Rendering parity (software ↔ GLES). The software backend now rounds glyph pen positions and image destinations to the nearest integer pixel, matching what the GLES backend already did; previously it truncated, so text and 1:1 images could land up to half a pixel off between the two backends and the bilinear sample read ~1 px softer than the source. Gradients and shadows are deliberately left unimplemented on the software backend, and the GLES multi-rect `glScissor` clip is left coarse on purpose: making it exact would need stencil bits the EGL config does not carry, or routing the partial-redraw path through the offscreen clip layer, which would break its `fill` / `clear_rects_transparent` scissor semantics. Adds software-backend pixel tests covering the snapping.
Font resolution unification. The system-font candidate chain, `find_font_opt` and `load_default_font_bytes` lived in two copies (`render/helpers` and `gles_render/helpers`) that had already diverged — one resolved through `find_font_opt`, the other inlined the candidate loop — and now live once in `system_fonts`. The two per-backend `OnceLock` default-font caches and `primary_handle` collapse into a single `system_fonts::default_handle`. Module docs and the `font_registry` caller are updated accordingly.
Shared image validation and rect inflation. `draw_image_data`'s dimension check and its one-line warning were byte-duplicated across both backends and are now `render::helpers::validate_rgba_dims`. The six manual symmetric `Rect`-inflate literals in the GLES primitives reuse the existing `Rect::expand`.
FrameState and DrawCtx de-duplication. The eleven `SurfaceState` fields the draw pass owns and threads through `DrawCtx` — `widget_rects`, the cursor / selection maps, the scroll state, `accessible_extras`, `prev_focused` / `prev_hovered` / `prev_pressed` — move into a `FrameState` sub-struct. The per-frame `build_draw_ctx` / `commit_draw_ctx` helpers can then borrow `&mut ss.frame`, disjoint from `ss.canvas` and `ss.pool`, so the four frame paths (software / GLES × full / partial) replace their duplicated `DrawCtx` construction and write-back with a single helper call each. A whole-`SurfaceState` borrow could not express this (partial borrows do not cross function boundaries), which is why the helpers take the sub-struct. `content_dirty` stays on `SurfaceState` — it is an invalidation flag, not frame state — and is reset at the call site.
rich_text tests. Adds the previously-missing `tests.rs` for the `RichText` widget: one hit rect per visual line a link spans (the widget's core invariant), the single-line and no-link cases, preferred-size growth with hard line breaks, and `map_msg` range preservation — all headless against a software `Canvas`, with line counts forced by `\n` so they do not depend on any system font's measured width.
Documentation. Fills the rustdoc gaps on the embedder-facing surface: `core::UiSurface` accessors, the `egl_context` public API, the `GlesCanvas` methods, `theme::typography` and `theme::error`, and the `RichText` / `Text` builders; adds a crate-level "Rendering backends" overview. `CHANGELOG.md` is added (0.2.0 / 0.1.0), and `docs/widgets.md` / `docs/cookbook.md` gain `rich_text`, `external`, and the CPU-draw / path-clip / externally-laid-out-tree recipes. `debian/changelog` gets the 0.2.0-1 entry. Private intra-doc links to `system_fonts` are demoted to code spans so `cargo doc` is warning-free.
Packaging. The `libltk-dev` registry crate shipped a `Cargo.toml` declaring the `lookup` bench while the `Makefile` install copied only `src/`, so Cargo refused to parse the manifest over a missing `benches/lookup.rs`; the install now ships `benches/` as well (the file alone satisfies the parse — criterion is a dev-dependency and is not resolved when the crate is consumed as a library). `Cargo.toml` is bumped to 0.2.0 to match the package version and the `ltk-0.2.0` registry directory.
Add the primitives rustdroid needs to project an Android view hierarchy onto an ltk surface, all kept general rather than Android-specific.
Canvas gains arbitrary vector path fill and stroke: `Canvas::fill_path` / `stroke_path` over a new `PathCmd` command list (MoveTo/LineTo/QuadTo/CubicTo/Close in surface coordinates). The software backend rasterises directly with tiny-skia; the GLES backend rasterises into a tiny-skia pixmap and blits it (CPU fallback, no GPU path shader). This is what renders an Android `Path`, a `VectorDrawable`, or a Lottie frame.
`ExternalSource::Cpu` (and the `External::cpu` constructor) adds an immediate-mode CPU drawing closure, invoked once per frame with the canvas and the widget's laid-out rect, working on both the GLES and software backends. It hosts a custom `View.onDraw` straight onto the ltk canvas without a GL texture round-trip, unlike the existing `Texture` source which only renders on GLES.
`Stack::push_placed` appends a child at an exact rect, bypassing alignment and intrinsic sizing. This lets a view tree whose geometry is computed elsewhere — Android's measure/layout pass, which yields an absolute rect per view — be projected onto a Stack in paint order.
New `RichText` widget: wrapped paragraph text carrying a message per clickable link range, with the layout pass emitting one hit rect per link line so taps land on the link rather than the whole paragraph. It is the ltk side of an Android `Spanned` carrying `URLSpan` / `ClickableSpan`.
gesture: only drive the horizontal pager once the axis locks horizontal
`on_move` emitted horizontal swipe progress whenever `swipe_axis != Vertical`, which includes the pre-lock window where `swipe_axis` is still `None` (the first ~8 px of travel). But `horizontal_drag_started` only flips once `dx.abs() > 8`. A vertical gesture that opened with a few pixels of lateral drift — a swipe-up to the launcher, a scroll — therefore emitted a tiny horizontal progress sample, which armed the consumer's pager, then locked vertical without `dx` ever passing 8, so `horizontal_drag_started` stayed false. On release the horizontal branch was skipped, no `on_swipe_horizontal_progress(0.0)` fired, and the pager stayed stuck active — on the crustace homescreen that froze the surface (the main stays motion-only behind stale page subsurfaces) until an unrelated gesture reset it.
Emit horizontal progress only once `swipe_axis == Some(Horizontal)`. Locking onto the horizontal axis already implies `dx.abs() > 8`, so `horizontal_drag_started` is set in the same step, restoring the invariant that a frame which drives the pager always has a matching release event to settle it. The cost is that the first ~8 px of a horizontal drag no longer move the page, which is the same deadband the axis lock already imposes on the vertical panels. Adds `pre_lock_lateral_drift_does_not_drive_horizontal_pager`.
