feat(html): an editing mode for a sheet, and the locks it refuses on

[andiwand]

🤖 Generated with Claude Code

Step 1.1 of docs/design/spreadsheet-editing.md — stage 0 is done, this opens the browser editor.

What the markup states

A sheet's editing is an overlay, so the page works out what it can and the markup carries only what it cannot:

where	what	why
<table>	data-odr-sheet="0"	every op names its sheet, and a view holds one
<table>	data-odr-editable="true"	so enable() can refuse before the user clicks anything
<td>	odr-locked + data-odr-lock="…"	only on a cell an edit would spoil

Three lock reasons, per decision 3:

• formula — overwriting it leaves every value computed from it stale.
• shapes — the cell's drawings are the cell, and an overlay would cover them.
• rich — a write replaces the cell's one run, so anything richer would be thrown away.

repeated is not among them: #858 made a repeated cell writable, so the lock the plan reserved for it is already gone.

A cell that can be written carries nothing at all — the cost is on the locked cells only, not on the half million others.

What the scripts do

sheet-editing.js holds odr.editing (enable, disable, isEnabled, isEditable, lockAt) and the three flat odr.on* callbacks from decision 7.

• Codes share one space with odr.onError's, appended and never renumbered — 1 is errorIllegalEditNewLine, so these start at 2.
• The message is for the console, the code is for the host. Nothing here is localised, so a host maps the code to its own catalogue; the English message is what a developer who wires nothing sees.
• The page suppresses its own repeats — an identical refusal within two seconds is dropped, so four taps on a locked cell are one snackbar rather than four.

spreadsheet.js publishes odr.sheet — cellAt(column, row), positionOf(cell), pinned(), pin(position) — and the editor addresses cells through it (decision 8, added to the design in this PR). It is the map, not a walk:

• a row is named by the label it carries, never by where it sits: sorting reorders the <tr>s, so body.rows[row] stops being row row the moment a column is sorted;
• the columns are walked with the rowspans above carried down: translate_sheet writes no <td> for a position a merge covers, so a colspan-only walk misreads every cell to the right of one;
• it is built once, by the script that owns the pin, the raise wrapper and the row order — a second copy would be a second owner of all three, and an overlay open over a cell the other script has just lowered.

Positions are the ones an op names (decision 1); the ruler's nth-child index stays private to the script, and a position a merge covers answers with the cell covering it — the one the file states. Merged sheets still get no wash and no sort control, as before.

The first two points were live bugs in the cellAt this PR previously carried inside sheet-editing.js: it indexed body.rows and walked colspan only. Neither had shipped.

One thing worth flagging

My first rich classifier assumed a cell wraps its text in a paragraph, which is odf's shape — ooxml hangs the run under the c directly. That locked 92,690 xlsx cells that are perfectly plain. Now a single paragraph is unwrapped once, and the corpus reads: 30,906 formula, 2,300 rich, 36 shapes. The remaining public rich is multiline.xlsx, which genuinely holds several runs.

Reference output

Regenerated and pinned — 355 output files (83 public, 272 private), plus resources/sheet-editing.js, a new file, and the resources/spreadsheet.js that publishes odr.sheet. Every ods page changes, since every <table> gains the two attributes and the page gains a script.

Gates: 0 of 355 changed files differ in visible text, and a fresh run against the new reference is 0 diffs in output/ and resources/.

Checks

• 1602 tests, 1596 passed, 6 pre-existing skips. 4 new, covering the table attributes, both lock reasons reachable from a fixture, and that a plain cell carries nothing.
• The position map is not reachable from odr_test, so test/browser/sheet grows two pages beside the raise checks: positions.html (a merged sheet — a colspan, a rowspan, and a rowspan past the row's last cell) and sorting.html (the same questions once the sort control has moved every row). 42 checks over the three pages, 0 failing. As a negative control the walk this PR replaced was run against the same fixtures: it misses 5 of the merged sheet's 20 positions and 4 of the 5 sorted rows, so the fixtures do discriminate.
• clang-tidy clean (the one deadcode.DeadStores it reports is pre-existing on main, identical code); gcc-15 -Wall -Wextra -Werror over four TUs. No C++ changed since — odr.sheet lives in a string literal, and fits_a_literal still holds at 14580 of 16380.

Next in the stack

1.2 the overlay editor, driving odr.sheet.pin for the arrow keys, 1.3 commit and row reflow (whose visibleRight/cutOff join odr.sheet), 1.4 undo/redo and committed(), 1.5 the browser tests and the wasm example button.