Okay. gnu gettext vendors libxml2. The following patch drops the dependency on gettext for libxml:

diff --git a/pkgs/development/libraries/libxml2/common.nix b/pkgs/development/libraries/libxml2/common.nix
index bc43cfd9d6e3..f16656b0798b 100644
--- a/pkgs/development/libraries/libxml2/common.nix
+++ b/pkgs/development/libraries/libxml2/common.nix
@@ -3,7 +3,9 @@
   darwin,
   lib,
   pkg-config,
-  autoreconfHook,
+  autoconf,
+  automake,
+  libtool,
   python3,
   doxygen,
   ncurses,
@@ -55,7 +57,9 @@ stdenv'.mkDerivation (finalAttrs: {
 
   nativeBuildInputs = [
     pkg-config
-    autoreconfHook
+    autoconf
+    automake
+    libtool
   ]
   ++ lib.optionals pythonSupport [
     doxygen
@@ -112,7 +116,11 @@ stdenv'.mkDerivation (finalAttrs: {
       ln -s cygxml2mod.dll python/.libs/libxml2mod.dll
     '';
 
-  preConfigure = lib.optionalString (lib.versionAtLeast stdenv.hostPlatform.darwinMinVersion "11") ''
+  # Uses `autoreconf -i` instead of autoreconfHook to avoid gettext dependency
+  preConfigure = ''
+    autoreconf -i
+  ''
+  + lib.optionalString (lib.versionAtLeast stdenv.hostPlatform.darwinMinVersion "11") ''
     MACOSX_DEPLOYMENT_TARGET=10.16
   '';
 

However, adding libxml2 to gettext buildInputs (to make it use thedynamic lib) throws obscure splicing errors, and i am not sure where to even start fixing it. I can fix builds, but eval in splicing and stdenv bootstrap is painful. Ideas welcome.

Looks like you can't use fetchFromGitLab in the bootstrap process (like fetchpatch)? With this patch, it evals (I haven't built):

diff --git a/pkgs/development/libraries/libxml2/default.nix b/pkgs/development/libraries/libxml2/default.nix
index 41d2b1f953f0..43df6360ade5 100644
--- a/pkgs/development/libraries/libxml2/default.nix
+++ b/pkgs/development/libraries/libxml2/default.nix
@@ -2,6 +2,7 @@
   lib,
   callPackage,
   fetchFromGitLab,
+  fetchurl,
   fetchpatch,
 }:

@@ -49,14 +50,12 @@ let
         ];
       };
     };
-    libxml2 = callPackage ./common.nix {
+    libxml2 = callPackage ./common.nix rec {
       version = "2.15.1";
-      src = fetchFromGitLab {
-        domain = "gitlab.gnome.org";
-        owner = "GNOME";
-        repo = "libxml2";
-        tag = "v${packages.libxml2.version}";
-        hash = "sha256-FUfYMq5xT2i88JdIw9OtSofraUL3yjsyOVund+mfJKQ=";
+      src = fetchurl {
+        name = "libxml2-${version}-source.tar.gz";
+        url = "https://gitlab.gnome.org/api/v4/projects/GNOME%2Flibxml2/repository/archive.tar.gz?sha=refs/tags/v${packages.libxml2.version}";
+        hash = "sha256-2py4DlRIlD4h1x8379d+whzsNbh8ofHpNAad/QFsTuw=";
       };
       extraMeta = {
         maintainers = with lib.maintainers; [

Does anyone know where the fact that the Darwin stdenv builds CMake twice comes from? As far as I can tell, it's from stage 0, and then just gets used in the other stages from there. Am I missing something here, is it something with the overrides? It looks like it might be, but then the fact that those are only allowed in the final stage (per booter.nix) (when that doesn't seem true, since then they wouldn't be defined...?) comes up.

(Isn't this the same pattern (defining in one stage and referencing in the others) that makes Glibc only build a limited number of times in the Linux stdenv?)

You think cmake should be rebuild less? Or more?

glibc's is probably a bit different as it's a part of stdenv.cc.libc and mainly used by that I would guess. Also, if depends if the package is used or not by other packages in the derivation would affect rebuild count as well.