fix(odf): address a repeated sheet cell by position, not by index

[andiwand]

🤖 Generated with Claude Code

The bug

One element stands for every position a repeat covers, so an index cannot say which of them a handle means:

SheetCell c = sheet.cell(5, 0);   // run anchored at column 2
c.position();                     // → (2, 0)
sheet.cell(5, 0) == sheet.cell(2, 0);  // → true

DocumentPath inherits it, since extract_path spells a cell by sheet_cell_position. This is the half of the repeated-cell change in v6.12.0 (#776) that a consumer could observe — it was marked Breaking at the time and justified with "spreadsheets are not editable", which #856 has now retired.

The fix

sheet_cell(sheet, column, row) hands out a positional id for a repeat:

tag(1) | ordinal(15) | column(24) | row(24)

the ordinal counting sheets in document order. A cell standing for one position alone keeps its index, so only a repeat pays for the indirection and the element store stays a flat array.

The decode is one hook — ElementRegistry::resolve_id, identity for every engine that does not shadow it. odf's shadow looks the position up in the sheet's existing run index (a binary search that was already there), and every payload accessor goes through it.

Two properties fall out:

• RegistryElementAdapter navigates entirely through element_at, so all navigation works untouched, and element_parent drops the position for free — a sheet's stored parent_id is a plain index.
• The id resolves through the index rather than pinning the element it found, so a handle follows the index. That is what will let the next step split a run under a handle without invalidating it.

extract_path/navigate_path needed no change: they already go out by position and back through sheet_cell.

The position stops at the cell — one paragraph and one run stand for every position — so a path into a repeated cell still names the anchor, and round-trips. A test pins that boundary.

One line goes with it: a cell's style resolved at the payload's position, which for a repeat is the anchor's column, and table:default-cell-style-name is a column's to state. Not reachable today (SheetCell exposes no style(), and children reach it through the plain anchor id), so this is consistency rather than a fix.

Not xlsx

Sheet.cells there is one entry per real <c>, each with its own element and its own true position. Neither bug exists, so nothing is added to it.

Checks

• 1582 tests, 1576 passed, 6 pre-existing skips. 4 are new. The reference HTML output is byte for byte unchanged.
• clang-tidy clean; gcc-15 -Wall -Wextra -Werror over seven touched TUs.

Next

Splitting a repeated run on write, which makes Sheet::set_cell accept a repeated cell instead of refusing it. It needs the DOM row/cell split plus a parser entry point for re-registering the copies, and it is safe only because a handle re-resolves — hence this first.