A RecordShift moves points, rects, clip rects and AccessKit bounds, and
nothing else: stroke widths, radii, blur widths, angles, opacity and
texture coordinates now compare exactly, so a replay that samples a
neighbouring atlas region or keeps a stale opacity is a difference.
Imports from core and alloc where clippy asks for them.
Records moved by a whole pixel from positions off the float grid differ from the part run at the new place by
rounding only. Positions, sizes, widths, angles, texture coordinates and AccessKit bounds compare within the given
tolerance; ids, counts, colours and text stay exact.
A new value whenever the atlas is made, cleared or rebuilt with new options, unique in the process. Kept paint that
holds text (a replayed record) is valid only under the generation it was made in.
A visible Ui's registration of a widget that only invisible (sizing-pass) Uis registered so far replaces the widget
rect and takes its place at the top of its layer. Widget rects carry when they took their place, and
pass_records_since finds a part's widgets by it, so a part holds the widgets a sizing pass registered before its
mark. Records leave out Shape::Noop shapes, layers left empty, and widgets neither visible nor enabled.
PassRecords::shifted gives the records as a replay at a RecordShift adds them. Ui::new disables before it registers
its widget, as new_child does.
replay_pass_records takes a RecordShift: shapes, widget rects and AccessKit bounds move by its
offset in layer space; a clip rect equal to the clip the part ran under becomes the new clip, and
one set inside the part moves with it, cut by the new clip. PassRecords::paints_callback says
whether records hold a paint callback, which a replay adds only where it was made.
`pass_records_since` gives shapes and widget rects with their layers sorted by order and id, since the maps they
come from have no stable order. `first_difference` reports a paint callback on either side as
`PassRecordsDiff::Callback`: a backend wraps every callback in one type, so neither its type nor its rect tells
whether two callbacks paint the same. Its doc says what the records cover and what they don't, and `Display` shows
the variant, index and rects only. PATCHES.md lists the pass records (E36), `HarnessBuilder::with_context` (E37) and
this comparison (E38).
Shapes layer by layer, then widget rects, AccessKit nodes and AccessKit parents compare with `==`, except a paint
callback, which compares by its rect and its callback's type: each run makes a new callback, and `PaintCallback`'s
`PartialEq` compares the callback by pointer. `PassRecordsDiff` gives the kind, the index and the record on each side.
Popup::show makes its area a sublayer of the layer it opened from when both share an Order (tooltips aside). Areas::end_pass follows each layer with its whole sublayer tree, depth first, and set_layer_order moves a listed layer with all of its unlisted descendants.
egui-winit sends it from State::new and on WindowEvent::ThemeChanged; eframe web
at start and on the prefers-color-scheme change. RawInput::system_theme keeps
the same value; the event is the change notice.
AppRunner forwards PlatformOutput::retired_ime_resets, and a reset the text
agent cannot complete (descriptor no longer matches, local replacement failed)
is reported retired in the next raw batch, with a repaint, as complete_reset
does for a completion.
egui-winit's ResetBridge reports every request it drops (focus loss, window
destroyed, IME disallowed, another IME context, a newer request), and
PlatformOutput::append reports a request a later pass drops in
PlatformOutput::retired_ime_resets, which egui-winit forwards. A custom editor
waiting for the completion asks again instead of waiting forever.
Add ScrollArea::horizontal_scroll_bar_label and vertical_scroll_bar_label.
The scroll bar AccessKit node had no name, which fails WCAG 4.1.2. The
labels are per axis because a name cannot be composed from a shared
string in every language, and egui has no built-in UI text.
A composition belongs to the focused IME context it began in. An editor
that comes back after a pass it was not shown in, or gains the focus,
starts a new context; its old composition ends there, the pre-edit text
kept in the buffer as typed and the cursor at its end. `preedit()` is
then `None`, so a widget wrapping `TextEdit` no longer sees stale
marked text in a field refocused in the pass it reappears.
The IME composition under way: the character range of the pre-edit
(marked) text in the buffer and that text, or `None` when nothing is
composing. While composing, the buffer holds the marked text beside the
committed text; a widget that wraps a `TextEdit` and reports only
committed text removes this range, and shows the same buffer in the next
pass so the composition goes on.
Enter and Space both act as a click on a focused widget. With
`activate_on_space: false`, only Enter does, as for a link: Space
doesn't activate a link, it scrolls the page in a browser. Defaults to
true, which keeps the current behaviour.
Requests focus together with the event filter the widget keeps while
focused. `set_focus_lock_filter` only applies once the widget had focus
for a pass, so a widget that moves focus to a sibling on an arrow key
(a radio group) lost it to egui's directional focus when a second arrow
key came in the pass right after the request.
`Ui::interact_opt` registers the widget as enabled only if both the `Ui`
and `InteractOptions::enabled` are. A widget can then draw its own
disabled look without a disabled child `Ui`, which would fade its
painter, add a node to the accessibility tree and shift the auto-ids of
the widgets after it.
TextFormat::font_features lists OpenType features (tag, value) that the
shaper applies to that section's runs, e.g. tnum for tabular figures.
The list is empty by default, which shapes exactly as before, and is part
of TextFormat's hash, so the galley cache keeps text with and without
features apart. RichText gains font_feature() and font_features().
Tests use a 3 KB digits-only subset of Space Grotesk (OFL), whose default
digits are proportional.
Preserve the reviewed 0.37 source tree and canonical dependency pins while
recording the existing Craft fixes from craft/0.36.2 in its ancestry.
All runtime files remain byte-identical to the published migration commit.
A zero-width glyph joined the block before it and took that block's level,
so an isolate initiator, PDI or RLM could no longer separate two runs: two
Latin isolates in right-to-left text read as one run, and a Hebrew isolate
followed by a Latin one came out left to right. A zero-width glyph now joins
the block before it only at the same level.
Port of emilk/egui#8657 (draft) to 0.36.2, the companion of W15 (rust-windowing/winit#4729): winit now binds a wl_data_device per seat for file drops, and a second device from smithay-clipboard loses the selection on compositors that serve only the first device per client (Hyprland). With winit's clipboard available, smithay-clipboard is not started. Only the clipboard.rs change is taken; the PR's temporary [patch] of egui's own winit is not.
On 0.37, `Clipboard::new` stays `unsafe`. While winit's clipboard is in use, the
PRIMARY selection goes through arboard: W15 offers only the regular selection.
storytold/winit craft/0.30.13 (W1–W15) with the language-support
patches on top (W17–W27). It has the same IME, Latin-layout and
delete-surrounding APIs as craft/suite, and W15's Wayland clipboard
(platform::wayland::Clipboard) that E4 uses.