Honestly, I struggle with this myself. On the one hand I like the diversity of clients; it feels like a sign of strength of the community and protocol that there are many options that have different values. But the cost of this diversity is that it makes things more complicated to coordinate, and different people with different values have different opinions on what a chat client should even want for features.

Something like Slack or Discord can roll out a server feature and client feature to all their clients all at the same time and have a unified experience. But the whole benefit of FLOSS is that anyone can fork the client to make changes, and the whole point of an open protocol is that multiple independent clients can interoperate, and so there's a kind of irony in me wanting those things, but those things producing a fractured output.

So I think XMPP, as a protocol, does the best compromise. These differences between clients and servers aren't just random changes in behaviour or undocumented features, they're named, numbered, alterations that live somewhere and are advertised in the built-in "discovery" protocols. The protocol format itself is extensible, so unexpected content can be passed alongside known content in a message or a server response and the clients all know to ignore anything they don't understand, and virtually all of the XEPs are designed with some kind of backwards compatibility in mind for how this feature might degrade when sent to a non-supported client.

It isn't perfect, but I think perfection is impossible here. A single server and client that everyone uses and keeps up to date religiously with forced upgrades is best for cohesiveness, but worst for "freedom", and a free-for-all where people just make random individual changes and everything is always broken isn't really a community, and XMPP sits in the middle and has a menu of documented deviations for clients to advertise and choose.

As for security, that can be mostly solved with libraries, independent of the rest of the client or server implementation. Like, most clients used libsignal for their crypto, so that could in theory be audited and bug-fixed and all clients would benefit. Again, not perfect, there's always room at the interface between the client code and the library code that's unique, but it's not as bad as rolling your own crypto.