feat: persistent disk wired + passthrough v3.1

GGUF_DIR routes to /data/gguf when HF Pro persistent disk is mounted. Passthrough mode only: qwen3:1.7b triage + nomic-embed-text.

startup.sh

@@ -22,6 +22,24 @@ echo "║   LiteLLM     :7860  (Claude-compat proxy) ║"
 echo "╚═════════════════════════════════════════════╝"
 echo ""
 
+# ── Persistent disk (HF Pro) ──────────────────────────────────
+# HF Pro persistent storage mounts at /data
+# GGUF models cached here survive Space restarts (no re-download)
+PERSISTENT_DISK="/data"
+GGUF_DIR="$PERSISTENT_DISK/gguf"
+OLLAMA_MODELS_DIR="$PERSISTENT_DISK/ollama-models"
+
+if [ -d "$PERSISTENT_DISK" ]; then
+    echo "✅ Persistent disk detected at /data"
+    mkdir -p "$GGUF_DIR" "$OLLAMA_MODELS_DIR"
+    export OLLAMA_MODELS="$OLLAMA_MODELS_DIR"
+else
+    echo "⚠️  No persistent disk — using ephemeral storage"
+    echo "   Enable in: Space Settings → Persistent Storage → mount /data"
+    GGUF_DIR="$HOME/.ollama/gguf"
+    mkdir -p "$GGUF_DIR"
+fi
+
 # ── 1. Start Ollama ───────────────────────────────────────────
 export OLLAMA_HOST=0.0.0.0:11434
 export OLLAMA_KEEP_ALIVE=5m