feat(evaluation): 成稿 100 分 rubric、证据评审与评测台调整

- 回忆录细项上限收紧为合计 100 分，去掉 110 折算与 raw_dimension_total
- judge_memoir 拼接原始访谈与可选导出基线；无证据时提示保守打真实性相关分
- 自动评测 run 与手动章节/故事评审统一带 transcript 证据（会话/用户聚合、截断）
- 访谈打分仍为情绪强化版 15 细项、总分 100
- 评测台默认基准改为 zuckxu 导出 MD；移除逐轮用户句对齐表及相关逻辑
- 新增 judge schema 与 memoir prompt 组装的单元测试

api/app/features/evaluation/judge_service.py

@@ -26,6 +26,37 @@ _CONV_MAX = 8192
 _CONV_JUDGE_JSON_MAX = 2048
 _MEMOIR_MAX = 12000
 _COMPARE_STREAM_MAX = 6144
+_MEMOIR_EVIDENCE_MAX = 12000
+
+
+def _build_memoir_judge_prompt(
+    *,
+    memoir_markdown: str,
+    source_transcript: str = "",
+    reference_memoir_markdown: str = "",
+    evidence_notes: str = "",
+) -> str:
+    """Assemble an evidence-aware memoir judging prompt."""
+    source = (source_transcript or "").strip()
+    reference = (reference_memoir_markdown or "").strip()
+    notes = (evidence_notes or "").strip()
+    sections = [MEMOIR_JUDGE_INSTRUCTIONS, ""]
+    if notes:
+        sections.extend(["【评审说明】", notes[:1200], ""])
+    if source:
+        sections.extend(["【原始访谈/证据】", source[:_MEMOIR_EVIDENCE_MAX], ""])
+    else:
+        sections.extend(
+            [
+                "【原始访谈/证据】",
+                "无可用原始访谈证据。对于记忆忠实度、事实准确性、事实覆盖率、记忆可追溯性，必须保守打分，不得凭空高分。",
+                "",
+            ]
+        )
+    if reference:
+        sections.extend(["【参考基线/导出成稿】", reference[:_MEMOIR_EVIDENCE_MAX], ""])
+    sections.extend(["【当前回忆录正文】", memoir_markdown[:_MEMOIR_MAX]])
+    return "\n".join(sections)
 
 
 class EvalJudgeService:
@@ -124,7 +155,7 @@ class EvalJudgeService:
 {r_json}
 
 请依次撰写：
-1) 两段对话在整体体验上的主要差异（共情、追问、重复感、自然度等）；
+1) 两段对话在整体体验上的主要差异（情绪承接、信息挖掘、人物建模、访谈结构、提问质量、上下文与重复盘问等）；
 2) B 相对 A 的优点与不足；
 3) 若 B 在关键维度明显弱于 A，给出可操作的改进方向（系统提示、访谈策略、模型或温度等）。
 
@@ -154,14 +185,22 @@ class EvalJudgeService:
             logger.warning("conversation compare stream failed: {}", e)
             yield f"\n\n[流式输出中断：{e}]"
 
-    async def judge_memoir(self, *, memoir_markdown: str) -> MemoirJudgeOutput | None:
+    async def judge_memoir(
+        self,
+        *,
+        memoir_markdown: str,
+        source_transcript: str = "",
+        reference_memoir_markdown: str = "",
+        evidence_notes: str = "",
+    ) -> MemoirJudgeOutput | None:
         if not self._llm:
             return None
-        prompt = f"""{MEMOIR_JUDGE_INSTRUCTIONS}
-
-【回忆录正文】
-{memoir_markdown[:_MEMOIR_MAX]}
-"""
+        prompt = _build_memoir_judge_prompt(
+            memoir_markdown=memoir_markdown,
+            source_transcript=source_transcript,
+            reference_memoir_markdown=reference_memoir_markdown,
+            evidence_notes=evidence_notes,
+        )
         try:
             return await allm_json_call(
                 self._llm,