codebee 0.1.103 → 0.1.105

package/app/core/novel_quality.py

@@ -1,181 +1,285 @@
-# -*- coding: utf-8 -*-
-"""Deterministic quality checks for signing-oriented Chinese web fiction.
-
-The LLM remains responsible for literary judgment.  This module supplies
-stable rubric defaults and small text signals so reviewers can cite evidence
-for the common rejection reasons instead of relying on vague impressions.
-"""
-from __future__ import annotations
-
-import re
-
-
-SIGNING_RUBRIC = [
-    "开篇吸引力",
-    "情节推进",
-    "文风统一",
-    "情感细腻度",
-    "节奏控制",
-]
-
-OPENING_DIM = "开篇吸引力"
-SIGNING_CHECKPOINT_WORDS = 20000   # 项目检查点默认值；不是通用平台规则，调用方应核对当前后台
-
-_SENTENCE_RE = re.compile(r"[^。！？!?；;\n]+[。！？!?；;]?")
-_DIALOGUE_RE = re.compile(r"[“‘\"「『](.*?)[”’\"」』]", re.S)
-_EMOTION_RE = re.compile(r"哭|泪|笑|怒|怕|颤|抖|咬|攥|握紧|心跳|呼吸|喉咙|指甲|沉默")
-
-
-def rubric_for(task):
-    """Return the configured rubric, or the signing rubric for serial novels."""
-    task = task if isinstance(task, dict) else {}
-    raw = task.get("rubric")
-    if isinstance(raw, str):
-        raw = [x.strip() for x in raw.replace("，", ",").split(",") if x.strip()]
-    if isinstance(raw, (list, tuple)):
-        values = [str(x).strip() for x in raw if str(x).strip()]
-        if values:
-            return values[:8]
-    if str(task.get("type") or "").lower() == "serial_novel":
-        return list(SIGNING_RUBRIC)
-    return []
-
-
-def opening_requirements(chapter, target_words=None):
-    """Return concise, chapter-aware acceptance requirements for prompts."""
-    try:
-        chapter = int(chapter)
-    except (TypeError, ValueError):
-        chapter = 1
-    if chapter == 1:
-        return ("首章观察点（非硬门槛）：记录读者何时能辨认视角人物、故事承诺与正在发生的事；"
-                "核心卖点是否通过具体场景兑现；铺垫是否符合题材且有阅读价值；结尾是否形成自然的继续阅读理由。"
-                "根据文本效果和作者目标评估，不因未在固定字数内出现冲突/卖点/钩子而自动判失败。")
-    return ("本章观察点：核对与前文的承接、局面/人物/信息/情绪的变化及本章叙事功能；"
-            "结尾是否提供符合本章节奏的收束或继续阅读理由。不强制每章反转或悬念断章。")
-
-
-def signal_summary(text, chapter=None):
-    """Extract explainable text signals; no claim is made that they are scores."""
-    text = str(text or "").replace("\r\n", "\n").strip()
-    paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
-    sentences = [s.strip() for s in _SENTENCE_RE.findall(text) if s.strip()]
-    sentence_lengths = [len(re.sub(r"\s+", "", s)) for s in sentences]
-    dialogue_chars = sum(len(x) for x in _DIALOGUE_RE.findall(text))
-    emotion_hits = len(_EMOTION_RE.findall(text))
-    chars = len(re.sub(r"\s+", "", text))
-    avg = round(sum(sentence_lengths) / len(sentence_lengths), 1) if sentence_lengths else 0.0
-    long_sentences = sum(1 for n in sentence_lengths if n >= 55)
-    hints = []
-    if avg >= 35 or long_sentences:
-        hints.append("存在偏长句，复查长定语、重复说明与动作拆分")
-    if len(paragraphs) >= 4 and chars / len(paragraphs) > 220:
-        hints.append("段落偏长，复查手机阅读下的节奏与信息密度")
-    # Dialogue/emotion-action counts are descriptive signals only. Their absence
-    # is not a quality defect: narrative voice, genre, and scene purpose vary.
