fix: copy per-chunk buffers and final tcm to avoid iterator-buffer reuse

src/pygama/evt/tcm.py

@@ -199,7 +199,10 @@ def _merge_sorted_tcms(a: ak.Array, b: ak.Array, coin_windows_local) -> ak.Array
                 row_in_table = np.arange(start, start + buf_len, dtype=int)
             buf_ak = ak.with_field(buf_ak, row_in_table, "row_in_table")
 
-            arrays.append(buf_ak)
+            # LH5Iterator reuses the same underlying buffer across iterations.
+            # Make an explicit copy so concatenation/merging isn't corrupted when the
+            # iterator advances and overwrites the buffer.
+            arrays.append(ak.copy(buf_ak))
 
         if at_end.all() and len(arrays) == 0 and tcm is None:
             break
@@ -214,6 +217,9 @@ def _merge_sorted_tcms(a: ak.Array, b: ak.Array, coin_windows_local) -> ak.Array
         else:
             tcm = _merge_sorted_tcms(tcm, new_tcm, coin_windows)
 
+        # Ensure tcm doesn't retain views into iterator-reused buffers.
+        tcm = ak.copy(tcm)
+
         if tcm is None or len(tcm) == 0:
             continue
 

tests/evt/test_build_tcm.py

@@ -378,3 +378,30 @@ def test_build_tcm_multiple_files(lgnd_test_data, tmp_dir):
 
     assert tcm_orig.fields == tcm.fields
     assert all(ak.all(tcm_orig[f] == tcm[f]) for f in tcm.fields)
+
+
+def test_build_tcm_buffer_reuse_copy_regression(lgnd_test_data):
+    """
+    Regression test for LH5Iterator's buffer reuse: build_tcm must not retain
+    views into per-chunk buffers across iterations.
+    """
+    f_raw = lgnd_test_data.get_path(
+        "lh5/prod-ref-l200/generated/tier/raw/cal/p03/r001/l200-p03-r001-cal-20230318T012144Z-tier_raw.lh5"
+    )
+
+    # small buffer_len forces multiple iterator iterations, which used to expose
+    # corruption if per-chunk buffers weren't copied before concatenation/merge.
+    tcm_small = evt.build_tcm(
+        [(f_raw, ["ch1084803/raw", "ch1084804/raw", "ch1121600/raw"])],
+        coin_cols="timestamp",
+        buffer_len=1,
+    ).view_as("ak")
+
+    tcm_large = evt.build_tcm(
+        [(f_raw, ["ch1084803/raw", "ch1084804/raw", "ch1121600/raw"])],
+        coin_cols="timestamp",
+        buffer_len=1_000_000,
+    ).view_as("ak")
+
+    assert tcm_small.fields == tcm_large.fields
+    assert all(ak.all(tcm_small[f] == tcm_large[f]) for f in tcm_small.fields)