fix(preflight): gate GPU runtime warning on detected backend

Fixes #748. The NVIDIA Docker runtime warning fired on AMD Strix Halo
systems because the default backend fallback was "nvidia" when no GPU
was detected via sysfs. Changed default to "cpu", added explicit
elif/else branches so NVIDIA warnings only show for NVIDIA backends,
and added /dev/dri positive check for AMD. CPU mode gets a clean pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Lightheartdevs

dream-server/dream-preflight.sh

@@ -48,8 +48,8 @@ detect_backend() {
         fi
     done
 
-    # 4. Default to nvidia (installer would have set .env anyway).
-    echo "nvidia"
+    # 4. No GPU detected — default to cpu.
+    echo "cpu"
 }
 
 BACKEND=$(detect_backend)
@@ -129,14 +129,17 @@ if [[ "$BACKEND" == "amd" ]]; then
             else
                 warn "/dev/kfd not found — ROCm containers may fail"
             fi
+            if [[ -d /dev/dri ]]; then
+                pass "AMD GPU device nodes available (/dev/dri)"
+            fi
             GPU_FOUND=true
             break
         fi
     done
     if [[ "$GPU_FOUND" == "false" ]]; then
         warn "No AMD GPU detected via sysfs"
     fi
-else
+elif [[ "$BACKEND" == "nvidia" ]]; then
     # NVIDIA: check nvidia-smi
     if command -v nvidia-smi &> /dev/null; then
         GPU_INFO=""
@@ -156,6 +159,8 @@ else
     else
         warn "nvidia-smi not found — NVIDIA GPU features unavailable"
     fi
+else
+    pass "CPU mode — no GPU runtime required"
 fi
 log ""