diff --git a/image_analysis/scale_merge/RotationScaleMerge.cpp b/image_analysis/scale_merge/RotationScaleMerge.cpp
index be65e5678..5312363a9 100644
--- a/image_analysis/scale_merge/RotationScaleMerge.cpp
+++ b/image_analysis/scale_merge/RotationScaleMerge.cpp
@@ -6476,7 +6476,13 @@ RotationScaleMerge::Result RotationScaleMerge::RunArm(bool scale_partials, bool
 
         // resize, not assign: the gather below writes every field of every full, so zeroing 80 bytes
         // each first is a few hundred megabytes of pure write traffic per chain. The staging arrays are
-        // members for the same reason.
+        // members for the same reason. Nothing in them survives the gather either, so when they have
+        // to grow they are freed first: growing in place would copy the previous chain's records into
+        // the new allocation only for the gather to overwrite them.
+        if (static_cast<size_t>(nf) > fulls.capacity()) {
+            fulls = {};
+            fulls_staging = {};
+        }
         fulls.resize(nf);
         FullsStaging &st = fulls_staging;
         st.Resize(nf);