scripts: golden-prompt runner — tolerate error_max_turns in capped routing runs

The 3-turn cap is the routing mode's cost control; hitting it is expected
once a skill starts executing roles, and activation is already visible in
the transcript. Only genuine run failures (rate limit, API error) are now
reported as errors. With this fix the full routing table passes 9/9 live.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: geledek

scripts/run_golden_prompts.py

@@ -104,8 +104,11 @@ def run_claude(prompt: str, model: str | None, max_turns: int | None, timeout: i
                     activated.add(block.get("input", {}).get("skill", "").split(":")[-1])
         elif event.get("type") == "result":
             result_text = event.get("result") or ""
-            if event.get("is_error") or event.get("subtype") != "success":
-                # A failed run must not masquerade as "skill stayed dormant".
+            # error_max_turns is expected in capped routing runs — activation is
+            # already visible in the transcript. Any other failure must not
+            # masquerade as "skill stayed dormant".
+            if (event.get("is_error") or event.get("subtype") != "success") \
+                    and event.get("subtype") != "error_max_turns":
                 result_error = f"{event.get('subtype')}: {result_text[:200]}"
     return activated, result_text, result_error