fix(vibe): warm-up runs a real text forward (absorb torch lazy-init)

Loading weights wasn't enough — the first forward still cost ~30s on torch's
one-time graph/thread init. Warm now runs a throwaway _embed_texts() so the first
USER query is ~1.5s, not 30s.

armada/serve.py

@@ -179,8 +179,9 @@ if __name__ == "__main__":
         def _warm():
             try:
                 print("  vibe  : warming CLAP…", flush=True)
-                _vibe_load()
-                print("  vibe  : ready (/vibe, /similar)", flush=True)
+                V = _vibe_load()
+                V["S"]._embed_texts(["warm up"])   # exercise the text tower (torch
+                print("  vibe  : ready (/vibe, /similar)", flush=True)  # lazy-inits once)
             except Exception as e:
                 print(f"  vibe  : disabled — {e}", flush=True)
         threading.Thread(target=_warm, daemon=True).start()