deploy(S4): Blender headless pipeline + WebAR client + backend fixes

backend/memory/episodic_memory.py

@@ -18,21 +18,41 @@ try:
 except ImportError:
     pass
 
+# "models/embedding-001" was RETIRED — it 404s ("not found for API version
+# v1beta"), which meant every embed() call fell through to the zero-vector
+# fallback below. That failed silently and is worse than an outright error:
+# identical all-zero vectors make every document equidistant, so semantic
+# retrieval returns essentially arbitrary results while still "working".
+# Verified live against this key — the only models exposing embedContent are
+# gemini-embedding-001, gemini-embedding-2 and gemini-embedding-2-preview.
+EMBED_MODEL = "models/gemini-embedding-001"
+EMBED_DIM = 768  # keep in step with the existing persona_* collections
+
+
 async def embed(text: str) -> list[float]:
     """Generates an embedding for the given text using Gemini."""
     try:
         def _get_embedding():
             result = genai.embed_content(
-                model="models/embedding-001",
+                model=EMBED_MODEL,
                 content=text,
                 task_type="retrieval_document",
+                output_dimensionality=EMBED_DIM,
             )
             return result['embedding']
-        
-        return await asyncio.to_thread(_get_embedding)
+
+        vec = await asyncio.to_thread(_get_embedding)
+        if not vec:
+            raise ValueError("empty embedding returned")
+        return vec
     except Exception as e:
-        logging.error(f"Embedding failed: {e}")
-        return [0.0] * 768
+        # Loud, not silent: a zero vector destroys ranking, so make it obvious
+        # in the logs that retrieval quality is degraded rather than fine.
+        logging.error(
+            f"Embedding failed via {EMBED_MODEL} ({e}); returning a ZERO vector — "
+            "semantic ranking is degraded until this is resolved."
+        )
+        return [0.0] * EMBED_DIM
 
 class EpisodicMemory:
     def __init__(self, persist_dir: str):