diff --git a/arborist/cold_pack_metadata.py b/arborist/cold_pack_metadata.py
index e8017f6..3954a7a 100644
--- a/arborist/cold_pack_metadata.py
+++ b/arborist/cold_pack_metadata.py
@@ -470,7 +470,19 @@ def _restore_generic_table(
     table: str,
     in_path: Path,
 ) -> int:
-    """Read array-per-line JSONL → INSERT OR IGNORE batches into `table`."""
+    """Read array-per-line JSONL → INSERT OR IGNORE batches into `table`.
+
+    Chunks special case (#53): the producer dumps a synthetic
+    ``_content_size`` column carrying each chunk's content byte-length
+    (the actual ``content`` BLOB is dropped from the metadata pack to
+    keep it small). Consumer drops the synthetic column from the INSERT
+    column list — chunks rows land with ``content IS NULL`` per
+    ``pull_metadata_pack``'s documented contract; phase 2
+    (``pull_chunk_pack``) fills the bytes via UPDATE WHERE leaf_hash=?.
+    The size hint is currently unused in the single-shard path; a
+    future zeroblob pre-allocation could read it back if page-split
+    cost becomes measurable.
+    """
     BATCH = 5000
     batch: list[tuple] = []
     cols: list[str] | None = None
@@ -478,7 +490,7 @@ def _restore_generic_table(
     insert_sql: str | None = None
     for row in read_columnar_jsonl(in_path):
         if cols is None:
-            cols = list(row.keys())
+            cols = [c for c in row.keys() if c != "_content_size"]
             placeholders = ", ".join("?" for _ in cols)
             col_list = ", ".join(f'"{c}"' for c in cols)
             insert_sql = (