Am pretty sure that the kind of complexity required for that kind of approach would make the whole VM thing far too expensive to consider running for probably many, many years (so "stop right there" unless you are planning years in advance now).
...then why cant all the peers conclude properly that the correct plugin was indeed run on the correct input data (which is the output of NXT VM script)?
What a peer would do is have to run the script itself (get used to the fact that the script will need to run on *every* peer except perhaps for some lightweight ones) and compare its result to that being suggested by another peer.
This is why the script's state must be "deterministic" if a peer tells you that the answer to running script A with state X results in state Y but you think it results in state Z then you ignore that block as being invalid and block the peer.