Install upgraded confs by atomic rename

A copy truncates the live conf and writes into it, so anything sourcing load_config.sh
during that window reads a partial file with empty path variables.

Deployment/conf_upgrade.sh

@@ -19,9 +19,73 @@
 # associative arrays (declare -A).
 #
 # ==============================================================================================
+# OPERATIONAL MODEL
+# ==============================================================================================
+#
+#   1. Validate both --template and --target exist
+#   2. Parse each into keys, handling scalars, indexed arrays and declare -A
+#   3. Classify every key: ADDED (template only) / REMOVED (target only) / KEPT (both)
+#   4. Emit the merged result — template structure, target values — to a temp file
+#      staged in the target's own directory
+#   5. --dry-run stops here and prints the report
+#   6. --backup copies the current target to .bak
+#   7. Install by atomic rename over the target
+#
+# Called automatically by git_pull_execute.sh after every pull, for master.conf and this
+# host's own host*.conf. The host template is HOSTN_-prefixed and the caller substitutes
+# the real MY_ID before merging.
+#
+# ==============================================================================================
+# DESIGN PRINCIPLES
+# ==============================================================================================
+#
+# The User's Value Always Wins
+#   For any key present in both files, the target's value is kept and the template's is
+#   discarded. The template supplies structure and new keys, never settings. This is what
+#   makes the upgrade safe to run unattended after every single pull.
+#
+# Structure Follows the Template
+#   Comments, ordering and blank lines come from the template, so an upgraded conf reads
+#   like the current version rather than accumulating layers of old formatting.
+#
+# Standalone by Design — Do Not Add load_config.sh
+#   This script sources nothing. It is the tool that repairs the conf that load_config.sh
+#   depends on, so it has to work when that conf is broken, partial, or missing keys.
+#   Sourcing load_config.sh here would make the repair tool fail in exactly the situation
+#   it exists for. That is also why it uses plain echo instead of log()/error(), and why
+#   there is no acquire_lock — common.sh is not available to it.
+#
+# Atomic Install, Never In-Place
+#   The merged conf is renamed over the target, not copied into it. Every watchdog sources
+#   load_config.sh on every run; a cp would truncate master.conf and write into it, and
+#   anything reading during that window gets a partial conf with empty path variables.
+#
+# Concurrency Handled by Atomicity, Not a Lock
+#   Two concurrent runs against the same target cannot corrupt it — each stages its own
+#   temp file and the rename is atomic, so the last writer simply wins. Since the merge is
+#   idempotent, that outcome is identical to running once.
+#
+# ==============================================================================================
 # OPERATIONAL SAFEGUARDS
 # ==============================================================================================
 #
+# Root Required to Write
+#   Installing over the target requires root. --dry-run deliberately does not, so the
+#   change report can be previewed by anyone.
+#
+# Atomic Rename
+#   The temp file is created in the target's own directory so the rename stays within one
+#   filesystem. On Unraid /tmp is rootfs while the confs live on flash, and a cross-device
+#   mv degrades to copy-then-unlink — precisely the torn write this avoids.
+#
+# Permissions Preserved
+#   mktemp creates 0600; the target's existing mode and owner are copied onto the temp file
+#   before it is installed, so a conf does not come back with different permissions.
+#
+# Temp File Cleanup
+#   An EXIT trap removes the staged file on any early exit, and is cleared once the rename
+#   has succeeded so the trap cannot delete the installed conf.
+#
 # Dry-run Mode
 #   --dry-run prints the full change report (ADDED / REMOVED / KEPT) then exits
 #   without writing anything. Always preview before applying to production confs.
@@ -34,10 +98,6 @@
 #   Both --template and --target are validated before any parsing begins.
 #   Missing files abort immediately with a clear error.
 #
-# Atomic Write
-#   Merged output is written to a tempfile first, then copied to the target.
-#   A partial write cannot corrupt the original.
-#
 # ==============================================================================================
 # CONFIGURATION
 # ==============================================================================================
@@ -225,15 +285,35 @@ if [[ "$DRY_RUN" == true ]]; then
     exit 0
 fi
 
-TMPOUT="$(mktemp)"
+# Checked here rather than at the top: --dry-run is a read-only report and is useful to
+# anyone, but installing over a conf under /boot needs root. Plain echo because this script
+# deliberately does not source common.sh — see the header.
+if [[ "$EUID" -ne 0 ]]; then
+    echo "ERROR: writing '$TARGET' requires root (use --dry-run to preview as any user)" >&2
+    exit 1
+fi
+
+# Staged beside the target, not in /tmp. mv is only atomic within one filesystem, and on
+# Unraid /tmp is rootfs while the confs live on flash — a cross-device mv silently degrades
+# to copy-then-unlink, which is exactly the torn write this is meant to prevent.
+TMPOUT="$(mktemp "${TARGET}.XXXXXX")"
 trap 'rm -f "$TMPOUT"' EXIT
 
+# mktemp creates 0600; carry the target's existing mode/owner across so the installed conf
+# does not come back with different permissions than it went in with.
+chmod --reference="$TARGET" "$TMPOUT" 2>/dev/null || true
+chown --reference="$TARGET" "$TMPOUT" 2>/dev/null || true
+
 _write_merged > "$TMPOUT"
 
 if [[ "$BACKUP" == true ]]; then
-    cp "$TARGET" "${TARGET}.bak"
+    cp -a "$TARGET" "${TARGET}.bak"
     echo "Backup: ${TARGET}.bak"
 fi
 
-cp "$TMPOUT" "$TARGET"
+# Atomic install. cp would truncate the live conf and write into it, leaving a window where
+# anything sourcing load_config.sh reads a half-written master.conf — every watchdog does
+# that constantly. A rename swaps the inode: readers get the old file or the new one.
+mv -f "$TMPOUT" "$TARGET"
+trap - EXIT
 echo "Updated: $TARGET"