The bare minimum would be one for the oldest GHC release that Nixpkgs wants to support.
Unless there's a serious problem with the tarball, there would be little need to change it.
Having tarballs for other major.minor releases could provide a shorter build path for newer versions.

We could drop the tarball once upstream provide a tarball that can build all Nixpkgs supported versions, or at the very least the active Nixpkgs default.
AFAIK upstream has made no effort to support booting GHC with a newer major.minor release, so I suspect that we'll need to keep around one of our own tarballs past the release of an official tarball.

GHC tarballs tend to be quite large, but if necessary, we can configure the build to be the bare minimum needed to boot GHC and make them much smaller.
I would need to experiment with this to find how small we can go, but I think it's reasonable to say that in the worst case it would be no larger than the current upstream bindists, which are meant to be used as fully functional GHCs.
For reference, upstream bindists for 9.12.1 seem to be around 300 MiB with xz compression.

When using the multiple outputs setup hooks, is it possible to define per-output propagated dependencies besides manually adding them to the relevant file in the nix-support directory of the output?

As an example, a header-only output may have a dependency on another derivation’s header-only output. My understanding is that the current way to handle this would be to add the dependency in propagatedBuildInputs, though that would add the dependency to every output, correct?

In reply to @truby:matrix.org
I wanted to test something from gcc's git master branch, is it difficult to get a stdenv with a gcc built from there? I tried the obvious (overriding gcc-unwrapped's src = ..) and it doesn't build, it seems to still be trying to apply patches even though I have patches = [];
Am I missing something to get this working, or is it just more difficult than this?

vcunat: do you remember what needed --oldincludedir= in https://github.com/NixOS/nixpkgs/commit/def75f028240ce2d38ec8e0707cc886c8b7596a4#diff-d5ce1dbe37e35a5baa932aac914a2f1a32ac46267f9400f2fff393f31d4aa44bR36 ? Looking at https://www.gnu.org/prep/standards/html_node/Directory-Variables.html it feels like it's something that predates gcc. It now breaks sqlite-3.49.0 https://github.com/NixOS/nixpkgs/pull/373944#issuecomment-2645813426

And I wonder if we should just drop --oldincludedir= completely.