框架: clean_apply加--report-md质检报告(每条删除带上下文对照,门①裁决证据面)

.claude/skills/clean/scripts/clean_apply.py

@@ -47,16 +47,24 @@ def flex_pattern(exact):
 
 
 def try_remove(text, exact):
-    """在 text 中删除 exact（精确→弹性两级）。返回 (新text, 命中次数, 实删字数, 方式)。"""
+    """在 text 中删除 exact（精确→弹性两级）。
+
+    返回 (新text, 命中次数, 实删字数, 方式, 首次命中前文, 首次命中后文)。
+    前后文各取 ~60 字：质检报告用——人工确认删除段是独立垃圾而非正文一部分。
+    """
     n = text.count(exact)
     if n:
-        return text.replace(exact, ""), n, len(exact) * n, "精确"
+        i = text.find(exact)
+        before, after = text[max(0, i - 60):i], text[i + len(exact):i + len(exact) + 60]
+        return text.replace(exact, ""), n, len(exact) * n, "精确", before, after
     pat = flex_pattern(exact)
     if pat:
-        new, n = pat.subn("", text)
+        m = pat.search(text)
-        if n:
+        if m:
-            return new, n, len(text) - len(new), "空白弹性"
+            before, after = text[max(0, m.start() - 60):m.start()], text[m.end():m.end() + 60]
-    return text, 0, 0, ""
+            new, n = pat.subn("", text)
+            return new, n, len(text) - len(new), "空白弹性", before, after
+    return text, 0, 0, "", "", ""
 
 
 @click.command()
@@ -65,7 +73,8 @@ def try_remove(text, exact):
 @click.option("--file", "file_", type=click.Path(exists=True), required=True)
 @click.option("--model", default="MiniMax-M3", show_default=True, help="建议来源模型（审计标注）")
 @click.option("--dry-run", is_flag=True, help="只对账不落库")
-def main(work_id, batch, file_, model, dry_run):
+@click.option("--report-md", type=click.Path(), help="人读质检报告输出路径（每条删除带上下文对照）")
+def main(work_id, batch, file_, model, dry_run, report_md):
     raw = json.loads(pathlib.Path(file_).read_text())
     if isinstance(raw, dict):
         raw = raw.get("deletions") or raw.get("删除") or []
@@ -126,7 +135,7 @@ def main(work_id, batch, file_, model, dry_run):
                             stats["拒绝"].append(f"#{no}: 含章题 {u_exact[:24]!r}")
                             hit_no = -1  # 标记已裁决，不再试邻章
                             break
-                        new, n, removed, how = try_remove(ch["text"], u_exact)
+                        new, n, removed, how, before, after = try_remove(ch["text"], u_exact)
                         if n:
                             ch["text"] = new
                             ch["removed"] += removed
@@ -137,7 +146,7 @@ def main(work_id, batch, file_, model, dry_run):
                             if cand != no:
                                 stats["邻章命中"] += 1
                                 note += f"(报#{no}实删#{cand})"
-                            ch["audit"].append((u_exact, n, note))
+                            ch["audit"].append((u_exact, n, note, before, after))
                             stats["删除段"] += 1
                             stats["删除字数"] += removed
                             hit_no = cand
@@ -163,7 +172,7 @@ def main(work_id, batch, file_, model, dry_run):
                 wc = len(re.sub(r"\s", "", ch["text"]))
                 conn.execute("UPDATE muse_content_block SET content_text=%s, word_count=%s, updater=%s WHERE id=%s",
                              (ch["text"], wc, ACTOR, ch["blk_id"]))
-                for exact, n, note in ch["audit"]:
+                for exact, n, note, _, _ in ch["audit"]:
                     conn.execute(
                         """INSERT INTO example_clean_log (work_id, chapter_id, batch, removed_text, occurrences,
                                reason, model, creator, updater, tenant_id)
@@ -184,8 +193,30 @@ def main(work_id, batch, file_, model, dry_run):
         # dry-run 打印完整删除明细：门①人工质检的数据源
         click.echo("── 删除明细（质检用） ──")
         for no, ch in dirty:
-            for exact, n, note in ch["audit"]:
+            for exact, n, note, _, _ in ch["audit"]:
                 click.echo(f"  [删] #{no} ×{n} {note or '(无理由)'}\n      {exact!r}")
+    if report_md:
+        # 人读质检报告：每条删除带前后文对照——放行裁决的证据面
+        lines_md = [f"# 清洗质检报告（work={work_id} batch={batch}{mode}）",
+                    "",
+                    f"拟删 **{stats['删除段']} 段 / {stats['删除字数']:,} 字**"
+                    f"（弹性命中 {stats['弹性命中']}、邻章命中 {stats['邻章命中']}）；"
+                    f"守卫拒绝 {len(stats['拒绝'])} 条（见文末）。",
+                    "",
+                    "> 看法：确认每条【删】段是独立垃圾、上下文（灰引文）是完整正文即可放行。", ""]
+        seq = 0
+        for no, ch in dirty:
+            for exact, n, note, before, after in ch["audit"]:
+                seq += 1
+                ctx_b = before.replace("\n", " ⏎ ").strip() or "（章首）"
+                ctx_a = after.replace("\n", " ⏎ ").strip() or "（章尾）"
+                body = exact.replace("\n", " ⏎ ")
+                lines_md += [f"### {seq}. 第{no}章《{ch['title']}》 ×{n} ｜ {note or '（无理由）'}",
+                             f"> …{ctx_b}", f"", f"**【删】** `{body}`", f"", f"> {ctx_a}…", ""]
+        lines_md += ["## 守卫拒绝清单（未删，含保护的疑正文行）", ""]
+        lines_md += [f"- {r}" for r in stats["拒绝"]] or ["- 无"]
+        pathlib.Path(report_md).write_text("\n".join(lines_md))
+        click.echo(f"质检报告已写 {report_md}")
 
 
 if __name__ == "__main__":