-    return {
-        "chapter": chapter,
-        "chars": chars,
-        "paragraphs": len(paragraphs),
-        "sentences": len(sentences),
-        "avg_sentence_chars": avg,
-        "long_sentences": long_sentences,
-        "dialogue_chars": dialogue_chars,
-        "emotion_action_hits": emotion_hits,
-        "hints": hints[:6],
-    }
-
-
-def format_signal_summary(summary):
-    """Render signal_summary output as a compact prompt block."""
-    if not isinstance(summary, dict):
-        return ""
-    parts = [
-        "文本信号（仅作证据，不替代文学判断）：约 %s 字；%s 段；%s 句；平均句长 %s 字；长句 %s；对白 %s 字；情绪动作线索 %s 处。"
-        % (summary.get("chars", 0), summary.get("paragraphs", 0),
-           summary.get("sentences", 0), summary.get("avg_sentence_chars", 0),
-           summary.get("long_sentences", 0), summary.get("dialogue_chars", 0),
-           summary.get("emotion_action_hits", 0))]
-    hints = summary.get("hints") or []
-    if hints:
-        parts.append("确定性复查提示：" + "；".join(str(x) for x in hints))
-    return "\n".join(parts)
-
-
-def style_anchor_block(anchor):
-    """Render the book-level style anchor as a prompt block; empty when absent."""
-    anchor = str(anchor or "").strip()
-    if not anchor:
-        return ""
-    return ("## 全书文风锚（大纲裁定，所有章节起草/修订/评审必须遵守；"
-            "与其冲突处以本锚为准）\n%s\n" % anchor[:400])
-
-
-def fingerprint(summary):
-    """Normalize a signal_summary into per-1000-char style metrics."""
-    if not isinstance(summary, dict):
-        return {}
-    chars = max(1, int(summary.get("chars") or 0))
-    return {
-        "avg_sentence_chars": float(summary.get("avg_sentence_chars") or 0.0),
-        "dialogue_per_k": round(1000.0 * float(summary.get("dialogue_chars") or 0) / chars, 1),
-        "emotion_per_k": round(1000.0 * float(summary.get("emotion_action_hits") or 0) / chars, 1),
-    }
-
-
-def _median(values):
-    xs = sorted(float(v) for v in values if isinstance(v, (int, float)))
-    if not xs:
-        return 0.0
-    mid = len(xs) // 2
-    return xs[mid] if len(xs) % 2 else round((xs[mid - 1] + xs[mid]) / 2.0, 1)
-
-
-def style_deviation_notes(baselines, summary, min_baseline=2):
-    """Cross-chapter style drift warnings from deterministic fingerprints.
-
-    baselines: signal_summary dicts of the book's previous chapters.  Needs at
-    least ``min_baseline`` chapters before speaking (开篇几章没有稳定基线)；
-    all notes are evidence hints for reviewers/revisers, never scores.
