feat(nav): enforce strict embedding length guards and autonomous brain-first navigation

GramAddict/core/qdrant_memory.py

@@ -125,10 +125,10 @@ class QdrantBase:
                 if key:
                     headers["Authorization"] = f"Bearer {key}"
                 # OpenAI/OpenRouter use 'input' instead of 'prompt'
-                payload = {"model": model, "input": str(text)[:8000]}
+                payload = {"model": model, "input": str(text)[:2000]}
             else:
                 # Local Ollama
-                payload = {"model": model, "prompt": str(text)[:8000]}
+                payload = {"model": model, "prompt": str(text)[:2000]}
 
             # Log to prevent user from thinking the bot is hung during model swap in VRAM
             if not getattr(self, "_has_logged_embedding", False):
@@ -361,7 +361,7 @@ class UIMemoryDB(QdrantBase):
         sig = re.sub(r"\s+", " ", sig).strip()
 
         # 3. Strict truncation for nomic-embed-text context window
-        return sig[:4000]
+        return sig[:2000]
 
     def _deterministic_id(self, intent: str) -> str:
         """