fix(eval_optimize_loop): address AI review round 12 — 1 critical + 3 warning + 1 suggestion

Critical:
- run_pipeline.py: _sys.stderr -> sys.stderr in _read_critical_case_ids
  except branch (NameError would crash pipeline on evalset read failure)

Warning:
- run_pipeline.py: add FileExistsError to lock rebuild except clause
  to handle race between stale-lock removal and O_EXCL recreation
- gate.py: baseline_cost <= 0 now returns explicit skip message instead
  of silently auto-passing cost gate (fake mode has simulated costs)
- baseline.py: remove unused import time

Suggestion:
- run_pipeline.py: document overfit_detection caveat in fake mode
  (flat +0.05 delta makes train_improved always true)

99 tests pass, pipeline 6 phases run end-to-end.

Author: woshidage77

examples/optimization/eval_optimize_loop/run_pipeline.py

@@ -35,7 +35,7 @@ def _read_critical_case_ids(val_path: Path) -> list[str]:
             data = json.load(f)
         return [c["case_id"] for c in data.get("cases", []) if c.get("critical", False)]
     except Exception as e:
-        print(f"Warning: cannot read critical case ids from {val_path}: {e}", file=_sys.stderr)
+        print(f"Warning: cannot read critical case ids from {val_path}: {e}", file=sys.stderr)
         return []  # empty: skip critical-case gate rather than guess wrong id
 
 
@@ -122,7 +122,7 @@ def _pid_alive(pid):
                 lf.flush()
                 _os.fsync(lf.fileno())
             acquired = True
-        except (FileNotFoundError, ValueError):
+        except (FileNotFoundError, ValueError, FileExistsError):
             pass
 
     if not acquired:
@@ -174,6 +174,12 @@ def _pid_alive(pid):
         # are simulated placeholder values. Gate decisions in fake mode are for
         # pipeline demo purposes only and do not reflect real optimization outcomes.
         # Real mode would re-evaluate with the optimized agent on the training set.
+        #
+        # Overfit detection caveat: candidate_train_scores uses a flat +0.05 delta
+        # which makes train_improved always true in fake mode.  This means
+        # overfit_detection degenerates to "reject iff val regresses" and is NOT
+        # a genuine overfit signal.  In real mode, train scores come from actual
+        # re-evaluation with the optimized prompt.
         # overfit detection: simulate candidate train scores with +0.05 delta
         candidate_train_scores = {
             cid: min(1.0, score + 0.05) for cid, score in train_bl.score_map.items()

examples/optimization/eval_optimize_loop/src/baseline.py

@@ -16,7 +16,6 @@
 from __future__ import annotations
 
 import json
-import time
 from dataclasses import dataclass, field
 import hashlib
 from pathlib import Path

examples/optimization/eval_optimize_loop/src/gate.py

@@ -196,8 +196,15 @@ def _check_cost(
     ) -> GateCheck:
         max_ratio = self.rules["cost_within_budget"].get("max_cost_ratio", 1.2)
         if baseline_cost <= 0:
-            passed = True
-            ratio = 1.0
+            # Fake mode: costs are simulated placeholders (see run_pipeline.py L173-176).
+            # Real mode: cost may be 0 if token_tracker is not connected.
+            # In either case, cost data is absent; skip this gate rather than auto-pass.
+            return GateCheck(
+                name="cost_within_budget",
+                passed=True,
+                description=f"cost budget skipped (baseline_cost={baseline_cost:.4f} <= 0, tracking inactive)",
+                detail="cost data unavailable; gate skipped",
+            )
         else:
             ratio = candidate_cost / baseline_cost
             passed = ratio <= max_ratio