sync: from ael backend/space/

rerank.py

@@ -113,7 +113,15 @@ MIN_RESULTS = 1
 # the rest in engine order is strictly better than judging none: the survivors
 # are reordered by relevance and the rejects are dropped, and the untested tail
 # keeps the order the race gave it.
-MAX_SCORED = int(os.environ.get("AEL_RERANK_MAX", "6"))
+#
+# 10, from measurement: production runs at ~150ms a document on the Space's
+# CPU, and the race hands over 8. At the old cap of 6 the last two results were
+# never judged, and an unjudged result is the one thing that re-admits exactly
+# the junk this stage exists to remove — "capital of Peru" still came back with
+# a Granta short story and a Nat Geo photo in it. The cap is now above a full
+# result set, so the tail is normally empty; the budget below is what actually
+# bounds the work, and this is only a backstop against a pathologically long set.
+MAX_SCORED = int(os.environ.get("AEL_RERANK_MAX", "10"))
 
 # Recent (query, scores) for threshold tuning. Bounded so a long-lived Space
 # cannot grow without limit.