The redis 4->6 jump regressed production: on the DO droplet the v6 client came up in a SocketClosedUnexpectedly reconnect loop and the server didn't bind :8889, taking chat down until rollback. redis v5+ changed socket/reconnect behavior and needs explicit config (pingInterval / reconnectStrategy), not a bare version bump — so this reverts session-server deps to the proven redis 4.7 / mongodb 6.20 (matching what's live on the droplet) until the upgrade is done properly. The `ip` removal rides along in the revert for coherence; it returns with the redo. deploy.fish is rewritten into a real pipeline: it installs (npm ci, never skipped), restarts via systemd (not pkill+nohup), waits out the ~45s history-load boot before health-checking :8889, and AUTO-ROLLS-BACK to the previous commit if the server doesn't come up — which is exactly what was missing when this broke. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>