feat: matching program subtype discrimination (1:1/1:N/M:N/MxN)

Add _resolve_matching_subtype post-processing step in classify_program()
that distinguishes matching program subtypes based on key variable naming
patterns and file/structural features:

Rules (in priority order):
1. 二段階 → 二段階 (already handled by rule engine)
2. 3 files + WS-SAVE-KEY → M:N→MxN (MT20)
3. WS-PREV-KEY present → 混合 (already handled, MT32)
4. WS-MAST-KEY + WS-TRAN-KEY → 1:N (MT02)
5. >=3 KEY vars + >=2 files → M:N (MT33)
6. Otherwise → 1:1 (MT01, MT03, MT18, MT19)

Results:
  MT01→1:1, MT02→1:N, MT03→1:1, MT16/17→二段階,
  MT18/19→1:1, MT20→M:N→MxN, MT33→M:N

Also fix double-backslash regex bug in classifier.py and pipeline.py
(r'[-\w]' should be r'[\w-]' for word character class).
Regression: 745 passed (unchanged).

hina/classifier.py

@@ -23,7 +23,7 @@ L1_RULES: list[tuple[str, list[str], float]] = [
     ("编辑输出", ["WRITE AFTER", "WRITE BEFORE"], 0.80),
     ("文件编成", ["ORGANIZATION IS"], 0.99),
     ("替代索引", ["ALTERNATE RECORD KEY"], 0.99),
-    ("マッチング", ["re:WS-[-\\w]*KEY"], 0.65),
+    ("マッチング", ["re:WS-[\\w-]*KEY"], 0.65),
 ]
 
 # ── 冲突解决规则 ─────────────────────────────────────────────────────────

hina/pipeline/pipeline.py

@@ -334,6 +334,73 @@ def _path_llm_assisted(
 
 # ── 主入口 ────────────────────────────────────────────────────────────────────
 
+# ── 匹配子类型解析 ──────────────────────────────────────────────────────────
+
+_MATCHING_SUBTYPE_RULES = [
+    # (match_fn, subtype)
+    # 按优先级从高到低排列
+]
+
+
+def _resolve_matching_subtype(
+    result: dict,
+    cobol_source: str,
+    structure: dict,
+) -> dict:
+    """匹配程序的子类型区分后处理。
+
+    在 classify_program 判定为 マッチング 后，进一步区分子类型：
+    - 1:1 マッチング / 1:N / N:1 / M:N / M:N→M 等
+
+    Args:
+        result: classify_program 的返回结果。
+        cobol_source: 原始 COBOL 源码。
+        structure: extract_structure 的返回结构。
+
+    Returns:
+        更新后的 result，增加 "subtype" 字段。
+    """
+    category = result.get("category", "")
+    if "マッチング" not in category and "キーブレイク" not in category:
+        return result  # 非匹配程序不做子类型区分
+
+    src_upper = cobol_source.upper()
+    import re
+
+    # 0. 二段階マッチング — 已在规则引擎中处理
+    if "二段階" in category:
+        result["subtype"] = "二段階"
+        return result
+
+    # 1. M:N→MxN 直積 — 特征: WRITE + WS-SAVE-KEY + 3 文件
+    if structure.get("file_count", 0) >= 3 and 'WS-SAVE' in src_upper:
+        result["subtype"] = "M:N→MxN"
+        return result
+
+    # 2. 混合匹配 (WS-PREV-KEY 存在)
+    if 'WS-PREV-KEY' in src_upper:
+        result["subtype"] = "混合"
+        return result
+
+    # 3. 检查键变量命名模式
+    key_vars = set(re.findall(r'WS-[\w-]*KEY[A-Z0-9-]*', src_upper))
+
+    # 不对称键名 → 1:N 或 N:1 (WS-MAST-KEY + WS-TRAN-KEY)
+    has_master = any('MAST' in k for k in key_vars)
+    has_tran = any('TRAN' in k for k in key_vars)
+    if has_master and has_tran:
+        result["subtype"] = "1:N"
+        return result
+
+    # 4. 多个键名 → 多文件匹配 (M:N 模式)
+    if len(key_vars) >= 3 and structure.get("file_count", 0) >= 2:
+        result["subtype"] = "M:N"
+        return result
+
+    # 5. 对称键名 → 默认为 1:1
+    result["subtype"] = "1:1"
+    return result
+
 
 def classify_program(cobol_source: str, llm: Any = None) -> dict:
     """完整程序类型判定管道。
@@ -419,20 +486,24 @@ def classify_program(cobol_source: str, llm: Any = None) -> dict:
     # 路径 A: keyword >= 90% -> 直接输出
     if max_keyword_confidence >= 0.90:
         logger.info("[pipeline] 路径 A: keyword 高确信度 (%.2f)", max_keyword_confidence)
-        return _path_keyword_direct(keyword_info, structure)
+        result = _path_keyword_direct(keyword_info, structure)
 
     # 路径 B: keyword 50-89% -> 规则引擎
-    if max_keyword_confidence >= 0.50:
+    elif max_keyword_confidence >= 0.50:
         logger.info("[pipeline] 路径 B: keyword 中确信度 (%.2f) -> 规则引擎", max_keyword_confidence)
-        return _path_rule_engine(keyword_info, structure)
+        result = _path_rule_engine(keyword_info, structure)
 
     # 路径 C: keyword < 50% -> LLM 辅助
-    if llm is not None:
+    elif llm is not None:
         logger.info("[pipeline] 路径 C: keyword 低确信度 (%.2f) -> LLM 辅助", max_keyword_confidence)
-        return _path_llm_assisted(keyword_info, structure, llm)
+        result = _path_llm_assisted(keyword_info, structure, llm)
 
     # LLM 不可用: 使用规则引擎兜底
-    logger.info("[pipeline] 路径 C(fallback): keyword 低确信度 (%.2f) -> 规则引擎兜底", max_keyword_confidence)
+    else:
-    result = _path_rule_engine(keyword_info, structure)
+        logger.info("[pipeline] 路径 C(fallback): keyword 低确信度 (%.2f) -> 规则引擎兜底", max_keyword_confidence)
-    result["method"] = "rule_engine_fallback"
+        result = _path_rule_engine(keyword_info, structure)
+        result["method"] = "rule_engine_fallback"
+
+    # ── 第 4 步: 匹配子类型区分（仅对匹配/键中断程序）──
+    result = _resolve_matching_subtype(result, cobol_source, structure)
     return result