fix: persist chat uploads on /data volume (broken images) + prevent duplicate chat sockets

- Broken images: UPLOADS_DIR/REC_DIR/TRANS_DIR were server/<dir> INSIDE the image,
  so every deploy.sh rebuild wiped uploaded files — old images 404'd ('broken
  image') though their DB rows survived. Make them env-overridable and point prod
  at /data/uploads|recordings|transcripts (persistent volume), matching DB/downloads.
  NOTE: files already lost to prior rebuilds can't be recovered; new uploads persist.
- Duplicate notifications: harden connectChatWs — close/detach any prior socket
  before opening a new one and keep a single pending reconnect timer, so a flaky
  reconnect can't leave two live sockets delivering every event/notification twice.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docker-compose.yml

@@ -14,6 +14,11 @@ services:
       # Desktop installers + auto-update feed live on the persistent volume so uploaded
       # builds survive image rebuilds (a plain image path would be wiped on every deploy).
       - DOWNLOADS_DIR=/data/downloads
+      # Chat uploads / recordings / transcripts on the persistent volume too, so they survive image
+      # rebuilds (otherwise old shared images 404 as "broken image" after every deploy).
+      - UPLOADS_DIR=/data/uploads
+      - REC_DIR=/data/recordings
+      - TRANS_DIR=/data/transcripts
     # Secrets (TURN credentials, SSO_SECRET, BIZGAZE_WEBHOOK_URL, etc.) live in
     # a .env file next to this compose file. It is gitignored — never committed.
     # See .env.example for the expected keys.