diff --git a/klausur-service/backend/cv_vocab_pipeline.py b/klausur-service/backend/cv_vocab_pipeline.py
index 1c0012e..837fcc6 100644
--- a/klausur-service/backend/cv_vocab_pipeline.py
+++ b/klausur-service/backend/cv_vocab_pipeline.py
@@ -4885,12 +4885,22 @@ def build_cell_grid_v2(
         logger.warning("build_cell_grid_v2: no usable columns found")
         return [], []
 
-    # Heal row gaps
-    _heal_row_gaps(
-        content_rows,
-        top_bound=min(c.y for c in relevant_cols),
-        bottom_bound=max(c.y + c.height for c in relevant_cols),
-    )
+    # Heal row gaps — use header/footer boundaries (NOT column bounds!)
+    # In Cell-First OCR, the crop IS the OCR input, so extending into
+    # header/footer means OCR'ing header/footer text ("VOCABULARY", page nums).
+    content_rows.sort(key=lambda r: r.y)
+    header_rows = [r for r in row_geometries if r.row_type == 'header']
+    footer_rows = [r for r in row_geometries if r.row_type == 'footer']
+    if header_rows:
+        top_bound = max(r.y + r.height for r in header_rows)
+    else:
+        top_bound = content_rows[0].y
+    if footer_rows:
+        bottom_bound = min(r.y for r in footer_rows)
+    else:
+        bottom_bound = content_rows[-1].y + content_rows[-1].height
+
+    _heal_row_gaps(content_rows, top_bound=top_bound, bottom_bound=bottom_bound)
 
     relevant_cols.sort(key=lambda c: c.x)
 
@@ -5006,11 +5016,20 @@ def build_cell_grid_v2_streaming(
     if not content_rows:
         return
 
-    _heal_row_gaps(
-        content_rows,
-        top_bound=min(c.y for c in relevant_cols),
-        bottom_bound=max(c.y + c.height for c in relevant_cols),
-    )
+    # Use header/footer boundaries for heal_row_gaps (same as build_cell_grid_v2)
+    content_rows.sort(key=lambda r: r.y)
+    header_rows = [r for r in row_geometries if r.row_type == 'header']
+    footer_rows = [r for r in row_geometries if r.row_type == 'footer']
+    if header_rows:
+        top_bound = max(r.y + r.height for r in header_rows)
+    else:
+        top_bound = content_rows[0].y
+    if footer_rows:
+        bottom_bound = min(r.y for r in footer_rows)
+    else:
+        bottom_bound = content_rows[-1].y + content_rows[-1].height
+
+    _heal_row_gaps(content_rows, top_bound=top_bound, bottom_bound=bottom_bound)
 
     relevant_cols.sort(key=lambda c: c.x)