I haven't used Discourse, but what you describe sounds like the way that Slashdot has been doing moderation since the late 90s, by randomly selecting users with positive karma to perform a limited number of moderation actions, including meta-moderation where users can rate other moderation decisions.
I always thought that this was the ideal way to do moderation to avoid the powermod problem that reddit and lemmy have, although I acknowledge the other comments here about neglecting minorities being a result of random sampling of the userbase, but it is likely that this also happens with self-selected moderation teams.
Within minority communities though, a plurality of members of that community will belong to that minority and so moderating their own community should result in fair selections. Another way to mitigate the exclusion of minorities might be to use a weighted sortition process, where users declare their minority statuses, and the selection method attempts to weight selections to boost representation of minority users.
A larger problem would be that people wanting to have strong influence on community moderation could create sock-puppet accounts to increase their chance of selection. This already happens with up/downvotes no doubt, but for moderation perhaps the incentive is even higher to cheat in this way.
I think a successful system based on this idea at least needs some strong backend support for detecting sock-puppetry, and this is going to be a constant cat and mouse game that requires intrusive fingerprinting of the user's browser and behaviour, and this type of tracking probably isn't welcome in the fediverse which limits the tools available to try to track bad actors. It is also difficult in an open source project to keep these systems secret so that bad actors cannot find ways to work around them.
You get incremental backups (snapshots) by using
To use this you pass in the previous snapshot location as DIR and use a new destination directory for the current snapshot. This creates hard links in the new snapshot to the files which were unchanged from the previous snapshot, so only the new files are transferred, and there is no duplication of data on disk (for whole-file matches).
This does of course require that all of the snapshots exist in the same filesystem, since you cannot hard-link across filesystems.