fix(ci): ensure /dev/fd symlinks exist for mkinitcpio and procsub

The chroot's /dev only gains fd/stdin/stdout/stderr symlinks at boot, and
container executors often lack them, which breaks mkinitcpio's preset step
(pipes through /dev/stdin) and bash process substitution
("/dev/fd/63: No such file or directory").

- build-iso workflow: create the links in the executor's /dev, then mount a
  fresh devtmpfs and create them there too so the chroot tree is covered
  whether the executor's /dev is devtmpfs or a plain tmpfs
- customize_airootfs.sh: self-heal the same links inside the chroot

.gitea/workflows/build-iso.yaml

@@ -43,13 +43,36 @@ jobs:
           set -euo pipefail
           pacman -Syu --noconfirm --needed archiso jq
 
-      - name: Ensure /dev/stdin exists for mkinitcpio
+      - name: Ensure /dev/fd symlinks exist for mkinitcpio and procsub
         run: |
           set -euo pipefail
-          # Docker/container environments may lack /dev/stdin; mkinitcpio needs it
-          [ -e /dev/stdin ]  || ln -sf /proc/self/fd/0 /dev/stdin
-          [ -e /dev/stdout ] || ln -sf /proc/self/fd/1 /dev/stdout
-          [ -e /dev/stderr ] || ln -sf /proc/self/fd/2 /dev/stderr
+          # mkinitcpio's preset step pipes through /dev/stdin and bash process
+          # substitution (< <(...)) uses /dev/fd; both break if these symlinks
+          # are missing (they are normally created once at boot by mkinitcpio's
+          # init). Two places need them, so cover both:
+          # 1. the executor's own /dev: mkarchiso reads /dev/stdin on the host
+          #    (machine-id, version, pkglist) and the workflow's own scripts use
+          #    process substitution here. A no-op if the links already exist.
+          ln -sfn /proc/self/fd    /dev/fd
+          ln -sfn /proc/self/fd/0  /dev/stdin
+          ln -sfn /proc/self/fd/1  /dev/stdout
+          ln -sfn /proc/self/fd/2  /dev/stderr
+          # 2. the chroot's /dev: devtmpfs is single-instance, so mounting it
+          #    here yields the same tree that pacstrap/arch-chroot expose as
+          #    <chroot>/dev. When the executor's /dev is itself devtmpfs (1)
+          #    already fixed that tree; this pass is what makes it work when
+          #    the executor's /dev is a plain tmpfs or directory instead.
+          d="$(mktemp -d)"
+          trap 'umount "$d" 2>/dev/null || true; rmdir "$d" 2>/dev/null || true' EXIT
+          mount -t devtmpfs devtmpfs "$d"
+          ln -sfn /proc/self/fd    "$d/fd"
+          ln -sfn /proc/self/fd/0  "$d/stdin"
+          ln -sfn /proc/self/fd/1  "$d/stdout"
+          ln -sfn /proc/self/fd/2  "$d/stderr"
+          ls -l "$d/fd" "$d/stdin" "$d/stdout" "$d/stderr"
+          umount "$d"
+          rmdir "$d"
+          trap - EXIT
 
       - name: Build Arch ISO
         run: |

airootfs/root/customize_airootfs.sh

@@ -1,6 +1,19 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
+# The chroot's /dev is the kernel's single shared devtmpfs, which only gets the
+# fd/stdin/stdout/stderr symlinks at boot (from mkinitcpio's init). If the
+# runner environment lacks them, process substitution (< <(...)) and upstream
+# tools (pacman-key) fail with "… /dev/fd/63: No such file or directory".
+# Self-heal here; the CI build-iso workflow does the same before pacstrap so
+# that mkinitcpio's hook (which pipes through /dev/stdin) works too.
+for _link in "fd:/proc/self/fd" "stdin:/proc/self/fd/0" \
+  "stdout:/proc/self/fd/1" "stderr:/proc/self/fd/2"; do
+  [[ -e "/dev/${_link%%:*}" || -L "/dev/${_link%%:*}" ]] ||
+    ln -sfn "${_link#*:}" "/dev/${_link%%:*}"
+done
+unset _link
+
 # Populate the keyring in the built root from the keyring bundles
 # (/usr/share/pacman/keyrings, incl. the copied cachyos keyring).
 # This mirrors the guide; it runs inside the image you are making.