wordcraft/docs/architecture.md

WordCraft architecture

Last reviewed: 2026-10-28 · Last updated: 2026-10-28 · Change: major (crate and layer inventory, source counts, proofing languages and module boundaries re-measured from this checkout; previously minor: proof is per language: English and Polish; previously major: first committed version; plan/architecture.md is local-only) · Target: Microsoft Word (Microsoft 365) for Mac 16.113.4

How WordCraft is built today. The longer design notes in plan/architecture.md are gitignored and local to the original author's machine; this file is the committed, current summary. Rules for working in the code are in AGENTS.md.

Layers

On 2026-10-28, find crates apps xtask -type f -name '*.rs' -print0 | xargs -0 wc -l counted 333,448 Rust source lines across 16 crates, 3 apps and xtask. Cargo.toml declares workspace members with crates/*, apps/* and xtask; cargo xtask layers verified 20 workspace packages (16 crates, 3 apps and xtask) and no layering violations. The layer and line counts below are from that command and per-directory find ... -name '*.rs' ... | wc -l counts on that date. Each crate name links to its source tree.

Layer	Crate	Lines	Job
L0 standalone	control-key	87	Per-window control-port key files shared by the app and MCP bridge
L0	geom	1,271	Units (points, twips, EMU, inches, cm), rectangles and measurement parsing
L1	doc	23,719	Document model: stories, paragraphs, runs, tables, sections, styles, numbering, fields, notes, comments, revisions and equations; edit primitives and bidi
L1	fonts	7,088	Font database, metrics, outlines, HarfRust shaping, Word font substitution and OpenType MATH
L1	metafile	3,614	Shared Windows Metafile parsing and drawing primitives used by format readers and renderers
L1	proof	16,326	Fifteen spelling dictionaries: English (US, GB, CA), Polish, Portuguese, Italian, Croatian, Turkish, Russian, Czech, Indonesian and Spanish (Spain, Mexico, Latin America, Uruguay); language modules provide spelling, suggestions and grammar, while Bulgarian and German have hyphenation only
L2	docbin	8,815	Word 97–2003 .doc reader (FIB, piece table, FKPs, SPRMs, stylesheet, lists, tables, notes, pictures) over a CFB container
L2	docx	37,376	OOXML WordprocessingML reader/writer (Transitional and Strict), OMML, custom properties and preserved VBA parts
L2	formats	23,058	TXT, Markdown, HTML, RTF, ODT and LaTeX import/export through the shared document model
L2	layout	35,990	Paragraph line layout (para), page/section flow and table layout, then display-list generation (display); hit testing maps points to document positions. para owns shaping, breaks, tabs and line geometry; page flow composes paragraphs, tables, headers/footers, notes and floats.
L3	pdf	4,513	krilla PDF export: font subsets, selectable text, links, outline and tagging
L3	render	4,499	vello_cpu rasterizer: pages and thumbnails from the display list; PNG/JPEG
L4	engine	72,851	Session and editing state; command registry in cmd/; file dispatch in io; catalog, keymap, recovery, ribbon model, speech and view state
L5	mcp	1,822	MCP server exposing commands as tools; headless or bridged to the running app
L5	zotero	2,372	Zotero word-processor integration protocol answered against a Session
L6	ui-egui	81,759	Front end: ribbon, keytips, mini-toolbar, Backstage, canvas, rulers, panes, dialogs, i18n, themes, control channel and headless ui_shot example
apps	wordcraft	3,740	Desktop app (eframe + wgpu; logging, window geometry and file associations)
apps	wordcraft-cli	1,901	convert, text, inspect, run, commands, parity, mcp and zotero
apps	wordcraft-web	1,159	Browser build (eframe web runner, WebGPU with WebGL2 fallback)
tooling	xtask	1,488	Workspace maintenance and quality-gate commands, including layers

Data flow

1. Open: engine::io picks a reader by extension (docx, docbin, formats, JSON) and produces a doc::Document.
2. Edit: every user-visible action is a command id run through Session::run (ribbon, shortcuts, command search, CLI, control channel and MCP all dispatch by id). Session::run snapshots for undo (blocks are Arc-shared, so snapshots are cheap) and catches an escaped panic, restoring the document (never-crash standard).
3. Layout: layout turns the document into pages of positioned glyphs, boxes and images (display list), incrementally: an edit relayouts only what changed (1.4 ms on a 188-page document, measured 2026-10-06).
4. Draw: the UI paints pages through render (CPU raster, cached per page, pixel-aligned) on an egui/wgpu canvas; PDF export walks the same display list through pdf.
5. Save: engine::io picks a writer by extension.

Testing and gates

• 821 #[test] functions and 10 property-test blocks: model, layout, DOCX round-trip and hostile input, .doc, formats, typing parity (engine/src/tests/typing.rs), bidi, i18n catalogs, UI interaction (egui_kittest), MCP acceptance; hostile_params_never_panic runs every command with junk params.
• cargo xtask ci: fmt, clippy -D warnings, tests, asset attribution, layering, wasm build.
• cargo xtask parity regenerates parity-checklist.md; a test keeps the catalog percentage above a floor that only rises.
• GitHub Actions on PRs: packaging lint, and FreeBSD / Windows ARM64 when build files change; release builds on pushes to release.

Revision history

Date	Change	Summary
2026-10-28	major	Recounted workspace crates, layers and Rust source lines; refreshed proofing languages and documented current engine/layout module boundaries and source links
2026-10-10	minor	proof: proofing languages (English, Polish), affix engine, Polish patterns; layout and engine route each run by its language
2026-10-10	major	First committed architecture doc, from the crates' sources and AGENTS.md