feat(odf): write a repeated sheet cell by cutting the run

[andiwand]

🤖 Generated with Claude Code

What it does

Sheet::set_cell refused a repeated .ods cell — one element stands for every position the repeat covers, so a write would have hit all of them. It now cuts the run:

<table:table-cell table:number-columns-repeated="4">x</…>
                        ↓  set_cell(2, 0, "y")
  repeated="2" → x  |  the one written → y  |  repeated="1" → x

split_repeat copies the table:table-row and the table:table-cell around the position, and leaves the original node as the one written — so its element, its paragraph and its run all survive, and the write goes through them unchanged. Only the copies need parsing.

reindex_sheet then rebuilds the sheet's position index off the dom. A cell node keeps the element it already carries (matched on the node), a copied one gets a fresh one. That replaces surgical insertion into the flat cells vector with its per-row first_cell offsets, which is the part that would have been easy to get subtly wrong.

It does not reintroduce the expansion

The cost follows the row the write touched, not the grid the repeat claims. A write into a 1048576 × 1024 repeat — #776's bomb — leaves the registry under 32 elements, with the neighbouring positions still reading their old value. A test pins exactly that.

Two costs worth stating, and they are in odf/AGENTS.md: cutting a repeated row copies every cell in it, so elements grow with the row rather than with the repeat; and the reindex walks the row nodes, which a repeat collapses, so it is bounded by the dom rather than the grid.

Refusals stay clean

The formula and markup checks now run before the split, so a refused write leaves the run uncut — the property #856 established, which the obvious ordering would have broken (splitting, then throwing, rewrites the file for nothing). only_text_run splits in two for this: holds_one_run decides without writing, text_run_of creates the run an empty paragraph needs.

A cell the file states no element for still refuses; that is materialising a missing <table:table-cell>, a separate thing.

Why it needed #857 first

A handle taken before the write keeps naming its position and re-resolves to the new element, so it shows the write:

const SheetCell held = sheet.cell(2, 0);
sheet.set_cell(2, 0, CellValue("y"));
held.value().text();   // "y"

Without positional ids that handle would silently have come to mean the left part of the cut run.

Checks

• 1588 tests, 1582 passed, 6 pre-existing skips. 8 are new: column split (middle and both ends), row split, handle-follows-the-write, save/reopen, refusal-leaves-it-uncut, no-expansion.
• Reference HTML output byte for byte unchanged. clang-tidy clean; gcc-15 -Wall -Wextra -Werror over five TUs.
• LibreOffice oracle, by hand on a real .ods: three repeated cells written, saved, and both the original and the result converted with soffice --convert-to csv. 121 lines each, and the only line that differs is the one holding the three cells written — nothing else moved. Not committed, soffice is not in CI.