feat: export core Hermes skills

2026-07-15 02:45:56 +00:00
parent a028b63eda
commit 54711fee2a
308 changed files with 41310 additions and 1 deletions
@@ -0,0 +1,127 @@
+# ContractEditor 原文Run属性污染诊断与修复
+
+## 2026-07-13 洋励合同实证
+
+### 问题描述
+
+ContractEditor（contract_docx_lib.py）在处理文档时，不仅给WB INS runs添加多余属性，还会**修改原文runs**的rPr——给本来靠docDefaults/style继承的orig runs添加显式eastAsia/cs/sz。
+
+### 典型污染模式
+
+| 属性 | 原文（待审查） | 被污染后（交付物中的orig run） | WB INS run |
+|------|--------------|-------------------------------|-----------|
+| eastAsia | None (继承minorEastAsia) | **宋体** (被加) | None |
+| cs | None | **宋体** (被加) | None |
+| sz | None (继承docDefaults=22) | **21** (被加且值错) | None |
+| ascii | 宋体 | 宋体 | None |
+| hint | eastAsia (部分有) | eastAsia | None |
+
+### 后果
+
+1. 原文所有文字从11pt(docDefaults sz=22)变成10.5pt(显式sz=21) — 整体缩小0.5pt
+2. INS文字没有任何属性 → 走docDefaults 11pt → 与被改小的原文不一致
+3. 字体验证脚本(wb-ins-font-verify.py)报INS缺属性，但实际问题是orig被污染
+
+### 诊断步骤
+
+```bash
+# 1. 读原文代表性段落run rPr
+python3 -c "
+import zipfile
+from lxml import etree
+WNS = '{http://schemas.openxmlformats.org/wordprocessingml/2006/main}'
+with zipfile.ZipFile('原文.docx', 'r') as z:
+    ...
+# 检查: eastAsia=None? sz=None? 
+# 如果是 → 原文靠继承
+
+# 2. 读交付物同段落orig run rPr
+# 检查: 是否多了eastAsia/cs/sz?
+# 如果是 → 被污染
+```
+
+### 修复代码模板
+
+```python
+import zipfile, tempfile, shutil
+from lxml import etree
+
+WNS = '{http://schemas.openxmlformats.org/wordprocessingml/2006/main}'
+
+with zipfile.ZipFile(filepath, 'r') as z:
+    all_files = {n: z.read(n) for n in z.namelist()}
+
+tree = etree.fromstring(all_files['word/document.xml'])
+body = tree.find(f'{WNS}body')
+
+# Step 1: Strip contaminated attributes from ALL orig runs
+for p in body.findall(f'{WNS}p'):
+    for r in p.findall(f'{WNS}r'):  # Only direct child runs (not inside ins/del)
+        rpr = r.find(f'{WNS}rPr')
+        if rpr is None:
+            continue
+        rf = rpr.find(f'{WNS}rFonts')
+        if rf is not None:
+            # Strip eastAsia (original didn't have it)
+            if f'{WNS}eastAsia' in rf.attrib:
+                del rf.attrib[f'{WNS}eastAsia']
+            # Strip cs (original didn't have it)
+            if f'{WNS}cs' in rf.attrib:
+                del rf.attrib[f'{WNS}cs']
+        # Strip sz (original relies on docDefaults)
+        sz = rpr.find(f'{WNS}sz')
+        if sz is not None:
+            rpr.remove(sz)
+
+# Step 2: Fix INS runs to match REAL original format
+for ins in body.findall(f'.//{WNS}ins'):
+    if ins.get(f'{WNS}author') != 'WB':
+        continue
+    for r in ins.findall(f'{WNS}r'):
+        rpr = r.find(f'{WNS}rPr')
+        if rpr is None:
+            continue
+        rf = rpr.find(f'{WNS}rFonts')
+        if rf is None:
+            rf = etree.SubElement(rpr, f'{WNS}rFonts')
+        # Match real original: ascii=宋体, hAnsi=宋体, NO eastAsia
+        rf.set(f'{WNS}ascii', '宋体')
+        rf.set(f'{WNS}hAnsi', '宋体')
+        if f'{WNS}eastAsia' in rf.attrib:
+            del rf.attrib[f'{WNS}eastAsia']
+        # Remove sz (let it inherit)
+        sz = rpr.find(f'{WNS}sz')
+        if sz is not None:
+            rpr.remove(sz)
+
+# Step 3: Per-paragraph hint matching
+for p in body.findall(f'{WNS}p'):
+    # Get orig run's hint
+    orig_hint = None
+    for r in p.findall(f'{WNS}r'):
+        rpr = r.find(f'{WNS}rPr')
+        if rpr is not None:
+            rf = rpr.find(f'{WNS}rFonts')
+            orig_hint = rf.get(f'{WNS}hint') if rf is not None else None
+        break
+    # Apply to INS runs in same paragraph
+    for ins in p.findall(f'.//{WNS}ins'):
+        if ins.get(f'{WNS}author') != 'WB':
+            continue
+        for r in ins.findall(f'{WNS}r'):
+            rpr = r.find(f'{WNS}rPr')
+            if rpr is None: continue
+            rf = rpr.find(f'{WNS}rFonts')
+            if rf is None: continue
+            if orig_hint:
+                rf.set(f'{WNS}hint', orig_hint)
+            elif f'{WNS}hint' in rf.attrib:
+                del rf.attrib[f'{WNS}hint']
+```
+
+### 注意事项
+
+1. **必须对比原文确定被污染了哪些属性** — 不同合同模板的原文属性不同
+2. **不是所有合同都有此问题** — 取决于原文是否靠继承（有显式属性的不会被"污染"，因为值相同）
+3. **Step 1必须在Step 2之前** — 否则wb-ins-font-verify仍会报INS与(被污染的)orig不一致
+4. **hint要逐段处理** — 同一文档不同段落的orig runs可能有的有hint有的没有