v0.2.0

导出 Excel、直接读 .zip 压缩包、一条命令看发票，命令行改成中文；阅读顺序更准，新增发票种类。
Excel export, ZIP archives, a one-command invoice view and a Chinese command line; better reading order and the invoice type.

pip install --upgrade ofdkit
ofdkit 发票.ofd                           # 看一张发票 / show an invoice
ofdkit batch 发票文件夹 -o 发票汇总.xlsx   # 汇总到 Excel / summarise into Excel

Added

• Excel output: ofdkit batch FOLDER -o 发票汇总.xlsx writes a workbook, still with only the standard library. Invoice numbers, codes and tax IDs stay text (in a CSV, Excel turns a 20-digit 数电票 number into 2.4E+19), amounts are numbers and dates are dates. A totals row follows filters; duplicates and unreadable files are coloured and explained; a second sheet lists every line item.
• ZIP archives: batch, show and the MCP tool list_ofd_invoices read the .ofd files inside .zip archives, such as a batch download from the tax service.
• ofdkit show, and the shortcut ofdkit PATH: an invoice as aligned Chinese text, another document's text, or a table of the invoices in folders and archives with totals.
• ofdkit.read_invoices() reads files, folders and archives into InvoiceFile results.
• Totals without duplicates, and sums of amounts and tax, in the batch summary and list_ofd_invoices; an item_names column in the CSV.
• ofdkit describe: the file's structure as JSON with all text masked, so problems can be reported without sharing an invoice. GitHub issue forms ask for it instead of the file.
• invoice_type: the invoice type as printed in its title, such as 电子发票（增值税专用发票） or 上海增值税电子普通发票.

Changed

• Better reading order. Vertical labels such as 购买方信息 and 备注 are read as words; columns of prose are read column by column, tables still row by row.
• Command-line help and messages are in Chinese. text and show say when a file has no text at all, which usually means a scan.
• The CSV from batch has new invoice_type and item_names columns.

Fixed

• Older invoices that share a number but have different invoice codes are no longer flagged as duplicates.
• Output piped into a program that stops reading early, such as head, no longer ends with an error.
• The MCP tool read_ofd_text with a page range no longer fails when one page in the range is damaged.

Four new reading-order samples in ofd-samples check these changes on every commit.