diff --git a/daft_lance/lance_merge_column.py b/daft_lance/lance_merge_column.py
index 3086156..8984120 100644
--- a/daft_lance/lance_merge_column.py
+++ b/daft_lance/lance_merge_column.py
@@ -314,8 +314,10 @@ def __call__(self, *cols: Any) -> list[dict[str, bytes]]:
                 writer.write_batch(b)
         file_size = os.path.getsize(filepath)
 
-        # Determine field IDs for the new columns
-        next_fid = max(f.id() for f in self.lance_ds.lance_schema.fields()) + 1
+        # Determine field IDs for the new columns. Lance's manifest-level
+        # max_field_id includes nested child fields and field IDs from dropped
+        # columns, so it is the correct high-water mark for dataset evolution.
+        next_fid = self.lance_ds.max_field_id + 1
 
         # Stitch new data file into fragment metadata
         new_file_entry = {
diff --git a/tests/io/lancedb/test_fast_path_merge.py b/tests/io/lancedb/test_fast_path_merge.py
index 9f69efb..2947ad5 100644
--- a/tests/io/lancedb/test_fast_path_merge.py
+++ b/tests/io/lancedb/test_fast_path_merge.py
@@ -648,3 +648,36 @@ def _make_vec(ids):
             assert len(emb) == N
             for v in emb:
                 assert pytest.approx(float(i), rel=1e-5) == v
+
+    def test_next_fid_uses_manifest_max_field_id(self, ds_path):
+        """Bug: next_fid only scanned top-level lance_schema.fields(), missing child IDs.
+
+        A struct column with M children occupies M+1 Lance field IDs (1 for parent +
+        M for children). With top-level-only scan: max_id = num_top_level_fields - 1,
+        next_fid can collide with a child field ID, and the new file's fragment metadata
+        maps to the wrong schema field → the new column reads back as null.
+
+        Example layout for table with id(0) + meta struct(1, children a=2, b=3):
+          top-level scan       → max=1, next_fid=2 (WRONG: collides with meta.a)
+          ds.max_field_id scan → max=3, next_fid=4 (CORRECT)
+
+        The fix uses Lance's manifest-level max_field_id, which includes child IDs.
+        """
+        struct_type = pa.struct([("a", pa.int64()), ("b", pa.utf8())])
+        rows = [{"a": i, "b": f"s{i}"} for i in [1, 2, 3]]
+        tbl = pa.table(
+            {
+                "id": pa.array([1, 2, 3], type=pa.int64()),
+                "meta": pa.array(rows, type=struct_type),
+            }
+        )
+        lance.write_dataset(tbl, ds_path)
+        ds = lance.dataset(ds_path)
+
+        df = read_with_metadata(ds_path)
+        df = df.with_column("score", daft.col("id").cast(daft.DataType.int64()) * 10)
+        ds2 = merge_columns_from_df(df, ds, ds_path)
+
+        result = ds2.to_table().sort_by("id").to_pydict()
+        assert result["score"] == [10, 20, 30], f"score is null or wrong (field ID collision): {result['score']}"
+        assert result["id"] == [1, 2, 3]