Rendering and observations#

Two pipelines end in one scene. A native application's is semantic application state → native page → layout → scene → optional RGBA. A web page's is HTML → DOM → cascade → layout → paint → scene → optional RGBA, run by cw-web (see architecture). Everything below the scene — hit testing, the semantic observation, the rasterizer, the font stack — is shared, and structured-only observations stop before rasterization in both. No Chromium process is involved in either.

cw-scene::Scene stores dimensions, background, revision, nodes and an optional typeface. A node has a stable ID, integer bounds, primitive, optional semantic role/label/value, interaction identifier, clipping, optional rounded_clip ({rect, radius}, honoured by hit testing), transform, z-order and opacity. A client that round-trips scene JSON must carry typeface and rounded_clip through, or it will silently lose font selection and window-corner clipping. A UiText/UiTextBold node may carry its own typeface (omitted when absent), which overrides the scene's for that node: browser pages whose theme names a font set every text node in the family the list resolves to (cw_scene::fonts::resolve_family).

The twelve primitives are Box, RoundedBox, UiText, UiTextBold, Text, AssetImage, Shadow, Image, Path, Symbol, Backdrop and Region (invisible interaction area). Symbol is a tintable vector glyph; Backdrop is a frosted-glass material. Affine coefficients use 1/1024 units; clipping is expressed in scene coordinates.

The scene supplies deterministic flow layout. Text measurement is per-primitive: UiText/UiTextBold are measured proportionally from the generated advance tables in cw-scene::metrics, with word wrapping and ellipsizing; the Text primitive (terminals, editors) keeps fixed-cell measurement and its golden pixels for plain rows. UiText/UiTextBold take two optional fields, omitted from scene JSON when unset so existing scenes serialize unchanged: italic (bool) and lang ("zh-Hans", "zh-Hant", "ja", "ko"), which picks regional Han forms. cw_scene::Style (bold, italic, lang) is what the metrics functions take; a plain bool still means bold.

Pair kerning: the web faces (Typeface::WEB — Arimo, Tinos, Carlito, Caladea, Lato, Source Sans 3, Source Serif 4, Montserrat and Playfair Display; Cousine, JetBrains Mono, Gelasio and Poppins have no kern pairs) kern as Chromium does. metrics::kern(typeface, style, left, right, size) is the adjustment between two adjacent characters in the same 1/64 px as advance (kern_units gives it in font units), and text_width, wrap, ellipsize and text::layout all add it, so the renderer, which places a run's glyphs from text::layout, draws a kerned run exactly as wide as it measures. The platform faces and DejaVu (Inter, OpenSans, Ubuntu, Roboto, DejaVu, Mono) are not kerned: the desktop shells and apps were laid out, and their golden frames pinned, on plain advances, and the bundled subsets carry no GPOS for the renderer to have applied. The web engine turns kerning off for a run with a non-zero letter-spacing, as CSS requires. Hit testing uses transformed coordinates, clipping, disabled state and z-order; later nodes win equal-z ties. It operates without generating pixels.

ScenePatch applies upserts, removals and background changes against an expected revision and produces damaged rectangles. Invalid revision/duplicate-ID updates are rejected atomically. The rasterizer can use damage to update the existing frame; compare full and incremental output in tests when adding primitives. Thirty-five font files are embedded, not one: DejaVu Sans, Sans Bold, Sans Oblique, Sans Bold Oblique and Sans Mono; regular, bold, italic and bold italic Inter, Open Sans, Roboto and Ubuntu for the per-platform shells; and regular and bold Noto Sans Hebrew, Arabic, Thai, Devanagari, Bengali, Georgian and Armenian. The CJK faces (Noto Sans SC and KR in two weights, and Traditional Chinese, Japanese and Korean locale forms), Noto Emoji, Noto Color Emoji and eight further scripts form a font pack that native builds embed and the Wasm build fetches on demand. They ship under three licenses (DejaVu, SIL OFL 1.1 and the Ubuntu Font Licence 1.0); all notices are in crates/graphics/render/assets, which is the authoritative list and documents the fallback chain, italics, colour emoji, locale forms and the pack.

UiText is laid out by cw_scene::text: a deterministic fallback chain, Unicode bidirectional reordering (right-to-left paragraphs and mixed runs), OpenType shaping with rustybuzz (Arabic joining and lam-alef, Indic reordering and conjuncts, Thai and Hebrew mark attachment, emoji ZWJ/flag/keycap/skin-tone ligatures) and line breaking between CJK characters with kinsoku. Scene metrics and the renderer share that one layout, so measured widths, wraps and ellipses are exactly what is drawn.

Emoji draw in colour from Noto Color Emoji's COLRv1 tables, painted by the renderer's own deterministic COLR rasterizer (gradients, clip boxes, transforms and composite modes), identically natively and in Wasm; the monochrome glyphs remain the fallback while the colour file is absent.

Scene layout is a deliberately smaller system than the web engine above it: it is flow layout for native pages, and it does not claim CSS, vertical or justified text or browser compositor compatibility. Native applications should be designed for that contract. Web pages are not: they go through cw-web, which is where CSS, the DOM and script live, and which is what every crates/services/* site is written against. The limits that remain — no vertical writing modes, no synthetic italics where a face lacks a slant — are listed with the supported property table in crates/web/engine/src/style/properties/mod.rs and enforced for authored sites by html::validate_strict.

Benchmarks must distinguish layout, scene patching, rasterization, readback and image encoding. See performance for reproducible commands and comparison limits; lower scene-update time alone does not establish faster full pixel production.

Raster details: source-over compositing uses straight RGBA. Image uses nearest neighbour sampling; AssetImage larger than 256x256 (wallpapers) uses a centred, aspect-preserving crop with bilinear filtering. Polygon paths support fills and strokes, antialiased. fontdue is pinned and uses the bundled fonts. Glyph/text-mask caches are bounded. The checked raster API rejects frames above 16,777,216 pixels. Text wraps and ellipsizes on measured advances for UiText, on fixed cells for Text, with fallback for unavailable glyphs. Text lays out as a terminal does (cw_scene::text::terminal): grapheme clusters in cells, combining marks in their base's cell, East Asian wide characters and emoji across two cells, right-to-left runs reordered within each row (VTE's implicit bidi mode) and script runs shaped, so Arabic joins cell by cell.

OS desktop scenes#

The optional native OS presentation composes application scenes with deterministic desktop/mobile chrome in Rust. UiText is proportional and picks its face from Scene::typeface — Inter, Open Sans, Ubuntu or Roboto per platform, DejaVu Sans as the fallback — or from its own typeface, which may also be one of the thirteen web families (Arimo, Tinos, Cousine, Gelasio, Carlito, Caladea, Lato, Source Sans 3, Source Serif 4, Poppins, Montserrat, Playfair Display, JetBrains Mono; see crates/graphics/render/assets/README.md, "Web faces"). RoundedBox supplies antialiased rounded surfaces with matching hit tests. Existing Text and its golden pixels remain unchanged. Application nodes can now have translated scene coordinates inside window content clips. The mouse pointer is part of a desktop frame: a Path glyph at the last pointer position, shaped for what is under it (see Pointer); phones draw none.