perf: batch embeddings in background thread to fix SSE timeouts

Root cause: 306 embedding calls at 61ms each blocked the request
thread for ~19s before forwarding to Anthropic.

- Batch all admitted objects into single embed_batch() call
- Run in background thread (non-blocking)
- store_object accepts pre-computed embeddings
- Goal detection uses turn heuristic instead of blocking embed

src/mnemosyne/gateway.py

@@ -1983,6 +1983,7 @@ def create_app(
                 new_objects = segmented[len(session._segmented_objects) :]
                 admitted_count = 0
                 rejected_count = 0
+                admitted_objects = []
                 for seg_obj in new_objects:
                     # Phase 4d: admission control — score and gate
                     has_dup = False
@@ -2005,23 +2006,45 @@ def create_app(
                         rejected_count += 1
                         continue
                     admitted_count += 1
-                    with Timer(session.benchmark.latency["embedding"]):
+                    admitted_objects.append(seg_obj)
+
+                # Batch embed + store in a single background thread.
+                # One embed_batch() call replaces N individual embed() calls.
+                if admitted_objects:
+                    _batch = list(admitted_objects)  # snapshot
+
+                    def _store_batch(s=session, objs=_batch):
+                        try:
+                            embedder = s.object_store._embedder
+                            embeddings: list[list[float]] = []
+                            if embedder is not None:
+                                texts = [o.content for o in objs]
+                                with Timer(s.benchmark.latency["embedding"]):
+                                    embeddings = embedder.embed_batch(texts)
+                            for i, so in enumerate(objs):
+                                emb = embeddings[i] if i < len(embeddings) else []
                                 _run_async(
-                            session.object_store.store_object(
+                                    s.object_store.store_object(
-                                session_id=session.id,
+                                        session_id=s.id,
-                                content=seg_obj.content,
+                                        content=so.content,
-                                object_type=seg_obj.object_type,
+                                        object_type=so.object_type,
-                                source_tool=seg_obj.source_tool,
+                                        source_tool=so.source_tool,
-                                source_key=seg_obj.source_key,
+                                        source_key=so.source_key,
-                                stub=seg_obj.stub,
+                                        stub=so.stub,
-                                tags=seg_obj.tags,
+                                        tags=so.tags,
-                                key_entities=seg_obj.key_entities,
+                                        key_entities=so.key_entities,
-                                turn=seg_obj.turn_start,
+                                        turn=so.turn_start,
+                                        embedding=emb,
                                     )
                                 )
+                        except Exception:
+                            pass  # non-critical — object store is for search
+
+                    _threading.Thread(target=_store_batch, daemon=True).start()
+                    for _so in admitted_objects:
                         session.benchmark.segmentation.record_object(
-                        seg_obj.object_type,
+                            _so.object_type,
-                        max(1, len(seg_obj.content) // 4),
+                            max(1, len(_so.content) // 4),
                         )
                 session._segmented_objects = segmented
                 parts = [
@@ -2104,21 +2127,13 @@ def create_app(
                         )
                     break
 
-            if user_text and session.object_store._embedder is not None:
+            # Skip embedding-based goal detection in the hot path — it takes
-                from mnemosyne.object_store import _cosine_similarity
+            # ~60ms per call and blocks the request thread.  Use turn-based
-
+            # heuristic instead: classify goal every 10 turns.
-                embedder = session.object_store._embedder
-                current_embedding = embedder.embed(user_text)
-
-                # Check for topic shift
             goal_changed = False
-                if session._last_user_embedding is not None:
+            turn = session.token_state.get("turn", 0)
-                    sim = _cosine_similarity(current_embedding, session._last_user_embedding)
+            if turn <= 1 or turn % 10 == 0:
-                    goal_changed = sim < 0.5
+                goal_changed = True
-                else:
-                    goal_changed = True  # First message — classify
-
-                session._last_user_embedding = current_embedding
 
                 if goal_changed and helper_llm is not None:
                     session.benchmark.goals.record_topic_shift()

src/mnemosyne/object_store.py

@@ -409,10 +409,12 @@ class ObjectStore:
         tags: list[str] | None = None,
         key_entities: list[str] | None = None,
         turn: int | None = None,
+        embedding: list[float] | None = None,
     ) -> StoredObject:
         """Create and store a new semantic object.
 
         Auto-generates: ID, timestamps, token estimates, embedding, stub.
+        Pass ``embedding`` to skip the per-object embed call (for batch use).
         """
         now = datetime.now(timezone.utc).isoformat()
         object_id = uuid.uuid4().hex[:16]
@@ -422,8 +424,9 @@ class ObjectStore:
             preview = content[:80].replace("\n", " ")
             stub = f"{object_type}: {preview}..."
 
-        # Compute embedding
+        # Use pre-computed embedding or compute one
-        embedding: list[float] = []
+        if embedding is None:
+            embedding = []
             if self._embedder is not None:
                 embedding = self._embedder.embed(content)