[agent:cc] benchmarks: allow null tok_s_prefill/power_w for concurrency>1 rows

validate.py

@@ -114,8 +114,14 @@ def row_errors(row: dict[str, Any], line_no: int) -> list[str]:
         if value not in allowed:
             errors.append(f"line {line_no}: {field}={value!r} not in {sorted(allowed)}")
 
+    # tok_s_prefill / power_w may be null (JSON null) on concurrency>1 rows where they are
+    # NOT separately measured - the row's headline is tok_s_decode (the aggregate). A null is
+    # the honest sentinel there; an arbitrary borrowed single-stream value would be misleading.
+    nullable_numeric = {"tok_s_prefill", "power_w"}
     for field in NUMERIC:
         value = row.get(field)
+        if value is None and field in nullable_numeric:
+            continue
         if not isinstance(value, (int, float)) or isinstance(value, bool):
             errors.append(f"line {line_no}: {field} must be numeric")
         elif value < 0: