Skip to content

Publish develop to main - #928

Open
github-actions[bot] wants to merge 264 commits into
mainfrom
release
Open

github-actions[bot] wants to merge 264 commits into
mainfrom
release

Conversation

@github-actions

Copy link
Copy Markdown
Contributor

Rebuilt from develop by .github/scripts/release/sync-release.sh.

The internal planning documents (PROJECT.md, AGENTS.md, CLAUDE.md, docs/INTENT.md, docs/SPEC.md) and the experiments/ folder are excluded; the source, README.md and docs/guide/ are not.

Merging this leaves main with exactly the tree the script produced.

Add IconRole::Delete = 40 to the schema and the generated IconRole enum
and tag decoder in both Protocol.gen.kt copies, draw a bin in
IconArtwork.kt (desktop and the staged Android copy), map it to the
`trash` SF Symbol in IosNavigationShell.kt, and extend
fr16_sample_icon_roles, SampleIconRolesTest and the every-system
IconButtonTest. samples/minimal now draws its delete action as the white
Delete icon on the red square instead of the word, with a test.

The schema hash in schema-hash.txt, tests/vectors/vectors.json and the
SCHEMA_HASH constants is not updated here: it needs codegen to run.

Covers FR-16 (IconRole::Delete acceptance criteria 1 to 4).
#55)

A panic in application code reached the exports' catch_unwind in no test:
the boundary tests only ever hit it through the Host's own code. One test
composes a composable that panics and expects init to return STATUS_PANIC;
the other clicks a button whose handler panics and expects dispatch_event
to return STATUS_PANIC. Each runs on its own thread under a launch lock.

Salvaged from dioxus-compose 58237cf1 (tests/boundary_panics.rs), rewritten
against the composable API and the compose_rust_host_* exports.
…-7, #55)

The existing targets fuzz the decoders alone. host_exports drives init,
dispatch_event, render_frame, release_batch and shutdown in arbitrary order
with arbitrary bytes and null pointers against a running composable
application; a valid-handshake call lets the rest reach a live Host.

Salvaged from dioxus-compose 58237cf1 (fuzz_targets/host_exports.rs),
rewritten against the composable API and the compose_rust_host_* exports.
….3.3, #914)

Material3 keeps 22dp above and below the drag handle, as Compose's
BottomSheetDefaults.DragHandle does; every other design system uses its
own sheet content padding. handle(false) leaves no gap. Criteria 5-7 name
the tests; the status goes back to Done with the commit that proves them.
… and Amper wasmJs

renderer/scripts/retry-resolution.sh reruns a Gradle or Amper build only when its output
shows a dependency or plugin resolution/download failure (3 attempts, 15s then 30s), as
the #907 loops do. build-compose.sh runs the fork's publish-local.sh and the Windows skiko
build through it, which covers every job that builds Compose from the fork (kn desktop
linux, windows static kit, packager linux static renderers). test.yml runs the wasmJs app
bundle and the Amper renderer build through it. Run 44692946847 lost the linux renderer
test to a transient 'Could not find com.google.protobuf:protobuf-java:3.25.5'.
The handle was the first row of the sheet's content column with nothing
after it, so the first line of the content sat right under the bar. The
handle's row now keeps a gap above and below the bar: 22dp in Material 3
(SheetStyle.handleGap, as Compose's DragHandle), the sheet's own content
padding in every other design system. Where a handle is drawn its top gap
replaces the sheet's top padding, so the two do not stack; without a
handle the content starts at the padding as before.

Covers FR-21.3.3 criteria 5-7: fr21_sheet_content_starts_below_the_handle,
fr21_sheet_handle_gap_is_the_sheet_padding_outside_material3,
fr21_sheet_without_handle_leaves_no_gap. The Android crate copy under
compose-rust/android-kotlin is updated to match the desktop sources.
…#301)

FR-20.10: a Host-only function that returns the index of the
app-defined width bucket the window is in (width >= b counts b),
reruns its caller only when that index changes, reads its own window,
and leaves report_size (20.8) to the app. No wire, schema or renderer
change. Approved by the leader on 2026-11-21.
current_window_breakpoint(&[f32]) -> usize returns how many of the
application's width breakpoints the window has reached (width >= b),
through a derived state remembered per call site, so the caller runs
again only when the bucket changes. It reads the window the composable
is in and does not turn on report_size.

Covers FR-20.10 criteria 1-5:
fr20_10_breakpoint_index_follows_the_window_width,
fr20_10_reader_reruns_only_on_a_bucket_change,
fr20_10_breakpoint_reads_its_own_window,
fr20_10_breakpoint_defaults_to_zero,
fr20_10_breakpoint_list_order_and_bad_values.
Not built here; the builder runs the gate.
With DXC_RESIZE_MARKER set (and neither empty nor 0), smoke_host.c sends
hand-check's marker scene instead of its default tree: the same window
(480x700, at least 320x320, resizable, the same title) and the same
records in the same order, so the Windows resize probe judges its frames
exactly as it judges hand-check's. It then answers every event and frame
with nothing and stays open until the window is closed. Its standard
output is unbuffered in that mode, because the probe's script stops the
process when it is done.

The scene is about 32 KB, so it has a 64 KB buffer of its own. The
default runs keep the 1024 byte buffer their recorded byte counts come
from, and a batch that would overflow its buffer now stops the run
rather than writing past it.

This is what lets the probe drag the GraalVM Windows renderer:
compose-rust links only the static Kotlin/Native kit on Windows, so
hand-check cannot link the GraalVM image, and the smoke host can.

smoke-host-marker-scene.test.sh holds every constant to hand-check's and
every tag and record length to compose-rust's encoder (FR-19.8.4, #687).
HostEvent and EventPayload carry #[non_exhaustive], so an event or a field
added after 1.0 is not a major version change. HostEvent::new builds an
event now that a struct literal cannot outside the crate, and every literal
in the tests, samples, benches and experiments uses it.

fr41_event_enums_are_non_exhaustive reads the sources and checks the
attribute sits on both declarations, with a self-test that the checker can
fail and a check that the compile-fail fixture names every event. The
trybuild fixtures in macros/tests/event_enums show that, outside the crate,
a match without a `_` arm, a struct literal and a pattern without `..` do
not build, and that the replacements do. Covers the FR-41 criterion on the
event enums. No wire or schema change.
Covers criterion 39.5.1 (no public builder chain outside the named value
types). The guard read only src/ui, so types `ui` re-exports from other
files (Message, KeyEvent) escaped it. It now reads every library source
under src, skipping src/bin, and counts a `pub` method as public when the
type it is on is declared plainly `pub` anywhere in the crate, a macro
metavariable, or a public trait.

It also catches `&mut self -> &mut Self`, a method returning its own type by
name, and a `pub fn` that is not at the start of its line inside a
macro_rules! body. The checks are functions over (path, text), and each one
has a self-test that feeds it a source breaking the rule.

The value-type list is the SPEC's seven plus BackgroundBrush, SemanticsState,
Message and KeyEvent, which await the owner's approval. Widening the check
found nineteen more public chaining types outside `ui`; they are listed with
a comment each so the check passes while each is decided.
Covers criterion 39.5.1 (the order follows the design rules): a widget
function takes its required parameters, then `modifier`, then the optional
ones, and `content` last when it has one. A parameter is optional when it
has a `#[default(..)]`.

The functions in src/ui/functions are read as text. A macro_rules!
repetition counts as optional when its parameters can carry their own
attributes, and every `params: { .. }` block passed to those macros must then
give each parameter a `#[default]`. `Window` is the one named exception to
having a `modifier`; its required parameters still come first. A self-test
feeds the check sources that break each part of the rule.
…-39)

Covers criterion 39.5.1 (no builder chains): the doc comments of Modifier,
the lazy list module, Separator, Window and Icon::asset still showed widget
builder chains. They now show the call macro, the form README.md and the
widget functions' docs use. IconAsset's docs no longer point at
`Icon::asset`, which is crate-private. Every example is `ignore`, so no
doctest changes.
…l 3's box by the corner modifier (FR-67, FR-68, #767)

Checkbox! takes indeterminate (property Indeterminate = 216, Bool, Checkbox
only, sent only while true) and outline_color (property OutlineColor = 217,
Paint, Checkbox only). A mixed box is drawn filled with a dash in the check
colour, is ToggleableState.Indeterminate to assistive technology, and a
press or Space asks for true. Material 3 draws it with TriStateCheckbox, or
with the shared drawing when the node names a colour, size, outline or
corner shape: ControlLook.appliesTo now counts a Checkbox's shape, so a
corner modifier alone rounds Material 3's box too.

The semantics checked field's unused value 3 is mixed
(SEMANTICS_CHECKED_MIXED, SemanticsState::mixed()); the mask is unchanged
and the constant is hashed into schema_hash().

The legacy CheckboxScope gets crate-private indeterminate/outline_color
setters only, so the call macro can reach them; the builder API gains
nothing public.

