run_benchmarks: publish the effective shot count with its provenance, so a 0-shot primary row can never be read as few-shot

eval/run_benchmarks.py

@@ -329,7 +329,15 @@ def main():
             rc, cmd = run_task(a.model, t, shots, d, 0, a.full_gpu_hours)
             h = harvest(d, t)
             rec[key] = {"rc": rc, "rows": h.get("rows", {}), "command": cmd,
-                        "shots": h.get("shots"), "harness_fewshot": h.get("config_fewshot"),
+                        "shots": (h.get("shots") if h.get("shots") is not None else 0),
+                        # Six of the nine YAMLs declare no num_fewshot (docs/05 §7, re-confirmed by E-049),
+                        # and the PRIMARY column deliberately passes no flag, so those six rows are
+                        # 0-shot. Recording `shots: 0` with its reason is what keeps a 0-shot number from
+                        # being read as the few-shot figure D-005 warned about -- the design of the columns
+                        # is frozen, the description of what happened is not allowed to be vague.
+                        "shots_source": ("harness config" if h.get("shots") is not None
+                                         else "harness default, none declared, no flag passed"),
+                        "harness_fewshot": h.get("config_fewshot"),
                         "n_input": h.get("n_input"), "n_effective": h.get("n_effective"),
                         "n_logged": h.get("n_logged"),
                         "limit": h.get("limit"), "model": h.get("model"), "dtype": h.get("dtype"),