-    """
-    notes = []
-    base = [fingerprint(b) for b in (baselines or []) if isinstance(b, dict)]
-    base = [b for b in base if b]
-    if len(base) < min_baseline:
-        return notes
-    cur = fingerprint(summary)
-    if not cur:
-        return notes
-    avg_base = _median([b["avg_sentence_chars"] for b in base])
-    if avg_base >= 12 and cur["avg_sentence_chars"] > avg_base * 1.45:
-        notes.append("本章平均句长 %s 字，此前各章中位 %s 字——句式明显变长，"
-                     "复查长定语、说明性插叙与节奏拖沓"
-                     % (cur["avg_sentence_chars"], avg_base))
-    elif avg_base >= 12 and cur["avg_sentence_chars"] < avg_base * 0.62:
-        notes.append("本章平均句长 %s 字，此前各章中位 %s 字——句式明显变碎，"
-                     "复查是否为凑字数拆散动作与因果"
-                     % (cur["avg_sentence_chars"], avg_base))
-    dial_base = _median([b["dialogue_per_k"] for b in base])
-    if dial_base >= 40 and cur["dialogue_per_k"] < dial_base * 0.45:
-        notes.append("对白占比骤降（本章 %s/千字，此前中位 %s/千字）——"
-                     "复查是否换成了大段叙述，文风与推进方式是否漂移"
-                     % (cur["dialogue_per_k"], dial_base))
-    emo_base = _median([b["emotion_per_k"] for b in base])
-    if emo_base >= 1.5 and cur["emotion_per_k"] < emo_base * 0.4:
-        notes.append("情绪动作密度骤降（本章 %s/千字，此前中位 %s/千字）——"
-                     "复查复杂情感是否退回「很伤心/很愤怒」式概括"
-                     % (cur["emotion_per_k"], emo_base))
-    return notes
-
-
-def opening_below(means, threshold):
-    """True when the signing opening dimension is scored below the line."""
-    try:
-        value = float((means or {}).get(OPENING_DIM))
-    except (TypeError, ValueError):
-        return False
-    try:
-        line = float(threshold)
-    except (TypeError, ValueError):
-        return False
-    return value < line
+# -*- coding: utf-8 -*-
+"""Deterministic quality checks for signing-oriented Chinese web fiction.
+
+The LLM remains responsible for literary judgment.  This module supplies
+stable rubric defaults and small text signals so reviewers can cite evidence
+for the common rejection reasons instead of relying on vague impressions.
+"""
+from __future__ import annotations
+
+import re
+
+
+SIGNING_RUBRIC = [
+    "开篇吸引力",
+    "情节推进",
+    "文风统一",
+    "情感细腻度",
+    "节奏控制",
+]
+
+OPENING_DIM = "开篇吸引力"
+SIGNING_CHECKPOINT_WORDS = 20000   # 项目检查点默认值；不是通用平台规则，调用方应核对当前后台
+
+_SENTENCE_RE = re.compile(r"[^。！？!?；;\n]+[。！？!?；;]?")
+_DIALOGUE_RE = re.compile(r"[“‘\"「『](.*?)[”’\"」』]", re.S)
+_EMOTION_RE = re.compile(r"哭|泪|笑|怒|怕|颤|抖|咬|攥|握紧|心跳|呼吸|喉咙|指甲|沉默")
+
+# 逐字重复检测阈值（2026-10-11 连载实案校准）：整段 ≥40 字重复两遍是 113/118 章
+# 实际事故形态；完整对白复现（≥6 字）同理——有意复诵必有动作/归属铺垫，评审员
+# 对 113 章「这杆子歪着」的判词就是「若有意让角色复诵咀嚼，缺少铺垫」。叙述性
+# 短句复沓（「早上探，往里。」5 字）是白描节奏手段，落在阈值之下不误伤。
+_DUP_MIN_PARA_CHARS = 40
+_DUP_MIN_SENT_CHARS = 12
+_DUP_MIN_DIALOGUE_CHARS = 6
+_DUP_MAX_NOTES = 4
+
+
+def duplicate_notes(text):
+    """章内逐字重复检测（确定性，成稿事故级）。
+
+    抓三类：整段逐字重复、整句（≥12 字去空白）逐字重复、完整对白（≥6 字去空白）
+    逐字重复。返回结构化证据列表 [{kind, count, quote}]，供落盘验收打回重起草与
+    评审提示词引用；短句复沓等有意的节奏手段不在阈值之内，不构成证据。
+    """
+    text = str(text or "").replace("\r\n", "\n").strip()
+    if not text:
+        return []
+    notes = []
+    # ① 整段逐字重复（去空白比对；间隔段落/连排都算）
+    para_counts = {}
+    for p in re.split(r"\n\s*\n", text):
