I'd fork debian. Technically, my installation already have several properties of a fork...
But what would I change?

* better integration of runit and svlogd, to allow getting rid of sysVinit and rc.d remains and benefit from logs in their own places by default (stuff I already have and just takes me, like, 3 minutes to do per daemon, but still);
* dodge the usrmerge bullshit: this change is due *only* because systemd is unable to properly handle mountpoints, which a lot of bug reports on the topic can be found (notably encrypted or network file systems create problems apparently);
* make the source builds easier. Right now, debian uses a pretty complicated and "rich" stack, mixing autotools, perl, python, bash, patch, and whatnot. I think there is no reason for a build mechanics to integrate all this, really. All one should do is rely on upstream's build system, take the *output* instead of messing with options, and build a set of packages from it...
* reduce the usage of pre/post inst/rm in packages. This is mostly done so that packages may support various configurations (and notably init systems) but really there is no need for scripting. Instead, avoiding automatic enabling of services at boot, and eventually move those files in a dedicated package. I suspect this way, minor patches would enable dpkg to install things in user's homedir.
* ban gtk4 and all applications depending on it. Purely and entirely. Gtk3 is still tolerable, gtk4 is not.
* maybe try to migrate to muslC, as this have a huge impact on memory usage. IIRC, we're talking about ~700KiB per process, here. Varying depending on random parameters (and thus between each launch of the same process), because glibc. I.e. `ps` reports that I have no `runsv` process taking less than 772KiB RSS, while using a statically linked or musl, it is exactly 4KiB, the minimum. That means that glibc allocates *a lot* of memory for nothing. May not be a problem on servers or personal computers, but for embedded systems, it's a lot less nice. My systems are relatively lightweight, only 30+ services running around. The total amount of wasted memory is far from negligible, when compared to the overall resource usage of my system. I'm not ecologist, but this kind of resource waste annoys me a lot. And for those who wonder, when I measured systemd's rss usage a whille ago, it was around 30megs, not changing much if adding or removing services.

There are probably some other things I'd change, but those are really the big spots. Ofc, I'd also change the installer, the one I remember is a huge bad joke, almost if done by someone out of spite because tired of reading it was hard to install. The thing forces the operator to stay close to computer by asking a question, doing something, asking next question, etc, except that "doing something" may very well take quite a time, and that questions are never really depending on the success or result of last operation: do you really need to have your partititions done to select what you want to install? Nope. That is just one example, of course.