.........
My statement in that gedanken experiment is that your 100 000 nodes will not get one single block, and will certainly not enforce their rules on the network. As such, they have no power to do so.
.........
Of course full nodes that don't mine will never get a block. 'Good' non-mining full nodes could enforce the rules, though, but only if it somehow happens that at least one 'good' node is between every conspiring 'evil' mining node so they would not be able to propigate blocks to eachother. Last I saw (which was a while ago), the large pools have a special semi-private "relay network" for their mining full nodes they use, and they could always directly connect to eachother. Users that actually want to get their transaction into a block (since non-mining full nodes don't make blocks) would have to connect to an evil node. There's also nothing stopping the 'evil' guys from popping 100 000 non-mining evil nodes on the network to make it easier for users to connect to the evil-net.
Indeed, you've got it. It is too much seen as "evil guys vs good guys". One should see it as "rule set A" vs "rule set B". If users and miners agree on rule set B, then no matter how many non-mining full nodes only accept rule set A, this doesn't enforce rule set A at all (whether rule set A is the "historical" or the "new" rule set doesn't matter here).
You are right that IF there's a vast majority of non-mining full nodes with rule set A, then a user connecting RANDOMLY to just a full node will most probably only see a "stopped" block chain. The user has to specify a miner node that follows rule set B in order to get his transactions through, and get the live block chain (according to rule set B). So a vast swarm of disagreeing non-mining full nodes can somehow perturb a bit the network, until users configure their wallets to ignore them, and only go to rule-B miner nodes.
And in fact, this is a very good thing, because otherwise, the attack of launching 100 000 full nodes with a different rule set would impose that different rule set, which is against the very idea of PoW securing the block chain: a sybil attack with full nodes would be sufficient if it were true that non-mining full nodes impose their rule set.
Purely technically, as you point out, only miners need to agree amongst themselves to use rule set B, and then only a block chain according to rule set B will be built. But, as was pointed out regularly, and as was used erroneously as an argument indicating the power of non-mining full nodes,
miners will not want to alienate users. Technically, they can, but economically, they would ruin themselves, because alienating the users who sustain the market cap, and who are finally the buyers of their minted coins would kill the revenue of the miners, as they will now technically mine coins nobody wants to buy.
But an army of non-mining full nodes,
in disagreement with the miners (and the users) has no power to impose its rule set, whether that rule set is a "new" one (a kind of sybil attack with an army of nodes) or the "old one" (as "guardians of immutability"). They can at most initially perturb the communication of transactions and blocks, and will in the end be ignored.
That was my point.