fix(eval_optimize_loop): address AI review round 8 — 1 critical + 7 warning

Critical:
- fake_judge: clamp response_quality to [0.2, 1.0] (was recognition*1.05
  could exceed 1.0, causing score overflow in gate and audit reports)

Warnings:
- PID lock: atomic acquire via os.O_CREAT|O_EXCL (truly atomic cross-platform,
  replaces read-check-replace TOCTOU pattern)
- auditor run_id: add microseconds (%f) to prevent same-second collisions
- gate critical_case: treat missing-from-candidate as regression
- call_agent: dedup sys.path.insert (only insert if path not present)
- baseline _run_real_split: add sys.path cleanup note in docstring
- candidate_train_scores simulated + image_id mapping: already documented
- Chinese comments: acknowledged, no functional change

Author: woshidage77

examples/optimization/eval_optimize_loop/fake/fake_judge.py

@@ -69,7 +69,7 @@ def evaluate(
         recognition = self._char_match_score(ground_truth, predicted)
         # 黑名单和回复质量随识别质量缩放（模拟真实场景）
         blacklist = max(0.1, recognition * 0.9)
-        response = max(0.2, recognition * 1.05)
+        response = min(1.0, max(0.2, recognition * 1.05))
 
         score = JudgeScore(
             recognition_quality=recognition,

examples/optimization/eval_optimize_loop/run_pipeline.py

@@ -96,24 +96,39 @@ def _pid_alive(pid):
         return True
 
     my_pid = _os.getpid()
+
+    # Atomic acquire: O_CREAT|O_EXCL fails if file exists (cross-platform)
+    acquired = False
     try:
-        with open(LOCK_FILE, "r", encoding="utf-8") as lf:
-            old_pid = int(lf.read().strip().split()[0])
-        if _pid_alive(old_pid):
-            print("another pipeline instance is running, aborting", file=sys.stderr)
-            sys.exit(75)
-        if not args.quiet:
-            print(f"Cleaning stale lock from dead PID {old_pid}", file=sys.stderr)
-    except (FileNotFoundError, ValueError):
-        pass
-
-    # atomic write: temp file + os.replace to avoid partial writes on crash
-    tmp_file = LOCK_FILE + ".tmp"
-    with open(tmp_file, "w", encoding="utf-8") as lf:
-        lf.write(f"{my_pid} {started_at}")
-        lf.flush()
-        _os.fsync(lf.fileno())
-    _os.replace(tmp_file, LOCK_FILE)
+        fd = _os.open(LOCK_FILE, _os.O_CREAT | _os.O_EXCL | _os.O_WRONLY, 0o644)
+        with _os.fdopen(fd, "w", encoding="utf-8") as lf:
+            lf.write(f"{my_pid} {started_at}")
+            lf.flush()
+            _os.fsync(lf.fileno())
+        acquired = True
+    except FileExistsError:
+        # Lock exists -- check if owner is alive
+        try:
+            with open(LOCK_FILE, "r", encoding="utf-8") as lf:
+                old_pid = int(lf.read().strip().split()[0])
+            if _pid_alive(old_pid):
+                print("another pipeline instance is running, aborting", file=sys.stderr)
+                sys.exit(75)
+            if not args.quiet:
+                print(f"Cleaning stale lock from dead PID {old_pid}", file=sys.stderr)
+            _os.remove(LOCK_FILE)
+            fd = _os.open(LOCK_FILE, _os.O_CREAT | _os.O_EXCL | _os.O_WRONLY, 0o644)
+            with _os.fdopen(fd, "w", encoding="utf-8") as lf:
+                lf.write(f"{my_pid} {started_at}")
+                lf.flush()
+                _os.fsync(lf.fileno())
+            acquired = True
+        except (FileNotFoundError, ValueError):
+            pass
+
+    if not acquired:
+        print("cannot acquire pipeline lock, aborting", file=sys.stderr)
+        sys.exit(75)
 
     # ---- try/finally: ensure lock release even on exception ----
     try:

examples/optimization/eval_optimize_loop/src/auditor.py

@@ -57,7 +57,7 @@ def save(self, audit_trail, baseline, attribution, optimization, validation=None
 
     def build_trail(self, pipeline_name, mode, random_seed, optimization, baseline_val, validation=None, gate_decision=None, started_at=""):
         now = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
-        run_id = datetime.now().strftime("%Y%m%d_%H%M%S") + f"_{random_seed}"
+        run_id = datetime.now().strftime("%Y%m%d_%H%M%S_%f") + f"_{random_seed}"
         entries = []
         for cand in optimization.candidates:
             entry = AuditEntry(

examples/optimization/eval_optimize_loop/src/gate.py

@@ -178,12 +178,13 @@ def _check_critical_cases(
             if cid in baseline and cid in candidate
             and candidate[cid] < baseline[cid]
         ]
-        passed = len(regressed) == 0
+        missing = [cid for cid in critical_ids if cid not in candidate]
+        passed = len(regressed) == 0 and len(missing) == 0
         return GateCheck(
             name="critical_case_no_regress",
             passed=passed,
-            description="关键 case 不退步",
-            detail=f"regressed: {regressed}" if regressed else "all critical cases stable",
+            description="关键 case 不退步且不丢失",
+            detail=f"regressed: {regressed}; missing: {missing}" if (regressed or missing) else "all critical cases stable",
         )
 
     def _check_cost(