feat(document)!: take an op envelope, not a map of modified text

[andiwand]

🤖 Generated with Claude Code

Step 0.4 of docs/design/spreadsheet-editing.md — the second-to-last item of stage 0.

The wire format

Document::edit understood one key, modifiedText: a map of DocumentPath → string. A sheet cell cannot be named that way — decision 1's whole point is that the cell a user types into may have no element at all. So it becomes the envelope decision 2 spells:

{"version": 1, "ops": [
  {"op": "setCell", "sheet": 0, "column": 1, "row": 2,
   "value": {"type": "number", "number": 12.5, "text": "12.5"}},
  {"op": "setCell", "sheet": 0, "column": 1, "row": 4, "value": {"type": "empty"}}
]}

• setCell names a position and carries a CellValue typed number, string or empty. number takes an optional text — without it the number spells itself, which is what CellValue(double) already does.
• setText carries what modifiedText carried, addressed by path. The plan has it becoming {id, text} with editing.md phase 1; keeping it path-addressed now means text editing does not break in the meantime.
• version must be 1; an unknown op, version, or sheet ordinal throws.

edit throws on the first op it cannot apply and leaves the ones before it applied — a host replays onto a fresh decode by construction (DocumentFile::document() decodes anew), so that is the contract, and it is now stated in the header rather than implied.

No rendered page moves

odr.generateDiff() emits the envelope instead of the map, so the browser editor keeps working. I diffed a full run against the reference tree:

• output/ — byte-identical, both repos.
• resources/document.js — the three lines of generateDiff. Not diffed by CI's compare, and the function is never called while a page renders, so it cannot move a pixel.

(resources/pdf-annotation.{css,js} already differ from the reference on main — pre-existing drift, not this branch.) The stale document.js is worth folding into the next PR's regen, which needs one anyway.

Also updated

The plan had gone stale in four places, now corrected: repeated ODS cells are no longer in step 0.2's refusal list, step 2.1 is marked landed for repeats with the open half (a position the file states no element for) called out, the "what the code gives us" table points at split_repeat/reindex_sheet and the envelope, and 0.4 notes that coalescing here is what would let a batch of ODS writes reindex once rather than per write — measured at ~0.18 µs per row node per write, so 100 writes on a 20000-row sheet is 0.36 s today.

Bindings pass the string through unchanged; their tests, the wasm test, and the three core _diff tests now speak the envelope.

Checks

• 1597 tests, 1591 passed, 6 pre-existing skips. 9 are new (document_edit_test.cpp).
• clang-tidy clean; gcc-15 -Wall -Wextra -Werror over four TUs.