be sufficient). Yes i'm serious, i hereby challenge you, and subsequently you also get your 'test'; five rules, 1) you are not allowed to update the Neural net in later stages (presuming you are going to use SF Nnue, 2) nor your current hardware (at least for the game you now can accept with Black, if you want and 3) you will not increase engine thinking time (beyond the 2 hrs) , 4) being honest and only plugging in the engine move, you (ofcourse) will not accept any human help or change strategy after reading my (or others) comments during the game, and 5) time forfeit for me is impossible (unless i would have abandoned the game or so eg. due to disease, in such a case it may count as draw), but generally i will try to respond with a maximum of -average- one week per move or so (probably faster).
Good luck !
NB although this forum is not for correspondence chess, such a test should be allowed imho, especially also with the comments/discussion which it can generate.

(*) for the opening stage, Uri Blass (GM correspondence player) wrote that SF-Nnue nowadays is good enough to get openings moves which cannot be beaten; but i don't think this is always true; make one suboptimal move (for Black) and still may lose the game, especially if for whatever reason you (ie the engine) got into a sharp, highly unbalanced line.
(**) probably 1.Nf3 or so, but i will decide on that after you accept my challenge.