Validate bulk binary sizes as integral metadata

Reject boolean and floating-point values for int64 bulk binary sizes instead of
silently converting them with int(). Keep integer strings accepted for existing
NVBench JSON compatibility, and add regression coverage for valid and malformed
size payloads.

python/scripts/nvbench_compare.py

@@ -845,6 +845,16 @@ def extract_binary_filename(summary):
 
 def extract_binary_size(summary):
     value = extract_summary_data_value(summary, "size", "int64")
+    if isinstance(value, bool):
+        raise ValueError(
+            f"summary {summary.get('tag', '<unknown>')!r} field 'size' "
+            f"value {value!r} is not an int64"
+        )
+    if isinstance(value, float):
+        raise ValueError(
+            f"summary {summary.get('tag', '<unknown>')!r} field 'size' "
+            f"value {value!r} is not an int64"
+        )
     try:
         return int(value)
     except (TypeError, ValueError) as exc:

python/test/test_nvbench_compare.py

@@ -114,6 +114,13 @@ def make_binary_summary(nvbench_compare, tag, filename, size):
     }
 
 
+def make_raw_binary_size_summary(nvbench_compare, value):
+    return {
+        "tag": nvbench_compare.SAMPLE_TIMES_TAG,
+        "data": [{"name": "size", "type": "int64", "value": value}],
+    }
+
+
 def make_reader_for_roots(ref_root, cmp_root):
     def read_file(path):
         if path == "ref.json":
@@ -463,6 +470,24 @@ def test_gpu_timing_data_loads_samples_and_frequencies_lazily(
     assert reader_calls == [str(samples_file), str(freqs_file)]
 
 
+@pytest.mark.parametrize("value", [3, "3"])
+def test_extract_binary_size_accepts_integral_values(value, nvbench_compare):
+    assert (
+        nvbench_compare.extract_binary_size(
+            make_raw_binary_size_summary(nvbench_compare, value)
+        )
+        == 3
+    )
+
+
+@pytest.mark.parametrize("value", [True, False, 3.0, 3.5, "3.5"])
+def test_extract_binary_size_rejects_non_integral_values(value, nvbench_compare):
+    with pytest.raises(ValueError, match="is not an int64"):
+        nvbench_compare.extract_binary_size(
+            make_raw_binary_size_summary(nvbench_compare, value)
+        )
+
+
 def test_compare_benches_collects_bulk_debug_rows(tmp_path, nvbench_compare):
     run_data = make_comparison_run_data(nvbench_compare)
     ref_samples_file = tmp_path / "ref-samples.bin"