fix: convert CJK built-in date formats in xlsx extraction (#736)

[kevin-hans]

Problem

Ingesting an .xlsx cell formatted with one of Excel's CJK built-in date formats — 27–36 in ECMA-376 §18.8.30, e.g. 31 = yyyy年m月d日, 32 = m月d日 — surfaces the cell as its raw Excel serial number in the extracted markdown. A column that displayed 2025年4月14日 in Excel becomes 45761, and the ingest LLM then flags the column as anomalous numeric data (#736).

Root cause is upstream in calamine: builtin_format_by_id (both 0.34 and 0.36) only knows the English built-in date IDs (14–22, 45, 47). Cells styled with 27–36 are tagged CellFormat::Other, so worksheet_range returns them as Data::Float(serial) and extract_spreadsheet in commands/fs.rs writes the serial verbatim. Chinese / Japanese / Korean Excel emits these IDs whenever a user picks a locale-native date format from the built-in list, so it hits real files in practice.

Custom <numFmt> codes with numFmtId ≥ 164 (e.g. m"月"d"日") are unaffected — calamine's detect_custom_number_format picks up the leading m/d/y/h/s tokens correctly. Only the built-in-ID branch was blind.

Fix

For .xlsx / .xlsm only, side-parse the workbook zip once before markdown emission:

1. Read xl/styles.xml and collect every cellXf index whose numFmtId is in 27..=36 and is not overridden by a <numFmt> entry (overrides are left to calamine's custom-format detector).
2. If that set is empty — the vast majority of workbooks — bail out immediately, so the common path pays zero per-cell cost.
3. Otherwise walk each sheet XML, collect the (row, col) of every <c s="..."> styled by a CJK date cellXf, and pre-convert its <v> serial to an ISO string (YYYY-MM-DD, or YYYY-MM-DD HH:MM:SS when a fractional time is present).
4. In extract_spreadsheet, a Data::Float cell with a matching coordinate uses the ISO override; every other cell goes through the existing formatter unchanged.

Handles the 1904 date system and Excel's phantom 1900-02-29 (serial 60 → no real calendar date, returns None so the raw serial stays visible rather than being mis-dated). .xls / .ods continue through calamine unchanged.

Verified

• Hand-built regression fixture (xlsx with numFmtId=31, serial 44562) in extract_spreadsheet_converts_cjk_builtin_date_format_ids: red before the fix (| 44562 | in output), green after (2022-01-01).
• cargo test --lib → 401 passed / 0 failed. All 33 commands::fs tests still pass, including anydoc_extracts_generated_structured_document_fixtures, which exercises the styles-less minimal_xlsx fixture and confirms the fast-path early-return does not misbehave.

Notes for review

• The XML parsing for styles.xml / sheet XML uses the same hand-rolled string-scanning style already used by extract_docx_markdown and extract_odf_text in this file, deliberately avoiding a new direct dependency on quick-xml (only transitively present through calamine today).
• The right long-term fix is upstream in calamine — teaching builtin_format_by_id about IDs 27–36. Once that lands and we bump calamine, the override map will simply be empty for every workbook and this code becomes dormant; no callers need to change.
• Output uses ISO dates (2022-01-01) rather than the workbook's original CJK glyphs (2022年1月1日). This matches what calamine emits for the dates it already recognizes and is what the ingest LLM interprets most reliably; happy to switch to locale-native rendering if preferred.

Fixes #736.

[kevin-hans]

@nashsu
hello,would you plz review this pr?