W5波+N-19收口：FTP鉴权落地、Trap25三连定案修复(N-30)、N-20/N-23/N-26

N-19 FTP 鉴权：
- lftpd_auth（USER/PASS 状态机 + 连续5次失败锁60s）+ SHA-256
  (user||pwd) 口令哈希；config.ini [ftp] 段 user/pwd_hash/write，
  出厂 admin/admin 在用有启动告警；lftpd_log.c 删除并入 lftpd_auth
- ftp_cfg 配置读写；台架 ftp_auth_bench.py 13/13

N-30 / Trap 25（CH395F_Trap_Records.md 四次定案全记录）：
- 修复1 closed-with-data accept：客户端发完正常 FIN 后芯片按
  FUN_PARA(0x09) §5.52② 保留 Socket+RX 缓冲，旧代码把 CLOSED 一律
  relisten 自毁缓冲 → STOR 必挂 425；改为接受"已关但可读"连接排空
- 修复2 丢 CONNECT 中断探针 net_lost_connect_probe + last_peer_port
  幽灵甄别（拆链窗口握手完成但中断丢失，双信源去抖补 accept）
- 修复3 send_list/send_nlst 批量单发：背靠背 ~60B 小段毒化芯片 TX
  （芯片报 SEND_OK 但对端全丢），整份目录一次性写入
- 证伪存档：固定沉降/Socket 轮换冷却/listen 门控（台架 v7 矩阵）
- 台架工具 test/board/ch395f_reuse_bench.{c,h} + host 脚本存档
  （不注册编译；实验分支 n30-chip-bench 保留）

N-20：Error_Handler 死循环降级（记录+复位路径）
N-23：scripts/check_registration.py 双端注册反向校验（pre-commit）
N-26：GD5F NAND 直 DMA 调用方缓冲窗口断言

验证：BUILD_OK 0/0；ftp_regression 连跑 4 轮 11/11（修复前
LIST#4/#7 100% 挂、STOR#2 曾 100% 425）；ftp_auth_bench 13/13；
@test 启动标记 PASS；ci_local L0+L0.5 全过；doc-links 通过。
登记簿 N-30 已销项。

scripts/ram_report.py

@@ -19,17 +19,23 @@ MAP_DEFAULT = os.path.join(
 )
 
 # Rows in "Image component sizes" that re-aggregate objects (would double-count sums).
-TOTALS_ROWS = {"Object Totals", "Library Totals", "User Totals", "mc_w.l", "mf_w.l"}
+TOTALS_ROWS = {"Object Totals", "Library Totals", "User Totals"}
+
+# ARM library aggregate rows (e.g. mc_w.l / mf_w.l / m_wm.l): the linker prints one
+# aggregate line per library in addition to the individual member .o lines, so any
+# "<lib>.l" row is a re-aggregation and must be dropped from per-file sums.
+# (2026-10-11: sha256's 64-bit shifts pulled in m_wm.l — the hardcoded name list
+#  missed it and the self-consistency check went +58; match the name pattern instead.)
+_LIB_AGG_RE = re.compile(r'^[a-z]{1,2}_[a-z0-9]+\.l$')
 
 
 def _is_aggregate(name):
     """Aggregate rows to drop from per-file sums: library-name summaries and Totals lines.
 
-    '(incl. Padding)' / '(incl. Generated)' are linker metadata; they are excluded from
-    per-file tables but their contribution IS counted in Grand Totals, so the per-file
-    sums will differ from Grand Totals by the alignment-padding delta (~300B here).
+    '(incl. Padding)' / '(incl. Generated)' are linker metadata ledger rows; they ARE
+    kept so that per-file sums reconcile exactly with Grand Totals (self-check Δ=0).
     """
-    return name in TOTALS_ROWS
+    return name in TOTALS_ROWS or bool(_LIB_AGG_RE.match(name))
 
 
 def parse_execution_region_header(line):