+        p = re.sub(r"\s+", "", p)
+        if len(p) >= _DUP_MIN_PARA_CHARS:
+            para_counts[p] = para_counts.get(p, 0) + 1
+    dup_paras = [p for p, c in sorted(para_counts.items(), key=lambda kv: -kv[1])
+                 if c >= 2]
+    for p in dup_paras[:2]:
+        notes.append({"kind": "整段重复", "count": para_counts[p], "quote": p[:40]})
+    # ② 整句逐字重复（已被整段重复覆盖的句子不再单列，避免一条事故报一串）
+    sent_counts = {}
+    for s in _SENTENCE_RE.findall(text):
+        s = re.sub(r"\s+", "", s)
+        if len(s) >= _DUP_MIN_SENT_CHARS:
+            sent_counts[s] = sent_counts.get(s, 0) + 1
+    for s, c in sorted(sent_counts.items(), key=lambda kv: -kv[1]):
+        if c < 2 or len(notes) >= _DUP_MAX_NOTES:
+            continue
+        if any(s in p for p in dup_paras):
+            continue
+        notes.append({"kind": "整句重复", "count": c, "quote": s[:40]})
+    # ③ 完整对白逐字重复（阈值更低：成稿事故最常以「对白整句连排两遍」出现）
+    seen_dialogue = {n["quote"] for n in notes}
+    dial_counts = {}
+    for d in _DIALOGUE_RE.findall(text):
+        d = re.sub(r"\s+", "", d)
+        if len(d) >= _DUP_MIN_DIALOGUE_CHARS:
+            dial_counts[d] = dial_counts.get(d, 0) + 1
+    for d, c in sorted(dial_counts.items(), key=lambda kv: -kv[1]):
+        if c < 2 or len(notes) >= _DUP_MAX_NOTES:
+            continue
+        if any(d in q for q in seen_dialogue):
+            continue
+        notes.append({"kind": "对白重复", "count": c, "quote": d[:40]})
+    return notes
+
+
+def format_duplicate_notes(notes):
+    """Render duplicate_notes evidence as human-readable lines."""
+    lines = []
+    for n in notes if isinstance(notes, list) else []:
+        try:
+            lines.append("%s %d 处：「%s」" % (n.get("kind", "重复"),
+                                              int(n.get("count") or 0),
+                                              str(n.get("quote") or "")[:40]))
+        except (TypeError, ValueError):
+            continue
+    return lines
+
+
+def rubric_for(task):
+    """Return the configured rubric, or the signing rubric for serial novels."""
+    task = task if isinstance(task, dict) else {}
+    raw = task.get("rubric")
+    if isinstance(raw, str):
+        raw = [x.strip() for x in raw.replace("，", ",").split(",") if x.strip()]
+    if isinstance(raw, (list, tuple)):
+        values = [str(x).strip() for x in raw if str(x).strip()]
+        if values:
+            return values[:8]
+    if str(task.get("type") or "").lower() == "serial_novel":
+        return list(SIGNING_RUBRIC)
+    return []
+
+
+def opening_requirements(chapter, target_words=None):
+    """Return concise, chapter-aware acceptance requirements for prompts."""
+    try:
+        chapter = int(chapter)
+    except (TypeError, ValueError):
+        chapter = 1
+    if chapter == 1:
+        return ("首章观察点（非硬门槛）：记录读者何时能辨认视角人物、故事承诺与正在发生的事；"
+                "核心卖点是否通过具体场景兑现；铺垫是否符合题材且有阅读价值；结尾是否形成自然的继续阅读理由。"
+                "根据文本效果和作者目标评估，不因未在固定字数内出现冲突/卖点/钩子而自动判失败。")
+    return ("本章观察点：核对与前文的承接、局面/人物/信息/情绪的变化及本章叙事功能；"
+            "结尾是否提供符合本章节奏的收束或继续阅读理由。不强制每章反转或悬念断章。")
+
+
+def signal_summary(text, chapter=None):
+    """Extract explainable text signals; no claim is made that they are scores."""
