feat(write_attachments): add --append mode, fix batch timing and SQL Server param limit

- Add --append flag: diff source-table 总排号 against existing
  Common.Attachment SN and classify/write only the missing ones. --limit
  caps the per-run append count, --order sets the direction. Backed by
  new fetch_existing_attachment_sns() in db.py.
- Fix batch elapsed-time accounting: summarize_results() summed each
  task's per-item elapsed_ms, which overcounts under ThreadPoolExecutor
  concurrency (cumulative work time, not real wall-clock — 100 tasks on
  8 workers reported ~5x the actual runtime). Callers now time
  classify_batch() via perf_counter and pass wall_clock_ms; both the
  write_attachments [汇总] line and main.py --summary report wall-clock
  separately from the cumulative sum.
- Format durations >=1s in seconds (88851.4 ms -> 88.85 s) in the
  human-readable [汇总] line; structured JSON --summary fields stay in ms.
- Chunk all IN (...) lists to 2000 items to respect SQL Server's 2100
  bind-parameter hard limit (previously --append --limit 5000 failed at
  the fetch step with "COUNT 字段不正确"). Applied to fetch_params_by_ids
  (sn / id paths) and to the DELETE inside upsert_attachments.

Co-Authored-By: Claude <noreply@anthropic.com>

README.md

@@ -173,9 +173,13 @@ python main.py --sn 26B742,26B743 --summary
 "meta": {"elapsed_ms": 1476.6, "attempts": 1, "llm": {"model": "deepseek-v4-flash", "elapsed_ms": 1472.8, "usage": {"prompt_tokens": 1297, "completion_tokens": 15, "total_tokens": 1312, "prompt_cache_hit_tokens": 1280, "prompt_cache_miss_tokens": 17, "cache_hit_rate": 0.9869}}}
 ```
 
-`--summary` 打印的整批汇总字段与 `meta.llm.usage` 对应：`count` / `total_elapsed_ms` / `llm_calls` /
-各类 `total_*_tokens` / `total_cache_hit_tokens` / `total_cache_miss_tokens` / `avg_cache_hit_rate`
-（按 token 加权的整体命中率）。
+`--summary` 打印的整批汇总字段与 `meta.llm.usage` 对应：`count` / `wall_clock_ms` / `total_elapsed_ms` /
+`llm_calls` / 各类 `total_*_tokens` / `total_cache_hit_tokens` / `total_cache_miss_tokens` /
+`avg_cache_hit_rate`（按 token 加权的整体命中率）。
+
+两个耗时字段的区别（`classify_batch` 是并发执行的）：
+- `wall_clock_ms`：整批分类的**真实墙钟**耗时（在 `classify_batch` 外层用 `perf_counter` 计时），即你实际等的时间。
+- `total_elapsed_ms`：各结果自身耗时 `elapsed_ms` 的**累加**（相当于把这些任务串行跑的总时长），并发下会明显大于墙钟；二者之比 ≈ 并发增益（≈ `max_workers`）。
 
 ## 对话日志
 
@@ -265,6 +269,11 @@ python write_attachments.py --range 800,805
 # 组合：区间内降序、再取前 3 个
 python write_attachments.py --range 800,805 --order desc --limit 3
 
+# 追加模式：仅补写源表中存在、但 Common.Attachment 还没有的总排号
+# --limit 此时是"本次最多追加多少个 SN"的上限；--order 决定从哪一端补起
+python write_attachments.py --append --limit 10
+python write_attachments.py --append --order desc --limit 10
+
 # 指定总排号文件（每行一个，键类型 sn）
 python write_attachments.py --ids-file ids.txt
 
@@ -293,7 +302,9 @@ python write_attachments.py --mode fine --enable-other
 ```
 
 运行结束后，除上面的 `ok/empty_param/not_found/error` 与"将写入行数/跳过数"统计外，还会打印一行
-`[汇总]`：本批总耗时、LLM 调用次数、token 总量（prompt/completion/total）与加权缓存命中率，便于核算成本。
+`[汇总]`：本批**总耗时**（整批分类的真实墙钟）、**LLM 累计**（各任务自身耗时的累加，并发下大于墙钟，
+括号标注并发数 `max_workers`）、LLM 调用次数、token 总量（prompt/completion/total）与加权缓存命中率，
+便于核算成本与并发效率。
 
 落库约定（与 `db.py` 常量、`write_attachments.py` 的转换逻辑保持一致）：