fix(eval_optimize_loop): address AI review round 9 — 2 critical + 6 warning

Critical:
- optimizer: cumulative multi-iteration (each iteration now uses previous
  prompt_after as prompt_before, not the original BASE_PROMPTS every time)
- baseline real mode image_id hash: ACK, docstring already marks PLACEHOLDER

Warnings:
- run_pipeline: remove unused import time and redundant sys import
- auditor: add None guard in save() baseline dict comprehension
- gate/attribution: extract PASS_THRESHOLD constant (0.6) from fake_judge,
  share across gate._check_no_new_hard_fail and attribution judge checks
- _pid_alive ctypes: already fixed in round 6, no change needed
- candidate_train_scores simulated: already documented

Suggestion:
- validator: warn on unknown failure_category fallback in CANDIDATE_PREDICTIONS

Author: woshidage77

examples/optimization/eval_optimize_loop/fake/fake_judge.py

@@ -37,6 +37,10 @@ class JudgeResult:
     failure_reason: str = ""
 
 
+# Shared threshold: gate _check_no_new_hard_fail and attribution judge checks
+# use the same value as FakeJudge.passed threshold.
+PASS_THRESHOLD = 0.6
+
 class FakeJudge:
     """基于规则的假 Judge。
 

examples/optimization/eval_optimize_loop/run_pipeline.py

@@ -7,7 +7,7 @@
     python run_pipeline.py --quiet            # minimal output
 """
 
-import argparse, asyncio, json, os as _os, sys, time
+import argparse, asyncio, json, os as _os, sys
 from pathlib import Path
 from datetime import datetime, timezone
 
@@ -35,7 +35,6 @@ 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:
-        import sys as _sys
         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
 

examples/optimization/eval_optimize_loop/src/attribution.py

@@ -12,6 +12,7 @@
 from typing import Optional
 
 from src.baseline import BaselineResult, BaselineCaseResult
+from fake.fake_judge import PASS_THRESHOLD
 
 
 @dataclass
@@ -189,13 +190,13 @@ def _attribute_case(
                 evidence.append(f"human_review with low conf={conf_val}")
 
         # Rule 3: Judge scores
-        if case.judge_recognition >= 0 and case.judge_recognition < 0.6:
+        if case.judge_recognition >= 0 and case.judge_recognition < PASS_THRESHOLD:
             candidates.append(("llm_rubric_fail", 0.80))
             evidence.append(f"judge_recognition={case.judge_recognition:.2f} < 0.6")
-        if case.judge_blacklist >= 0 and case.judge_blacklist < 0.6:
+        if case.judge_blacklist >= 0 and case.judge_blacklist < PASS_THRESHOLD:
             candidates.append(("knowledge_recall_insufficient", 0.75))
             evidence.append(f"judge_blacklist={case.judge_blacklist:.2f} < 0.6")
-        if case.judge_response >= 0 and case.judge_response < 0.6:
+        if case.judge_response >= 0 and case.judge_response < PASS_THRESHOLD:
             candidates.append(("llm_rubric_fail", 0.65))
             evidence.append(f"judge_response={case.judge_response:.2f} < 0.6")
 

examples/optimization/eval_optimize_loop/src/auditor.py

@@ -39,7 +39,7 @@ def save(self, audit_trail, baseline, attribution, optimization, validation=None
         ts_dir = audit_trail.run_id
         audit_path = self.output_dir / "audit" / ts_dir
         audit_path.mkdir(parents=True, exist_ok=True)
-        full = {"audit_trail":audit_trail.to_dict(),"baseline":{k:v.to_dict() for k,v in baseline.items()},"attribution":attribution.to_dict(),"optimization":optimization.to_dict()}
+        full = {"audit_trail":audit_trail.to_dict(),"baseline":{k:v.to_dict() if v else {} for k,v in baseline.items()}},"attribution":attribution.to_dict(),"optimization":optimization.to_dict()}
         if validation: full["validation"] = validation.to_dict()
         if gate_decision: full["gate_decision"] = gate_decision
         with open(audit_path/"optimization_report.json","w",encoding="utf-8") as f:

examples/optimization/eval_optimize_loop/src/gate.py

@@ -8,6 +8,8 @@
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Optional
+from fake.fake_judge import PASS_THRESHOLD
+
 
 
 @dataclass
@@ -149,8 +151,8 @@ def _check_no_new_hard_fail(
         candidate: dict[str, float],
     ) -> GateCheck:
         max_new = self.rules["no_new_hard_fail"].get("max_new_fails", 0)
-        base_fails = sum(1 for s in baseline.values() if s < 0.6)
-        cand_fails = sum(1 for s in candidate.values() if s < 0.6)
+        base_fails = sum(1 for s in baseline.values() if s < PASS_THRESHOLD)
+        cand_fails = sum(1 for s in candidate.values() if s < PASS_THRESHOLD)
         new_fails = max(0, cand_fails - base_fails)
         passed = new_fails <= max_new
         return GateCheck(

examples/optimization/eval_optimize_loop/src/optimizer.py

@@ -224,10 +224,13 @@ def optimize(
             prompt_type = target["prompt_target"]
             confidence = target["confidence"]
 
-            # ???????
-            prompt_before = self._get_base_prompt(prompt_type)
+            # prompt_before: use previous iteration's result for cumulative optimization
+            if candidates:
+                prompt_before = candidates[-1].prompt_after
+            else:
+                prompt_before = self._get_base_prompt(prompt_type)
 
-            # ???????
+            # generate optimization append
             prompt_after, change_log = self._generate_optimization(
                 prompt_type, category, prompt_before, confidence
             )

examples/optimization/eval_optimize_loop/src/validator.py

@@ -68,6 +68,9 @@ def run(self, val_baseline, optimization_result, simulate_regression=False):
         return self._run_real(val_baseline, candidate)
 
     def _run_fake(self, val_baseline, candidate, simulate_regression=False):
+        if candidate.failure_category not in CANDIDATE_PREDICTIONS:
+            import warnings
+            warnings.warn(f"Unknown failure_category '{candidate.failure_category}', falling back to final_answer_mismatch")
         pred_map = REGRESSION_PREDICTIONS if simulate_regression else CANDIDATE_PREDICTIONS.get(
             candidate.failure_category, CANDIDATE_PREDICTIONS["final_answer_mismatch"])
         deltas = []