One problem I'd like to solve is making a way for users to override
the system resolv.conf;

 The s6-dns client library uses the DNSCACHEIP environment variable for
this: if it contains a list of DNS caches, this list will override the
/etc/resolv.conf-provided one. (The idea comes from djbdns, but has been
extended to a full list instead of a single cache address.)
 Same thing with the DNSQUALIFY environment variable, which can have
a list of suffixes that overrides resolv.conf. (djbdns had a complex
rules-rewriting-based qualification mechanism that nobody ever used,
so the simpler approach was easier and better.)

 Maybe musl could use the same approach: environment variables are a
reasonable place for hardcoded-path overrides. But it has to be balanced
against namespace pollution.

This seems like a good foundation for a package system. I've looked
into Nixos before but never really tried it out, and got the
impression that the concept was very good but it might not be the best
implementation. So something similar to Nixos sounds interesting. :-)

 I've always believed that the filesystem itself should be used as a
packaging system: every package should have its own system user and reside
in its own directory, and /usr/bin and friends should only contain
symlinks. Native isolation via Unix permissions, atomic package replacement,
easy package management. But for some reason, people seem absolutely
reluctant to do this.

The philosophy used in musl, which is somewhat different from the sort
of philosophy you might have when designing a new distribution, is not
to invent new policy but to avoid policy and build on existing,
already-widely-accepteed policy when it's unavoidable.

 I don't agree with all decisions in musl, but this one I can definitely
stand for.

There are LOTS of ways one could extend hostname lookups, ranging from  NSS modules to
hosts.d and resolv.d, but rather than trying to support everything
imaginable (result: bloat and serious security considerations) in
libc, the musl approach to hostname lookup is that libc contains the
basics that are suitable for most/all simple systems, and anything
more advances can be provided by an external daemon running on
localhost that speaks DNS protocol and provides whatever lookup
semantics you desire.

 In the DNS case, the flexible - and best, IMNSHO - approach is to run a
small local DNS cache on localhost indeed; but the problem is that there's
an existing codebase that sometimes insists on clobbering /etc/resolv.conf,
which adds to the packaging burden when your purpose is to create or maintain
a distribution. Having extension mechanisms at the libc level can help in
that situation.

--
 Laurent