So there has to be one winner that dominates, and that has to be the majority implementation.
Yes the majority will overwhelm the minority no change to that. The motivation for this also include use cases like:
1. You do not know in a transition period if the new version is already majority
2. regression testing of pre-releases
If I find one and I'm malicious, then I have no incentive to use it and reveal the issue via the parallel-impl test nodes. Rather, I'd wait until some big chunk of the network have made the re-implementation the dominant (or only) codebase on their nodes, and then I'd split.
A testing protocol makes testing easier, so chances are less a malicious one is the only one who finds the split and can exploit it.
I plan to make my implementation available for testing to anyone through this without requiring to downland install and bootstrap one.