diff --git a/tests/bench/collection_interning_ab.sh b/tests/bench/collection_interning_ab.sh
--- a/tests/bench/collection_interning_ab.sh
+++ b/tests/bench/collection_interning_ab.sh
@@ -37,6 +37,10 @@
 # identical in both arms, so the ratio stays valid even though the absolute
 # numbers are not prod-tuned.
 MEMTABLE_MB=${MEMTABLE_MB:-8}
+# consecutive samples with an empty queue AND a static record count before an
+# arm counts as drained
+SETTLE_SAMPLES=${SETTLE_SAMPLES:-12}
+DRAIN_TIMEOUT=${DRAIN_TIMEOUT:-10800}
 # plc.klbr.net is dawn's unthrottled mirror; falls back if unreachable
 PLC_URL=${PLC_URL:-https://plc.klbr.net}
 
@@ -140,19 +144,35 @@
         --data-binary "@$SAMPLE" -o "$dir/put-repos.json" \
         -w 'put /repos -> %{http_code}\n'
 
-    # drain: pending must hit zero and stay there, since a repo can be requeued
+    # drain. pending is only the QUEUE depth: it reads zero while in-flight
+    # backfills are still writing records, so waiting on it alone stops the arm
+    # mid-flight and the arms end up having indexed different repo sets. require
+    # the record count to have stopped growing as well.
     echo "draining..."
-    local zero=0 elapsed=0
-    while [ "$zero" -lt 4 ]; do
+    local settled=0 elapsed=0 last_records=-1
+    while [ "$settled" -lt "$SETTLE_SAMPLES" ]; do
         sleep "$INTERVAL"
         elapsed=$((elapsed + INTERVAL))
-        local pend
-        pend=$(curl -s --max-time 5 "http://127.0.0.1:$API_PORT/stats" \
-            | jq -r '.counts.pending // 0' 2>/dev/null || echo 1)
-        if [ "$pend" = "0" ]; then zero=$((zero + 1)); else zero=0; fi
-        if [ $((elapsed % 60)) -lt "$INTERVAL" ]; then echo "  ${elapsed}s pending=$pend"; fi
-        if [ "$elapsed" -gt 7200 ]; then echo "  giving up after 2h with pending=$pend" >&2; break; fi
+        local st pend recs
+        st=$(curl -s --max-time 10 "http://127.0.0.1:$API_PORT/stats" 2>/dev/null || echo '{}')
+        pend=$(echo "$st" | jq -r '.counts.pending // 1' 2>/dev/null || echo 1)
+        recs=$(echo "$st" | jq -r '.counts.records // -1' 2>/dev/null || echo -1)
+        if [ "$pend" = "0" ] && [ "$recs" = "$last_records" ]; then
+            settled=$((settled + 1))
+        else
+            settled=0
+        fi
+        last_records=$recs
+        if [ $((elapsed % 60)) -lt "$INTERVAL" ]; then
+            echo "  ${elapsed}s pending=$pend records=$recs settled=$settled"
+        fi
+        if [ "$elapsed" -gt "$DRAIN_TIMEOUT" ]; then
+            echo "  giving up after ${DRAIN_TIMEOUT}s with pending=$pend records=$recs" >&2
+            echo "INCOMPLETE" > "$dir/INCOMPLETE"
+            break
+        fi
     done
+    echo "  drained at records=$last_records after ${elapsed}s"
     date +%s > "$dir/end_epoch"
 
     curl -s --max-time 60 "http://127.0.0.1:$API_PORT/stats" > "$dir/stats-predrain.json"
@@ -261,7 +281,12 @@
                 printf "  counts_b_per_rec     %.2f -> %.2f  (%+.2f%%)\n", acb, bcb, (acb>0?(bcb-acb)/acb*100:0)
                 printf "  total_b_per_rec      %.2f -> %.2f  (%+.2f%%)\n", atb, btb, (btb-atb)/atb*100
                 if (arss>0) printf "  peak_rss_mb          %.1f -> %.1f  (%+.2f%%)\n", arss, brss, (brss-arss)/arss*100
-                if (ar != br) printf "  !! record counts differ (%d vs %d): bytes/record still\n     comparable but the arms did not index the same set\n", ar, br
+                if (ar != br) {
+                    d = (br-ar)/ar*100; if (d<0) d=-d
+                    printf "  !! RECORD COUNTS DIFFER: %d vs %d (%.1f%%)\n", ar, br, d
+                    printf "  !! the arms did not index the same set, so this is NOT a\n"
+                    printf "  !! valid paired comparison -- rerun with a longer drain\n"
+                }
             }
         }
     ' "$RUN_DIR/summary.tsv"