baseline: setup failure is not recorded as a result; repair_stats tolerates a missing trajectory

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat

harness/runner.py

@@ -96,6 +96,8 @@ def repair_stats(path):
     import hashlib
     last = None  # (object, md5, failed)
     after = changed = 0
+    if not os.path.exists(path):  # setup failed: no agent run, no trajectory
+        return {"pushes_after_error": None, "pushes_changed_after_error": None, "repair_rate": None}
     for line in open(path):
         e = json.loads(line)
         if e.get("tool") != "sap_push_source" or (e["result"].startswith("Tool ") and e["is_error"]):

train/baseline.py

@@ -52,6 +52,10 @@ def main():
             json.dump(res, open(out_path, "w"), indent=1)
             print(json.dumps({"task": tid, "error": str(e)[:300]}), flush=True)
             continue
+        if rep.get("setup_failed"):  # no agent run; do not record, so a restart tries the task again
+            bad = [x for x in rep.get("setup", []) if not x.get("ok")]
+            print(json.dumps({"task": tid, "setup_failed": True, "details": bad, "run_dir": run_dir}, default=str)[:1500], flush=True)
+            continue
         h = rep.get("hidden_tests") or {}
         res["tasks"][tid] = {"category": t["category"], "object_type": t["object_type"],
                              "score": (rep.get("score") or {}).get("total"), "parts": rep.get("score"),