diff --git a/setup/scripts/install-packages b/setup/scripts/install-packages
index 0884224..dab3f48 100755
--- a/setup/scripts/install-packages
+++ b/setup/scripts/install-packages
@@ -1318,6 +1318,18 @@ fi
 
 if [[ "$ROLE" == server ]]; then
 	echo -e "\n--- Installing packages (server) ---"
+	# A server never installs Terra, but one that already has it (installed by
+	# hand, or by an older Panama) has it enabled for the update below. Pin it
+	# first, exactly as a desktop would, rather than update through a stock
+	# terra-release file with metadata checks off.
+	if rpm -q terra-release >/dev/null 2>&1; then
+		terra_status=0
+		_terra_effective_status || terra_status=$?
+		if (( terra_status != 1 )); then
+			log "Pinning the Terra repository this server already has"
+			install_terra_repository > /dev/null
+		fi
+	fi
 	log "Updating all packages. This may take a while"
 	sudo dnf update -y --refresh > /dev/null
 	install_list core-packages "Core"
diff --git a/tests/setup/desktop-first-contract b/tests/setup/desktop-first-contract
index 55cd360..bd63798 100755
--- a/tests/setup/desktop-first-contract
+++ b/tests/setup/desktop-first-contract
@@ -90,6 +90,15 @@ PY
 sed -n '/^if \[\[ "\$ROLE" == server \]\]; then/,/^fi/p' "$installer" | grep -q '^\s*exit 0' \
     || note 'the server branch does not exit before the desktop section'
 
+# A server pins a Terra it already has, but never installs one: the Terra call
+# in the server branch must sit behind the terra-release check.
+server_branch="$(sed -n '/^if \[\[ "\$ROLE" == server \]\]; then/,/^fi/p' "$installer")"
+if grep -q 'install_terra_repository' <<<"$server_branch"; then
+    awk '/rpm -q terra-release/ { guarded = 1 } /install_terra_repository/ && !guarded { bad = 1 }
+         END { exit bad }' <<<"$server_branch" \
+        || note 'the server branch can install Terra on a server that never had it'
+fi
+
 # ── Everything fragile is actually tolerated ────────────────────────────────
 #
 # Being below the desktop is only half of it. A bare `dnf swap` below the