+    text = str(text or "").replace("\r\n", "\n").strip()
+    paragraphs = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
+    sentences = [s.strip() for s in _SENTENCE_RE.findall(text) if s.strip()]
+    sentence_lengths = [len(re.sub(r"\s+", "", s)) for s in sentences]
+    dialogue_chars = sum(len(x) for x in _DIALOGUE_RE.findall(text))
+    emotion_hits = len(_EMOTION_RE.findall(text))
+    chars = len(re.sub(r"\s+", "", text))
+    avg = round(sum(sentence_lengths) / len(sentence_lengths), 1) if sentence_lengths else 0.0
+    long_sentences = sum(1 for n in sentence_lengths if n >= 55)
+    hints = []
+    if avg >= 35 or long_sentences:
+        hints.append("存在偏长句，复查长定语、重复说明与动作拆分")
+    if len(paragraphs) >= 4 and chars / len(paragraphs) > 220:
+        hints.append("段落偏长，复查手机阅读下的节奏与信息密度")
+    # 逐字重复（整段/整句/完整对白）是成稿事故而非文风选择，直接点名证据；
+    # 起草验收闸（pipeline._chapter_state）用同一判定打回，评审面据此复核
+    for line in format_duplicate_notes(duplicate_notes(text)):
+        hints.append("逐字重复（成稿事故，须删并保留信息量最高的一处）：" + line)
+    # Dialogue/emotion-action counts are descriptive signals only. Their absence
+    # is not a quality defect: narrative voice, genre, and scene purpose vary.
+    return {
+        "chapter": chapter,
+        "chars": chars,
+        "paragraphs": len(paragraphs),
+        "sentences": len(sentences),
+        "avg_sentence_chars": avg,
+        "long_sentences": long_sentences,
+        "dialogue_chars": dialogue_chars,
+        "emotion_action_hits": emotion_hits,
+        "hints": hints[:6],
+    }
+
+
+def format_signal_summary(summary):
+    """Render signal_summary output as a compact prompt block."""
+    if not isinstance(summary, dict):
+        return ""
+    parts = [
+        "文本信号（仅作证据，不替代文学判断）：约 %s 字；%s 段；%s 句；平均句长 %s 字；长句 %s；对白 %s 字；情绪动作线索 %s 处。"
+        % (summary.get("chars", 0), summary.get("paragraphs", 0),
+           summary.get("sentences", 0), summary.get("avg_sentence_chars", 0),
+           summary.get("long_sentences", 0), summary.get("dialogue_chars", 0),
+           summary.get("emotion_action_hits", 0))]
+    hints = summary.get("hints") or []
+    if hints:
+        parts.append("确定性复查提示：" + "；".join(str(x) for x in hints))
+    return "\n".join(parts)
+
+
+def style_anchor_block(anchor):
+    """Render the book-level style anchor as a prompt block; empty when absent."""
+    anchor = str(anchor or "").strip()
+    if not anchor:
+        return ""
+    return ("## 全书文风锚（大纲裁定，所有章节起草/修订/评审必须遵守；"
+            "与其冲突处以本锚为准）\n%s\n" % anchor[:400])
+
+
+def reuse_guard_block(samples):
+    """Render prior chapters' opening/closing excerpts as a negative-example block.
+
+    samples: [{"chapter": int, "head": str, "tail": str}] — 起草面据此避开同款
+    开笔句式与同型收束（跨章复用红线：同一动作句式至多两次，第二次必须换载体
+    或升级信息量）。无样例（首批第 1 章等）返回空串，不占提示词字节。
+    """
+    rows = []
+    for s in samples if isinstance(samples, list) else []:
+        if not isinstance(s, dict):
+            continue
+        try:
+            ch = int(s.get("chapter"))
+        except (TypeError, ValueError):
+            continue
+        head = str(s.get("head") or "").strip()
+        tail = str(s.get("tail") or "").strip()
+        if head:
+            rows.append("- 第 %d 章开笔：「%s」" % (ch, head[:60]))
+        if tail:
+            rows.append("- 第 %d 章收束：「…%s」" % (ch, tail[-70:]))
+    if not rows:
+        return ""
+    return ("## 跨章复用红线（起草前必读）\n"
+            "此前各章已用的开笔/收束形态如下——本章不得再以同款句式开笔或同型"
+            "句式收束；同一动作句式全书至多两次，第二次必须换载体或升级信息量，"
+            "连读时章尾节奏不得趋同：\n" + "\n".join(rows) + "\n")
+
+
+def fingerprint(summary):
+    """Normalize a signal_summary into per-1000-char style metrics."""