Covers FR-67 "Checkbox의 모서리, 혼합 상태, 테두리 색" criteria 1-7 and the
FR-68 mixed state. Protocol.gen.kt (both copies) is edited by hand in the
codegen form; SCHEMA_HASH, schema-hash.txt, vectors.json and the
highlight.rs tables hash still hold the old values and are regenerated by
the builder.
The probe job's last step judged the drags with a 140 line script
written inside the workflow. It is now .github/scripts/resize-probe-gate.sh,
which takes the renderer's name as run-probe-v2.ps1 labels it (kn, or
graalvm for the GraalVM image), so a second probe job can be held to the
same checks instead of to a copy of them. For kn it checks exactly what
the step checked; the closing line now also names the renderer.

windows-resize-gate.test.sh runs the script as the step calls it, and
checks that the step hands it the display scale's status and the
runner. windows-resize-probe.test.sh reads the gate's checks from the
script (FR-19.8.4, #687).
Covers criterion 39.5.2 (both forms give the same mutations and the same
group key at one call site).

The mixed-forms test said it recomposed but composed once. It now changes
each call's remembered count and the parent's state, renders two frames, and
checks that the parent recomposed, that each call kept its own count, and
that no call remembered a new state.

A new test composes two programs that differ only in the form of one call
site, and asserts that they send identical batches for the first
composition and for the frame after the state at that site and its
parent's changed.
Covers criterion 39.5.1. The guard skipped methods on any type it did not
see declared `pub`, so a type made by a macro, such as the wire enums
`define_wire_enum!` declares, escaped it. A method is now skipped only when
its type is declared in the text without `pub` (private or restricted) and
never with it; a type no text declares counts as public. The self-test has a
case for it.
Covers criterion 39.5.2 (both forms give the same mutations). Only Column
and Row were compared byte for byte. Each fr39_widget_functions file now has
a `functions` module that writes every full-parameter screen of its `macros`
module in the function form, every argument in declaration order and the
ones the macro call leaves out given their defaults, and one test per screen
asserts the two first batches are equal. Together they cover all 51 widget
kinds.
…-39)

Covers criterion 39.5.9 (a missing required argument is reported at the
macro call when there are no arguments). `Field!()` leaves out the required
`value`. The expected `.stderr` is not in this commit: it is generated by
running the trybuild test with TRYBUILD=overwrite and has to be reviewed
before it is committed, so the first run of this test fails until then.
Covers criterion 39.5.13 (the generated docs are what the public name
shows). The generated macro_rules! is `#[doc(hidden)]` and its public name is
`pub use __compose_call_X as X`. rustdoc leaves out a re-export of a hidden
item unless the re-export is `#[doc(inline)]`, so the public name had no page
and the parameter list written into the macro's docs was never shown.

The re-export of a `pub` composable now carries `#[doc(inline)]`; a private
one's does not. `ui` re-exports the widgets from a private module, and rustdoc
follows that re-export to the hidden macro as well, so it carries
`#[doc(inline)]` too.

An expansion test asserts the docs sit on the macro the public name resolves
to, that the re-export renames exactly that macro and is inlined, and that
a private function's re-export is not. A test reads src/ui/mod.rs for the
attribute on the widget re-export. Neither runs rustdoc; `cargo doc` and a
look at compose_rust/ui/macro.Column.html is the check that remains.
The resize workflow now builds the GraalVM native image renderer from
this commit and drags it the way the probe job drags hand-check:

- graalvm-fork-desktop builds the Compose fork's desktop modules with
  fork-desktop-modules.yml, under an artifact name of its own (the
  reusable workflow gains an artifact input, compose-fork-desktop by
  default), since a run holds one artifact of a name.
- graalvm builds the image with build-native-windows.ps1, runs the
  Windows smoke test, and links the C smoke host to it as
  graalvm\bin\hand-check.exe, a GUI program with the probe's
  per-monitor DPI manifest, kept as hand-check-windows-graalvm.
- probe-graalvm (windows-2022, windows-2025) drags it beside the late
  control at 100% and 200%: every edge at three rates and 25 window
  state runs, judged by resize-probe-gate.sh graalvm, with its frames
  and contact sheets kept as resize-probe-frames-graalvm-<os>.

run-probe-v2.ps1 repeats the window states for every renderer on the
new path, so the GraalVM ones can be judged by rate too, and accepts a
comma list for -Apps and -Edges. The chain does not wait for the
renderer kit, so it runs beside the Kotlin/Native one.

