csv_to_mdb --verify: 四档汇总(一致/缺口/边界/超日历), 不再把少行说成一致

scripts/csv_to_mdb.py

@@ -188,7 +188,8 @@ def verify(out: pathlib.Path, cls_filter: str | None, refresh: bool, deep: bool)
     src = src_month_rows(cls_filter, out / '_csv_month_rows.json', refresh) if deep else {}
     print(f'== CSV → MDB 全量核对 · {out} ==')
     print(f'   {"类":6s} {"月份":8s} {"mdb 各表行数":18s} {"应得行数":10s} 判定   zip')
-    bad = []
+    bad, short, over, minor = [], [], [], []
+    BOUNDARY = 0.005            # 超日历 ≤0.5% 视为**月末边界行**(有些台次月 00:00 那一行落在本月)
     for m in mdbs:
         cls = m.stem.rsplit('-', 1)[1]
         ym = '-'.join(m.stem.split('-')[:2])
@@ -199,20 +200,37 @@ def verify(out: pathlib.Path, cls_filter: str | None, refresh: bool, deep: bool)
             want = CAD.get(cls, 0) * calendar.monthrange(y, mo)[1] * 38
         aligned = len(set(rows)) == 1 if rows else False
         d = (rows[0] - want) if rows else -want
-        # 日历模式: 少于日历=数据缺口(正常), 多于日历才是错; 深核对模式: 必须与源逐行相等
-        ok = bool(rows) and aligned and (rows[0] == want if deep else rows[0] <= want)
-        tag = ('OK' if ok else '异常') if deep else ('一致' if d == 0 else (f'缺口{d:+d}' if d < 0 else f'超日历{d:+d}'))
+        # 深核对: 必须与源逐行相等; 日历模式: 少行=该月数据本身不齐(正常), 多行超过 0.5% 才算转换错
+        if deep:
+            ok = bool(rows) and aligned and rows[0] == want
+        else:
+            ok = bool(rows) and aligned and d <= want * BOUNDARY
+        tag = (('OK' if ok else '异常') if deep else
+               ('一致' if d == 0 else (f'缺口{d:+d}' if d < 0 else
+                                      (f'超日历{d:+d}' if d > want * BOUNDARY else f'边界{d:+d}'))))
         zp = m.parent / f'{ym}-{cls}.zip'
         ztxt = f'有 {zp.stat().st_size / 1e6:.1f} MB' if zp.is_file() else '无'
         print(f'   {cls:6s} {ym:8s} {",".join(map(str, rows)):18s} {want:<10d} '
               f'{tag:12s} {ztxt}')
         if not ok:
             bad.append((cls, ym, rows, want))
-    print(f'   合计: {len(mdbs)} 个月库 · 一致 {len(mdbs) - len(bad)} · 差异 {len(bad)}'
-          + ('' if deep else '（按日历推算; 差异多为该月首尾不完整, 属正常 ⇒ 要按源逐行核对加 --deep）'))
+        if rows and d < 0:
+            short.append((cls, ym, rows, want))
+        elif rows and d > 0:
+            (over if d > want * BOUNDARY else minor).append((cls, ym, rows, want))
+    # ★2026-09-19 修: 原来汇总只报"一致 N / 差异 K", 而日历判据是 `rows <= want` ⇒ **任何缺口都算一致**,
+    #   于是 1min 那张表有 8 个月短几百到一万多行, 汇总却是"一致 18 · 差异 0"(自相矛盾, 只有看表才发现)。
+    #   现在四档分开报: 一致 / 缺口(少, 数据本身) / 边界(多但 ≤0.5%, 月末那一行) / 超日历(多且超 0.5%, 才是转换错)。
+    print(f'   合计: {len(mdbs)} 个月库 · 一致 {len(mdbs) - len(bad)} · 缺口 {len(short)} · '
+          f'边界 {len(minor)} · 超日历 {len(over)}'
+          + ('' if deep else '（按日历推算; 少行=该月数据本身不齐, 多行>0.5% 才是转换错; 逐行对源加 --deep）'))
+    for cls, ym, rows, want in sorted(short, key=lambda x: x[2][0] - x[3])[:3]:
+        print(f'      [i] 缺口最大: {cls} {ym}: {rows[0]:,} 行 (差 {rows[0] - want:+,} = '
+              f'{100 * (rows[0] - want) / want:.2f}%)')
     for cls, ym, rows, want in bad[:8]:
         print(f'      [i] {cls} {ym}: mdb={rows} vs 应得 {want}')
-    print('   结论: ' + ('全部一致（各分片表行数相同 ⇒ 逐行对齐）' if not bad else '有差异, 见上'))
+    print('   结论: ' + ('全部一致（各分片表行数相同 ⇒ 逐行对齐）' if not bad else
+                       ('有差异, 见上' if over or deep else '分片表行数整齐; 少行属当月数据不齐(见上), 无转换错')))
     return 0 if not bad else 5