+    if not isinstance(summary, dict):
+        return {}
+    chars = max(1, int(summary.get("chars") or 0))
+    return {
+        "avg_sentence_chars": float(summary.get("avg_sentence_chars") or 0.0),
+        "dialogue_per_k": round(1000.0 * float(summary.get("dialogue_chars") or 0) / chars, 1),
+        "emotion_per_k": round(1000.0 * float(summary.get("emotion_action_hits") or 0) / chars, 1),
+    }
+
+
+def _median(values):
+    xs = sorted(float(v) for v in values if isinstance(v, (int, float)))
+    if not xs:
+        return 0.0
+    mid = len(xs) // 2
+    return xs[mid] if len(xs) % 2 else round((xs[mid - 1] + xs[mid]) / 2.0, 1)
+
+
+def style_deviation_notes(baselines, summary, min_baseline=2):
+    """Cross-chapter style drift warnings from deterministic fingerprints.
+
+    baselines: signal_summary dicts of the book's previous chapters.  Needs at
+    least ``min_baseline`` chapters before speaking (开篇几章没有稳定基线)；
+    all notes are evidence hints for reviewers/revisers, never scores.
+    """
+    notes = []
+    base = [fingerprint(b) for b in (baselines or []) if isinstance(b, dict)]
+    base = [b for b in base if b]
+    if len(base) < min_baseline:
+        return notes
+    cur = fingerprint(summary)
+    if not cur:
+        return notes
+    avg_base = _median([b["avg_sentence_chars"] for b in base])
+    if avg_base >= 12 and cur["avg_sentence_chars"] > avg_base * 1.45:
+        notes.append("本章平均句长 %s 字，此前各章中位 %s 字——句式明显变长，"
+                     "复查长定语、说明性插叙与节奏拖沓"
+                     % (cur["avg_sentence_chars"], avg_base))
+    elif avg_base >= 12 and cur["avg_sentence_chars"] < avg_base * 0.62:
+        notes.append("本章平均句长 %s 字，此前各章中位 %s 字——句式明显变碎，"
+                     "复查是否为凑字数拆散动作与因果"
+                     % (cur["avg_sentence_chars"], avg_base))
+    dial_base = _median([b["dialogue_per_k"] for b in base])
+    if dial_base >= 40 and cur["dialogue_per_k"] < dial_base * 0.45:
+        notes.append("对白占比骤降（本章 %s/千字，此前中位 %s/千字）——"
+                     "复查是否换成了大段叙述，文风与推进方式是否漂移"
+                     % (cur["dialogue_per_k"], dial_base))
+    emo_base = _median([b["emotion_per_k"] for b in base])
+    if emo_base >= 1.5 and cur["emotion_per_k"] < emo_base * 0.4:
+        notes.append("情绪动作密度骤降（本章 %s/千字，此前中位 %s/千字）——"
+                     "复查复杂情感是否退回「很伤心/很愤怒」式概括"
+                     % (cur["emotion_per_k"], emo_base))
+    return notes
+
+
+def opening_below(means, threshold):
+    """True when the signing opening dimension is scored below the line."""
+    try:
+        value = float((means or {}).get(OPENING_DIM))
+    except (TypeError, ValueError):
+        return False
+    try:
+        line = float(threshold)
+    except (TypeError, ValueError):
+        return False
+    return value < line