I hesitate to call "the shit that keeps stuff older than a year working on that kernel" a tech debt. It's necessity in any general use OS.

The stuff that can be changed without breaking compatibility is, well, it was the developer's best idea at the time and some of them turned out to be good, some turned out to be bad.

Going "but on paper that idea would be better, let's make OS around it" rarely ends up being plain better. It's not one dimensional.

For example if your CPU scheduler makes all jobs run faster (less cache thrashing while keeping CPUs busy etc.) but is unfair, that might be great for people running batch jobs, but shit for desktop users.

Scheduler wasting cycles but making sure interactive tasks get the right level of priority might be improvement for desktop users but not for server use etc.

Or you figure out that the reason something was "unnecessarily complex", was actual necessary complexity that wasn't obvious at first.

Also Linux isn't exactly stranger to taking the whole bunch of code and replacing it with something better, I think we're at 3th firewall stack now (ipchains -> iptables -> nftables)