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_manual_service.py

@@ -10,6 +10,7 @@ from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.core.dependencies import get_eval_judge_langchain_llm
 from app.core.logging import get_logger
+from app.features.conversation import repo as conversation_repo
 from app.features.evaluation.errors import (
     EvaluationBadRequestError,
     EvaluationNotFoundError,
@@ -27,6 +28,8 @@ logger = get_logger(__name__)
 _MAX_JUDGE_MARKDOWN_CHARS = 20_000
 _MAX_EVAL_CHAPTERS = 30
 _MAX_EVAL_STORIES = 40
+_MAX_EVIDENCE_CONVERSATIONS = 8
+_MAX_EVIDENCE_TRANSCRIPT_CHARS = 16_000
 
 
 def _clip_md_for_judge(text: str, max_chars: int = _MAX_JUDGE_MARKDOWN_CHARS) -> str:
@@ -48,6 +51,41 @@ def _transcript_from_export_turns(turns: list[tuple[str, str]]) -> str:
     return "\n\n".join(parts)
 
 
+def _trim_evidence_text(text: str, max_chars: int = _MAX_EVIDENCE_TRANSCRIPT_CHARS) -> str:
+    s = (text or "").strip()
+    if len(s) <= max_chars:
+        return s
+    return f"{s[:max_chars]}\n\n…（访谈证据已截断）"
+
+
+async def _conversation_transcript_for_manual(
+    db: AsyncSession, conversation_id: str
+) -> str:
+    rows = await conversation_repo.get_conversation_messages(conversation_id, db)
+    parts: list[str] = []
+    for row in rows:
+        role = (row.role or "").lower()
+        body = (row.content or "").strip()
+        if not body:
+            continue
+        label = "用户" if role == "human" else "AI"
+        out = _assistant_text_for_eval_display(body) if role != "human" else body
+        parts.append(f"{label}: {out}")
+    return "\n\n".join(parts)
+
+
+async def _user_transcript_evidence(db: AsyncSession, user_id: str) -> str:
+    conversations = await conversation_repo.get_user_conversations(user_id, db)
+    if not conversations:
+        return ""
+    parts: list[str] = []
+    for conv in reversed(conversations[:_MAX_EVIDENCE_CONVERSATIONS]):
+        transcript = await _conversation_transcript_for_manual(db, str(conv.id))
+        if transcript:
+            parts.append(f"## 会话 {str(conv.id)}\n{transcript}")
+    return _trim_evidence_text("\n\n".join(parts))
+
+
 def _normalize_title_key(title: str) -> str:
     t = (title or "").strip().lower()
     t = re.sub(r"^#+\s*", "", t)
@@ -271,6 +309,7 @@ class EvalJudgeManualService:
         judge_llm = get_eval_judge_langchain_llm()
         judge = EvalJudgeService(judge_llm)
         baselines = list(baseline_sections or [])
+        evidence_transcript = await _user_transcript_evidence(self._db, uid)
 
         chapter_results: list[dict[str, Any]] = []
         try:
@@ -281,7 +320,7 @@ class EvalJudgeManualService:
                 body = (ch.canonical_markdown or "").strip()
                 if not body:
                     continue
-                bl = _baseline_for_chapter_title(baselines, ch.title or "", i)
+                bl = _baseline_for_chapter_title(baselines, str(ch.title or ""), i)
                 baseline_excerpt = ""
                 if bl and (bl.body or "").strip():
                     baseline_excerpt = _clip_md_for_judge(bl.body, max_chars=6000)
@@ -289,7 +328,14 @@ class EvalJudgeManualService:
                 if baseline_excerpt:
                     md += f"## 导出基线（节选）\n\n{baseline_excerpt}\n\n"
                 md += f"## 当前成稿\n\n{_clip_md_for_judge(body)}"
-                cj = await judge.judge_memoir(memoir_markdown=md)
+                cj = await judge.judge_memoir(
+                    memoir_markdown=md,
+                    source_transcript=evidence_transcript,
+                    reference_memoir_markdown=baseline_excerpt,
+                    evidence_notes=(
+                        "严格按文档打分；真实性、事实覆盖率、可追溯性必须优先对照该用户历史访谈证据。"
+                    ),
+                )
                 chapter_results.append(
                     {
                         "id": ch.id,
@@ -310,7 +356,13 @@ class EvalJudgeManualService:
                 if not body:
                     continue
                 md = f"# 故事：{st.title}\n\n{_clip_md_for_judge(body)}"
-                sj = await judge.judge_memoir(memoir_markdown=md)
+                sj = await judge.judge_memoir(
+                    memoir_markdown=md,
+                    source_transcript=evidence_transcript,
+                    evidence_notes=(
+                        "严格按文档打分；真实性、事实覆盖率、可追溯性必须优先对照该用户历史访谈证据。"
+                    ),
+                )
                 story_results.append(
                     {
                         "id": st.id,