Read OpenDocument spreadsheets (.ods)

[vaceslav]

After the first release, once it is stable. OpenDocument is the native format of LibreOffice/OpenOffice and is increasingly required in public administration across Germany and the EU, so files will arrive in it.

It fits the architecture as one more cursor behind ITabularCursor; analysis, mapping and import do not change. Rough size: a day or two including tests.

Why it is simpler than xlsx

• Types are explicit per cell: office:value-type (float, percentage, currency, date, time, boolean, string) with office:value / office:date-value (ISO) / office:boolean-value — no number-format guessing, no styles lookup, no 1900/1904 epoch.
• No shared-string table: text is inline in the cell.
• Detected from content like the others: a zip whose first entry mimetype reads application/vnd.oasis.opendocument.spreadsheet.
• SheetScanner already strips prefixes (table:, office:); its kept-attribute set (today r, t, s) needs to become configurable.

The traps

• Repeat attributes. table:number-columns-repeated / table:number-rows-repeated — LibreOffice writes "the next 1,048,000 rows are empty" as one element. Never expand literally: represent empty repeats lazily, drop trailing empty repeats, and add a bound to the guide's Bounds table (it is also an attack vector).
• All sheets live in one content.xml. Reading sheet 3 means scanning past 1 and 2; forward reading is fine, there is no random access per sheet.
• Cell text assembly: text:p paragraphs (joined with a line break), nested text:span, text:s c="n" (n spaces), text:tab, text:line-break; annotations (office:annotation) excluded from the value.
• Covered cells of merges (table:covered-table-cell) count as empty cells in their position.
• Formulas: use the cached value, as for xlsx.
• Hidden sheets/rows, table:table-header-rows, flat ODS (.fods) — decide in or out.
• TabularFormat.Ods, TabularFile.Detect/Open, docs and README format list.

How to verify

Differential, as in #8: convert xlsx fixtures and other readers' corpora to ods with LibreOffice (soffice --headless --convert-to ods), read both through TriasDev.Tabular and compare cell by cell; add hand-built minimal fixtures for each trap above. Benchmark on the large fixtures converted to ods (peak memory must stay flat despite content.xml holding every sheet).