dream.py: cache the SentenceTransformer embedder across retrieve() calls

Pipeline mode calls retrieve() three times (NREM, Early REM, Late REM).
Previously each call re-imported and re-instantiated SentenceTransformer
("all-MiniLM-L6-v2"), allocating ~200MB and spending 30-60s on disk->CPU
init three times sequentially. lru_cache(maxsize=1) makes the load happen
once per process.

Expected: pipeline runtime drops ~100-180s, removes 2x redundant 200MB
allocations, and reduces transient memory pressure during the same window
when other nightly jobs may run.

scripts/dream.py

@@ -16,6 +16,7 @@ import os
 import json
 import sqlite3
 import argparse
+from functools import lru_cache
 from collections import Counter
 from pathlib import Path
 from datetime import datetime, timedelta
@@ -283,6 +284,11 @@ def retrieve_graphiti(mode, task=None, n_results=8, excluded_sources=None):
         print(f"[Graphiti retrieval error: {e}] — falling back to empty.")
         return []
 
+@lru_cache(maxsize=1)
+def _get_embedder():
+    from sentence_transformers import SentenceTransformer
+    return SentenceTransformer("all-MiniLM-L6-v2")
+
 def retrieve(mode, task=None, n_results=8, excluded_sources=None, type_filter=None):
     # E3 experiment: DREAMER_SUBSTRATE=graphiti routes retrieval to Graphiti /search
     # Default behavior: pgvector similarity search (unchanged)
@@ -291,8 +297,7 @@ def retrieve(mode, task=None, n_results=8, excluded_sources=None, type_filter=No
     substrate = os.getenv("DREAMER_SUBSTRATE", "pgvector")
     if substrate == "graphiti":
         return retrieve_graphiti(mode, task=task, n_results=n_results, excluded_sources=excluded_sources)
-    from sentence_transformers import SentenceTransformer
+    embedder = _get_embedder()
-    embedder = SentenceTransformer("all-MiniLM-L6-v2")
     low, high = MODE_RANGES[mode]
 
     if task: