build(nix): declare build inputs explicitly and fix dev shell env - #1781

Summary

I've been building Vicinae from the flake and working in nix develop, and went looking for why a couple of things behaved oddly in the dev shell. That turned into four small commits: two on the package, two on the dev shell. The only change to the built output is the removed (inert) modules-load.d file; every other installed byte is identical. The evidence is in the collapsed section at the bottom so you can reproduce it rather than take my word for it.

Each commit stands alone, so happy to split this, drop any part, or reshape it entirely.

1. Undeclared dependencies (0ab07cfe)

nix/vicinae.nix sets strictDeps = true, so I expected buildInputs to be the complete picture. The CMake actually requires eight things that aren't in it: wayland-scanner and wayland-protocols (cmake/Wayland.cmake), OpenSSL (src/lib/crypto, vendor/sqlcipher), X11::xcb / X11::xcb_keysyms (src/server), and the bare xkbcommon / udev link names (src/server, src/lib/linux-utils, src/snippet).

They resolve today because qt6.qtbase and layer-shell-qt propagate all eight transitively.

Nothing is broken. The reason to declare them anyway is that the arrangement is invisible exactly when it matters. Two cases from this repo, and neither is a criticism, they're the point:

• feat: encrypt databases using sqlcipher if enabled #1578 added find_package(OpenSSL REQUIRED), and did update nix/vicinae.nix in the same change (swift / apple-sdk for Darwin), yet openssl still didn't get added, because propagation meant nothing ever failed.
• refactor: unvendor merged wayland-protocols #1753 made system wayland-protocols a hard configure-time requirement (FATAL_ERROR when absent). nix/ wasn't touched, and stayed green because layer-shell-qt happened to carry it.

If a nixpkgs bump ever changes what Qt propagates, configure fails with wayland-protocols not found in a diff that has nothing to do with Wayland, and whoever bisects it has a bad afternoon. I read this as finishing what strictDeps already opted into, but if you'd rather deliberately lean on Qt's propagation and keep the list minimal, that's a fair call and I'll drop this commit.

2. modules-load.d (5943d423)

INSTALL_MODULES_LOAD_CONFIG defaults to ON, so the package installs $out/lib/modules-load.d/vicinae.conf. systemd only reads modules-load.d from /etc, /run and /usr(/local)/lib, and NixOS generates /etc/modules-load.d from boot.kernelModules, so in a store path the file is inert, and it reads as though the package handles loading uinput when it doesn't. Same shape of problem as #1607 (a package manager wanting to own installed files, there for Gentoo); the two compose rather than conflict. That PR adds a coarse switch, and the Nix build wants the existing fine-grained one, since it does want the themes, desktop file and icon.

3. Dev shell (441b65a4, 85521f69)

Both are artifacts of ordering rather than anything anyone got wrong.

441b65a4: #1129 added export CC=${pkgs.gcc15}/bin/gcc (and CXX) to get gcc 15 into the shell, which was the right fix at the time. #1545 later rebuilt the shell as mkShell.override {stdenv = package.stdenv;}, and since gcc15Stdenv.cc is pkgs.gcc15 (verified below), the exports became a second copy of what the stdenv already does. The two CMAKE_*_COMPILER exports were no-ops all along: CMake reads CC/CXX from the environment and never the CMAKE_*_COMPILER variables (verified below, and documented). With the compiler exports gone the hook only carries the QML import paths, so this commit also drops the isLinux guard on the shellHook: qmlls is equally useful on macOS. Say the word if you'd rather keep it guarded.

85521f69: #1189 introduced the aggregate qtEnv; #1636 later added the QML import paths so qmlls could see Nix-installed QML modules, but pointed them at pkgs.qt6.qtdeclarative rather than the qtEnv that already existed, so qmlls saw a subset of what the shell provides (org.kde.layershell and QtWayland were missing). This is really just finishing #1636.

# bash, from the branch checkout root, clean tree
diff <(nix-store -qR "$(nix eval --raw "git+file://$PWD?ref=main#packages.x86_64-linux.default.drvPath")" | sort) \
     <(nix-store -qR "$(nix eval --raw .#packages.x86_64-linux.default.drvPath)" | sort)
# -> 8 paths: one vicinae-0.24.0.drv and one *-source per side, two npm-deps.drv per side

main=$(nix build "git+file://$PWD?ref=main" --no-link --print-out-paths)  # full source build
branch=$(nix build . --no-link --print-out-paths)
tmp=$(mktemp -d)
cp -r "$main" "$tmp/main" && cp -r "$branch" "$tmp/branch" && chmod -R u+w "$tmp"
export LC_ALL=C  # sed must treat the ELF binaries as bytes, not UTF-8
find "$tmp/main"   -type f -exec sed -i "s/$(basename "$main"   | head -c32)/@self@/g" {} +
find "$tmp/branch" -type f -exec sed -i "s/$(basename "$branch" | head -c32)/@self@/g" {} +
diff -r "$tmp/main" "$tmp/branch"
# -> Only in .../main/lib: modules-load.d   (the intended removal; nothing else)

Happy to clarify any of these further.