If one implementation goes down or has a bug, then only the users of that implementation are affected.
This is true for many things but not for consensus systems.
In a consensus system with multiple major implementations if _any_ implementation has an issue everyone has an issue-- and if there is just a "disagreement" but none are objectively wrong, then there is still an issue.
When Bitcoin.com mined a 1+ MB block because of a bug in their code, who was affected? When people found and exploited bugs in the BU client, who were affected?
You wrote, "in a consensus system with multiple major implementations, ..." My ideal is a consensus system
without multiple major implementations. As I wrote earlier, there are 83 different implementation being used right now and yet Bitcoin is humming along. The real problem is that a few of them dominate the rest, and a person striving to make Bitcoin more robust would be working to eliminate the dominance of any particular implementation because that implementation is a point of failure.
The reason for this is that for a consensus system the primary purpose is to be consistent, so any inconsistency is a failure, regardless of whom is "right" vs "wrong" or even an otherwise harmless difference. Inconsistency immediately creates an opportunity for funds theft, when you think a state is final but it really isn't and it gets reversed.
An implementation is a _precise_ definition. A pile of paper is not, or to the extent that it is precise it's still not that relevant: if paper says X and the network does Y, then Y it is, you can't generally just undo Y later without undoing transactions that people thought were final and irreversible. Good specs are useful to aid understanding and analysis, but cannot define the network because they cannot control its action from moment to moment.
It is true that an implementation is a precise definition, but is such a precise definition necessary? You write as if Bitcoin does not have a mechanism for resolving disputes between nodes, but it does of course.