fix: preserve real exit code in EXIT trap and fold parallel job state

The EXIT trap ended with '[ $ec -ne 0 ] && cleanup_notify_failure ...',
so its last command returned 1 on success and the notify status on
failure — every run exited with code 1 regardless of the actual result.
Rewrite cleanup_on_exit to return the real status.

Parallel jobs run in subshells, so HAD_UPDATES/UPDATED_REPOS/NITRO_BUILT
set inside update_renderer/update_client were lost. Persist per-job
deltas to a temp state dir and re-source them in parallel_wait so the
summary reflects renderer/client updates. Also keep the per-repo yarn
cache between updates (only removed on explicit Clean) and drop the
dead clean_node_modules helper.

update-Nitrov3.sh

@@ -860,7 +860,13 @@ cleanup_on_exit() {
         fi
         restore_binaries
     fi
-    [ $ec -ne 0 ] && cleanup_notify_failure "Exit code $ec"
+    # Preserve the real exit status: the last command in an EXIT trap defines
+    # the script's exit code, so never let a guard/notify call overwrite it.
+    if [ $ec -ne 0 ]; then
+        cleanup_notify_failure "Exit code $ec"
+        return $ec
+    fi
+    return 0
 }
 trap cleanup_on_exit EXIT
 trap 'release_lock; cursor_show; exit 1' SIGINT SIGTERM
@@ -924,14 +930,6 @@ git_update() {
     )
 }
 
-clean_node_modules() {
-    if [ ! -f "package.json" ]; then
-        warn "clean_node_modules: no package.json found, skipping"
-        return
-    fi
-    rm -rf "node_modules" 2>/dev/null || sudo rm -rf "$(pwd)/node_modules" 2>/dev/null || true
-}
-
 # yarn install scoped to a per-repo cache folder. Parallel installs (renderer +
 # client) must never share a cache — yarn 1 corrupts the shared cache, leaving
 # node_modules incomplete (vite/pixi.js missing) while still exiting 0.
@@ -965,11 +963,35 @@ PARALLEL_NAMES=()
 PARALLEL_FAILED=()
 PARALLEL_RESULTS=()
 
+# Each parallel job runs in a subshell, so variable changes (HAD_UPDATES,
+# UPDATED_REPOS, NITRO_BUILT) made inside update_renderer/update_client would
+# otherwise be lost. Jobs persist their state to a per-job file which
+# parallel_wait re-reads back into the parent shell.
+PARALLEL_STATE_DIR="${TMPDIR:-/tmp}/nitro-parallel-$$"
+mkdir -p "$PARALLEL_STATE_DIR" 2>/dev/null || true
+
 parallel_run() {
     local name="$1"; shift
     PARALLEL_NAMES+=("$name")
     (
+        # Snapshot inherited state so we only persist the delta back: the
+        # emulator (run before the parallel phase) may already be in
+        # UPDATED_REPOS, so emitting the full array would duplicate it.
+        local _orig="${UPDATED_REPOS[*]:-}"
         "$@" >/dev/null 2>&1
+        local rc=$?
+        {
+            # Note: exit status is captured via `wait`; only fold state vars back.
+            echo "HAD_UPDATES=$HAD_UPDATES"
+            echo "NITRO_BUILT=$NITRO_BUILT"
+            for r in "${UPDATED_REPOS[@]:-}"; do
+                case " $_orig " in
+                    *" $r "*) ;;
+                    *) printf 'UPDATED_REPOS+=(\047%s\047)\n' "$r" ;;
+                esac
+            done
+        } > "$PARALLEL_STATE_DIR/$name"
+        exit $rc
     ) &
     PARALLEL_PIDS+=($!)
 }
@@ -987,10 +1009,17 @@ parallel_wait() {
             PARALLEL_RESULTS+=("__PARALLEL_FAIL__${name}")
             PARALLEL_FAILED+=("$name")
         fi
+        # Fold the job's persisted state back into the parent shell so the
+        # summary reflects renderer/client updates, not just the emulator.
+        if [ -f "$PARALLEL_STATE_DIR/$name" ]; then
+            . "$PARALLEL_STATE_DIR/$name"
+            rm -f "$PARALLEL_STATE_DIR/$name"
+        fi
         i=$((i+1))
     done
     PARALLEL_PIDS=()
     PARALLEL_NAMES=()
+    rm -rf "$PARALLEL_STATE_DIR" 2>/dev/null || true
 }
 
 parallel_failed() {
@@ -1327,10 +1356,10 @@ do_cleanup() {
         ok "Old config backups pruned"
     fi
     [ -d "$BACKUP_DIR" ] && find "$BACKUP_DIR" -maxdepth 1 -name '*.sql' -printf '%T@ %p\n' 2>/dev/null | sort -rn | tail -n +6 | sed 's/^[0-9.]* //' | xargs -r rm -f || true
-    # Clean both the global yarn cache and the per-repo caches used by the
-    # parallel renderer/client installs.
+    # NOTE: the per-repo yarn cache ($YARN_CACHE_DIR) is intentionally kept
+    # between updates — it is what makes parallel installs fast and reliable.
+    # It is only removed on an explicit "Clean" (menu) action.
     yarn cache clean 2>/dev/null || true
-    rm -rf "$YARN_CACHE_DIR" 2>/dev/null || true
     rm -rf /tmp/nitro-* /tmp/epicnext-* 2>/dev/null || true
     # Clean old Maven build artifacts (older than 7 days)
     find "$EMULATOR_MODULE" -maxdepth 3 -name '*.jar' -mtime +7 2>/dev/null -exec rm -f {} \; || true