feat(odf): a write takes the cached result of what reads it away

[andiwand]

🤖 Generated with Claude Code

Rebased on main, now that the rest of the step-3 stack has landed. One commit.

Step 3.4 of docs/design/spreadsheet-editing.md.

The problem

An edited input leaves every formula reading it computing an old number, and the saved .ods went on showing the result its producer cached. An .xlsx has a switch for exactly this — calcPr/@fullCalcOnLoad, which every save already sets. ODF has none.

What it does

So the result goes instead. drop_stale_results asks SheetDependencies which cells read the written position — directly or through another formula — and takes the attributes stating a value off each. The text:p showing it goes as an element rather than as a node, so the registry keeps no dangling one.

<!-- before --><table:table-cell table:formula="of:=SUM([.A1:.A2])" office:value-type="float" office:value="3"><text:p>3</text:p></table:table-cell>
<!-- after  --><table:table-cell table:formula="of:=SUM([.A1:.A2])"/>

The formula, the cell's style and a drawing anchored in the cell all stay. A cell stating a formula and no result is one a reader has to compute, and none can show a wrong number for.

A formula the graph could read no position out of keeps its result. Nothing links it to the write. Document::unresolved_formulas() is what names them.

The cost until step 4: a cleared cell renders empty here too. That is the trade the design accepted — an empty cell says "not computed", a stale one says nothing.

The spike, as far as a CLI can answer it

LibreOffice's --convert-to recomputes whatever a file cached — a file whose cached result I hand-edited to 999 converted back to 3 — so it cannot show what a reader that does not recompute would display, and setting ODFRecalcMode to "never" in a scratch profile changed nothing. ODF states no recalculation switch at all, so the rule the file itself can carry is what is left.

End-to-end check: a real LibreOffice-written .ods, edited through Sheet::set_cell and saved by us, opens as a valid package and LibreOffice computes the cell we cleared (10 / 2 / 12, where the file states no result for the third).

Test

test/src/internal/odf/odf_sheet_stale_test.cpp, 8 cases from inline fixtures: the direct reader, the chain, a formula reading something else, a reader on another sheet, what the cell keeps, a drawing anchored in a cleared cell, the saved file, and a write nothing reads.

Reference output: this PR emits no different html — it changes the write path only. compare-html stays red until the output repos take the commit #887 still owes them.

https://claude.ai/code/session_01VVmjmddv2Ui17Nptc1ggue