Spreadsheet editing: start with plain ods cells, then solve repeated and empty ones

[andiwand]

Editing is off for spreadsheets everywhere: odf::Document::is_editable returns true only for text/presentation/drawing (odf_document.cpp, // TODO fix spreadsheet editability), and xlsx returns false outright — and cannot save at all (ooxml_spreadsheet_document.cpp, UnsupportedOperation). So the practical target is ods, where save already rebuilds the package from content.xml.

The reason it is off is real, and it is the sparse sheet model. parse_sheet (odf_parser.cpp) only creates a sheet_cell Element for a cell that is non-empty; empty cells are recorded as a range in the sheet's position-keyed maps with null_element_id. And a cell with table:number-columns-repeated / number-rows-repeated gets one Element per column repeat, all pointing at the same pugi::xml_node, flagged is_repeated — which element_is_editable already refuses. Both are exactly the cases where writing through the current editor would be wrong: text_set_content splices the DOM of one text run, so writing a repeated cell writes all of its repeats, and an empty cell has no text run to splice at all.

First step: the cells that already work

A non-empty, non-repeated cell already has a paragraph → text subtree, element_is_editable already walks the parent chain and clears it, and DocumentPath already addresses a cell by position (DocumentPath::Cell, "cell" prefix) rather than by sibling index — so the path survives a sheet that is mostly holes. The missing piece is mostly the document-level gate.

• Let is_editable be true for DocumentType::spreadsheet, so translate emits contenteditable + data-odr-path for those cells and html::edit can navigate back to them.
• office:value has to move with the text. A cell carries office:value-type plus a typed attribute; the <text:p> is only the display form. Typing 42 into a float cell and rewriting only the paragraph leaves the old office:value behind, and LibreOffice reads the attribute — the cell silently reverts. So a cell edit is not a text edit: it needs to re-infer the type (or drop to string and clear the typed attributes), which is the first thing that pushes past modifiedText.
• Formula cells (table:formula) have the same problem, worse: the text is a cached result. Simplest honest behaviour is to refuse, or to clear the formula and keep the literal.
• A round-trip test through soffice --convert-to is the check that the saved package is actually valid, not just that the bytes changed.

The broader solution: repeated and empty cells

Both are the same shape — the element tree does not have a place to write to — and both are fixed at write time, not at parse time. Neither should inflate the model: expanding repeats or materialising every empty cell up front is what #762 is already dying from.

• Repeated cells: split on write. Editing one cell of a number-columns-repeated="10" run means rewriting that run as up to three nodes (before / the edited one / after) and fixing the repeat counts, then repointing the affected registry entries. Same on the row axis with number-rows-repeated, which is the harder one because a row split rewrites whole <table:table-row> nodes.
• Empty cells: materialise on write (edit empty spreadsheet cells #236). An empty cell has a position and a style but no Element. The write path needs to be able to take a DocumentPath::Cell position, create the <table:table-cell> (splitting the surrounding repeat run to make room), give it a <text:p>, and register the new Element — i.e. the same splitting primitive as above, plus insertion.
• That primitive — give me a writable, unshared xml node for position (c, r) — is the piece worth designing first; both cases and any later style edit go through it. // TODO covered cells / // TODO mark as repeated in odf_parser.cpp sit next to it and are probably worth cleaning up in the same pass, since covered cells must never become writable targets.

What the diff protocol needs

html::edit accepts one thing: {"modifiedText": {path: string}}, and it hard-requires element.as_text(). An empty cell has no text element, so it can never appear in that map — the browser has nothing to attach data-odr-path to either, since translate_sheet writes empty cells as bare <td>. That means:

• the renderer must mark editable cells (not only text runs) with a path, and
• the diff needs a cell-shaped entry — value plus type — that edit can apply to a position that has no Element yet.

Worth deciding whether that is a new key alongside modifiedText or a general "modifiedCells" that supersedes the text case for sheets.

Scope

Step one is small and shippable on its own (gate + value/type handling + a round-trip test). The splitting/materialising work is the real body and only makes sense for ods until xlsx can save at all. Related: #236 (empty cells, subsumed by this), #238 (cell overflow), #762 (why nothing here may expand the model).