Covers FR-19.8.4 for the GraalVM window (#687, refs #60 #17).
…R-39)

The error is reported on the #[Composable] attribute, with the Field!()
call only as the macro invocation note: with no arguments no caller
token reaches the check, so this records the current output.
…ey (FR-39, #23)

The status note now says what PR #906 left: widgets are public functions
plus call macros, the *Scope builders are pub(crate), and ui::next existed
only in v0.0.12. The group key paragraph described Location::caller(); the
key is the hash of the function's module path and name, with a counter for
branch groups and an occurrence count for repeated keys. The byte
comparison against public builders went with 9ef0d72. The 1.0.0 date is
2026-11-28 (owner, 2026-11-21). The value-type list is left as it was.

Leader-approved wording [leader, 2026-11-21].
…-39, FR-90, #23)

The owner deferred the rsx! expressiveness check past 1.0 on 2026-11-20
("1.0 이후로 미루기"). The criterion moves verbatim into FR-90, status
Draft, with the owner's decision and the state of parity.sh. 39.4.7 stays
in FR-39 as a one-line pointer so 39.4.8 keeps its number. FR-39 stays
Agreed: the 39.4.8 measurement is still to be recorded.

Leader-approved wording [leader, 2026-11-21].
…(FR-39, #23)

Issue #23 asked for a milestone; M9 is it, with what is done (runtime in
PR #112, call macros and Copy handles from #278, builders out of the
public API in PR #906) and what remains (the 39.4.8 measurement). The
scope no longer names a Dioxus custom renderer as the Rust side, the
FR-39 row no longer points at the draft PR, and the 1.0.0 date is
2026-11-28 (owner, 2026-11-21).

Leader-approved wording [leader, 2026-11-21].
…39, #23)

Factual only: the runtime merged with PR #112 on 2026-10-17 and the
builders left the public API with PR #906; the 1.0.0 date is 2026-11-28
(owner, 2026-11-21, by way of 2026-11-23); the expressiveness criterion
D2 calls "FR-39 기준 6" is 39.4.7, now FR-90. The decision itself is
unchanged.

Leader-approved wording [leader, 2026-11-21].
…R-4, NFR-6, PR-3)

NFR-6 goes back to Agreed: 15 files under renderer/desktop/src use
@ExperimentalComposeUiApi, @ExperimentalTextApi or @InternalComposeUiApi,
not one adapter file. NFR-4 goes back to Agreed: its own record says
Linux and Windows were only built and smoke-tested, and the macOS check
predates the Kotlin/Native renderer.

PR-3 stays Done. Its "(미구현)" and "(CI 확인 전)" markers were stale:
the web page check is .github/scripts/check-sample-pages.mjs, the stopped
clock is FrameRequests.kt with StoppedFrameClockTest, and the Windows
3e step passed in ci.yml run 44692946847 (2026-11-20).

Leader ruling "status must be true" [leader, 2026-11-21].
github-actions Bot and others added 30 commits November 22, 2026 19:55
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Feat: Render real preview widget trees via desktop ImageComposeScene and measure save-to-frame (FR-54 PR 3)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
…points-panics

Feat: Checkbox mixed state and outline, sheet handle gap, window breakpoints, NFR-7 panic tests (CR-7)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Fix: Blur box shadows with a mask filter instead of a blurred layer (#976)
Chore: Route Amper modules through Google Maven mirror and retry ./kotlin test steps (#919)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
…st-batch

Fix: keep ScrollIntoView placement alive across concurrent nested scroll containers (ember#838)
Linux menus: wire K/N X11 clipboard for text menu Paste state and strengthen FR-33.1 tests (#19)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Packagers: Linux .deb/.rpm from Dioxus.toml and Mac App Store .pkg builder (#4)
Feat: Close sample reference-parity and layout defects across all 9 samples (FR-22, #2)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Fix: Remove implicit clipToBounds on WeightedStack so padding_xy matches uniform padding + spacer (FR-70, #982)
Web symbol fallback for Monospace/Serif/named families + macOS #180/#146 regression guards (#525, #180, #146)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Feat: Add Host-only current_node_breakpoint over remember_node_size (FR-28.1, #967)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
…nothing

Fix: Send no ButtonKind record for a button that names no kind (FR-30.1)
Fix: Send a CodeEditor's scroll reports in one frame and animate ScrollTo the right way (FR-80)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md
CI: Read the .msi back and require WACK's DPI check on the .msix (#3)
Generated by .github/scripts/release/sync-release.sh. The internal planning documents
and experiments/ listed below are kept on develop only:
  PROJECT.md
  AGENTS.md
  CLAUDE.md
  experiments
  docs/INTENT.md
  docs/SPEC.md
  docs/platforms.md

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant