“Collapse” meaning what, exactly? Do you mean run out of storage from the volume of content, or that processing all the messages is too taxing?

Years back, I setup a Synapse’s server on my personal server (Yunohost). At some point, I joined the “big” Matrix room. Bad idea: RAM and CPU usage went through the roof. I had to kill the server but even that took forever as the system was struggling with the load.

But don’t just take my words for it:

https://github.com/matrix-org/synapse/issues/7339 [...]

It appears that issue is closed as per this comment:

This should hopefully be significantly improved in the upcoming v1.36.0 release. I'm going to close this for now, if people still see issues after updating then feel free to make a new issue.

So pehaps this issue that you are describing is now fixed?

How does it scale differently than Matrix?

[...] XMPP rooms are more conventional: a room is located on one server. That’s an “old” model, but it scales. [...]

This only scales so long as the single server is able to keep up with all of the requests. In the replication, as you have described, all the instances sort of act like load balancers -- they spread the individual requests, and concentrate them into single links between the instances.

And another reason is I may not want to be bothered by people I don’t know, regardless how much I could appreciate reading and/or exchanging with them in the Fediverse.

I think I see what you are getting at with this. Would it be like, for example, if your Lemmy account is also tied to Matrix, then someone on Lemmy could send you a request to talk on Matrix? Granted this could already be assumed to occur if one uses the same username for all of their accounts, but it could possbily be more of an issue if it was more directly integrated. That being said, I'm not sure how realistic this scenario would be since the Matrix protocol is completely independent of Activity Pub. The only connection between accounts that I can think of is OAuth.