diff --git a/lib/expressir/express/remark_attacher.rb b/lib/expressir/express/remark_attacher.rb
index a5934192..4e62c321 100644
--- a/lib/expressir/express/remark_attacher.rb
+++ b/lib/expressir/express/remark_attacher.rb
@@ -11,6 +11,17 @@ module Express
     # 2. Proximity-based matching for simple tags
     # 3. NOT creating spurious schema-level items for ambiguous tags
     class RemarkAttacher
+      # Child collection attributes to walk when traversing model elements.
+      # These are the named collections on model elements (e.g., entity.attributes,
+      # schema.entities, function.parameters). Extracted to a constant to avoid
+      # duplication between calculate_children_end_line and collect_children.
+      CHILD_COLLECTION_ATTRIBUTES = %i[
+        schemas types entities functions procedures rules constants
+        attributes derived_attributes inverse_attributes
+        where_rules unique_rules informal_propositions
+        parameters variables statements items remark_items
+      ].freeze
+
       def initialize(source)
         @source = source
         @attached_spans = Set.new
@@ -23,8 +34,14 @@ def initialize(source)
       def attach(model)
         @model = model
         remarks = extract_all_remarks
-        attach_tagged_remarks(model, remarks)
-        attach_untagged_remarks(model, remarks)
+
+        # Build nodes_with_positions ONCE for both tagged and untagged remark passes.
+        # This avoids double tree walk (381K nodes × 2 = 762K visits) which was
+        # the largest memory overhead in remark attachment (~430MB for large files).
+        nodes_with_positions = build_sorted_nodes_with_positions(model)
+
+        attach_tagged_remarks(model, remarks, nodes_with_positions)
+        attach_untagged_remarks(remarks, nodes_with_positions)
 
         # Free expensive data structures after attachment is complete.
         # These are only needed during the attach process.
@@ -133,7 +150,7 @@ def get_line_number(position)
 
       alias get_line_number_from_offset get_line_number
 
-      def attach_tagged_remarks(model, remarks)
+      def attach_tagged_remarks(model, remarks, nodes_with_positions)
         tagged = remarks.select { |r| r[:tag] }
         return if tagged.empty?
 
@@ -143,12 +160,6 @@ def attach_tagged_remarks(model, remarks)
         # This is the key optimization that makes scope lookup O(1) per remark
         @scope_map ||= build_scope_map
 
-        # Collect nodes with positions for position-based fallback
-        nodes_with_positions = []
-        collect_nodes_with_positions(model, nodes_with_positions)
-        # Use stable sort to ensure deterministic ordering across Ruby versions
-        nodes_with_positions.sort_by!.with_index { |n, i| [n[:position] || Float::INFINITY, i] }
-
         tagged.sort_by { |r| r[:position] }.each do |remark|
           next if @attached_spans.include?(remark[:position])
 
@@ -917,15 +928,10 @@ def create_remark_item(parent, id)
         item
       end
 
-      def attach_untagged_remarks(model, remarks)
+      def attach_untagged_remarks(remarks, nodes_with_positions)
         untagged = remarks.reject { |r| r[:tag] }
         return unless untagged.any?
 
-        nodes_with_positions = []
-        collect_nodes_with_positions(model, nodes_with_positions)
-        # Use stable sort to preserve original order for equal keys
-        nodes_with_positions.sort_by!.with_index { |n, i| [n[:position] || Float::INFINITY, i] }
-
         untagged.each do |remark|
           next if @attached_spans.include?(remark[:position])
 
@@ -1092,10 +1098,7 @@ def calculate_children_end_line(node)
         end
 
         # Check specific child collections
-        %i[schemas types entities functions procedures rules constants
-           attributes derived_attributes inverse_attributes
-           where_rules unique_rules informal_propositions
-           parameters variables statements items remark_items].each do |attr|
+        CHILD_COLLECTION_ATTRIBUTES.each do |attr|
           collection = safe_get_collection(node, attr)
           collection&.each do |child|
             if child.is_a?(Model::ModelElement) && child.source_offset && child.source
@@ -1114,10 +1117,7 @@ def collect_children(node, result, visited)
           collect_nodes_with_positions(child, result, visited)
         end
 
-        %i[schemas types entities functions procedures rules constants
-           attributes derived_attributes inverse_attributes
-           where_rules unique_rules informal_propositions
-           parameters variables statements items remark_items].each do |attr|
+        CHILD_COLLECTION_ATTRIBUTES.each do |attr|
           collection = safe_get_collection(node, attr)
           collection&.each do |item|
             collect_nodes_with_positions(item, result, visited)
@@ -1125,6 +1125,16 @@ def collect_children(node, result, visited)
         end
       end
 
+      # Build sorted nodes_with_positions ONCE for both tagged and untagged remark passes.
+      # This merges the two separate tree walks into one, cutting node visits in half.
+      def build_sorted_nodes_with_positions(model)
+        nodes_with_positions = []
+        collect_nodes_with_positions(model, nodes_with_positions)
+        # Stable sort: nil positions last, ties broken by insertion order
+        nodes_with_positions.sort_by!.with_index { |n, i| [n[:position] || Float::INFINITY, i] }
+        nodes_with_positions
+      end
+
       def find_nearest_node(remark, nodes)
         remark_line = remark[:line]