fix: set production VITE_API_URL/VITE_WS_URL explicitly in CI build

The Gitea Actions workflow checks out a fresh copy of the repo via
actions/checkout, which never has frontend/.env.local (it's gitignored,
only exists in the manually-maintained /root/realtime-collab clone).
Without it, Vite silently fell back to the localhost dev defaults and
baked a broken API URL into the production bundle — every CI-triggered
build overwrote a working manual build with a broken one.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

.gitea/workflows/deploy.yml

@@ -15,6 +15,13 @@ jobs:
       - name: Build backend image
         run: docker build -t realtime-collab-backend:latest ./backend
 
+      - name: Import image into k3s containerd
+        # docker build only populates the Docker daemon's store; k3s runs its
+        # own containerd (imagePullPolicy: Never, never pulls a registry), so
+        # the image must be explicitly imported or the pod keeps running
+        # whatever was last imported.
+        run: docker save realtime-collab-backend:latest | k3s ctr -n k8s.io images import -
+
       - name: Roll out backend
         run: |
           kubectl rollout restart deployment/realtime-collab-backend
@@ -22,6 +29,14 @@ jobs:
 
       - name: Build frontend
         working-directory: frontend
+        env:
+          # frontend/.env.local holds these for the manually-maintained
+          # /root/realtime-collab checkout, but it's gitignored, so this
+          # actions/checkout-based workspace never has it — without setting
+          # them here, Vite silently falls back to the localhost dev
+          # defaults and bakes a broken API URL into the production bundle.
+          VITE_API_URL: https://collab.m1ngdaxie.com/api
+          VITE_WS_URL: wss://collab.m1ngdaxie.com/ws
         run: |
           npm ci
           npm run build

.gitignore

@@ -40,3 +40,5 @@ postgres_data/
 loadtest/pprof
 
 /docs
+
+.playwright-mcp/