Don't normalize centroids with zero samples. Fixes #15209 (#15213)

Author: msokolov

lucene/CHANGES.txt

@@ -89,6 +89,8 @@ Bug Fixes
 * GITHUB#14847: Allow Faiss vector format to index >2GB of vectors per-field per-segment by using MemorySegment APIs
   (instead of ByteBuffer) to copy bytes to native memory. (Kaival Parikh)
 
+* GITHUB#15213: Don't normalize centroids with zero samples, which caused divide-by-zero. (Mike Sokolov)
+
 Changes in Runtime Behavior
 ---------------------
 * GITHUB#14187: The query cache is now disabled by default. (Adrien Grand)

lucene/core/src/java/org/apache/lucene/codecs/lucene104/Lucene104ScalarQuantizedVectorsWriter.java

@@ -529,7 +529,9 @@ static int mergeAndRecalculateCentroids(
         mergedCentroid[j] += centroid[j] * vectorCount;
       }
     }
-    if (recalculate) {
+    if (totalVectorCount == 0) {
+      return 0;
+    } else if (recalculate) {
       return calculateCentroid(mergeState, fieldInfo, mergedCentroid);
     } else {
       for (int j = 0; j < mergedCentroid.length; j++) {

lucene/core/src/java/org/apache/lucene/util/VectorUtil.java

@@ -65,7 +65,14 @@ public static float dotProduct(float[] a, float[] b) {
       throw new IllegalArgumentException("vector dimensions differ: " + a.length + "!=" + b.length);
     }
     float r = IMPL.dotProduct(a, b);
-    assert Float.isFinite(r);
+    assert Float.isFinite(r)
+        : "not finite: "
+            + r
+            + " from <"
+            + java.util.Arrays.toString(a)
+            + ","
+            + java.util.Arrays.toString(b)
+            + ">";
     return r;
   }