Skip to content

Sync with expo/exact a107143a1 - #10

Merged
janicduplessis merged 2589 commits into
mainfrom
@janic/sync-upstream
Nov 15, 2026
Merged

janicduplessis merged 2589 commits into
mainfrom
@janic/sync-upstream

Conversation

@janicduplessis

@janicduplessis janicduplessis commented Nov 15, 2026 •

Copy link
Copy Markdown

Description

Merges expo/exact main at a107143 into the fork's main. It's a merge, not a rebase, so the fork's history stays and nothing is force-pushed. Merge this PR with a merge commit, not a squash; otherwise the next sync starts from the old merge base.

Our fixes

Kept:

Superseded by upstream:

Conflicts resolved

  • kernel/tables/schema.json: upstream used prop ids 300 to 309, the ids our fadeDuration and loading had. Ours are now 310 and 311, and the digest in kernel/src/wire/codec.rs is recomputed (0xc44798a91fd796dc). Wire ids move, so an app must be rebuilt against this checkout. Plans baked against the old one don't load.
  • host/web-js/rt.js: upstream split it, and PropHooks now lives in tree.js. The bound-placeholder hook and the image fade glue moved there unchanged, reformatted in the new file's style.
  • host/web/src/element.rs, host/web-js/src/emit.rs, contract/lower/src/lint.rs: upstream's versions, plus our placeholder background, FadeDuration/Loading attributes and rows::lazy choice, and without the lint hint that refused loading on an image.
  • host/apple/Sources/ExactKit/IOS/NavigationBarIOS.swift: upstream's makeNavigation(first:), with RouteNavigationBar as the bar class.
  • host/apple/build.mjs: upstream's entitlement lines (grantEntitlements, signInEntitlement), plus iCloudKeyValueEntitlement. The file stays at its 1,500-line cap.
  • scripts/app-apple.test.mjs went over its 1,500-line cap, so host.ios.iCloudKeyValue signs the iCloud key-value store entitlement #6's test moved to host/apple/presence.test.mjs.
  • host/apple/tests/ExactKitTests/NavigationBarItemsIOSTests.swift: keeps both upstream's shortcut test and A touch on an authored header control under the iOS bar reaches it #5's touch test.

Test plan

  • The pinned versions haven't changed: Bun 1.4.2, Rust 1.97.0, and the web nightly nightly-2026-08-21. Linking the wasm web artifact on a Mac now also needs that nightly's llvm-tools component (scripts/cargo-check.mjs). A checkout also needs bun install --frozen-lockfile for the new workspace packages.

  • bun scripts/caps.mjs: within every cap. cargo fmt --check and cargo check --keep-going are clean.

  • git diff --stat origin/main HEAD (expo/exact's main) shows only the fork's delta: 21 files, +297 −29.

  • cargo test -p exact-kernel --lib wire::codec: passes with the new digest.

  • bun host/apple/build.mjs --test --ios --only ContentOwnersIOSTests,NavigationBarIOSTests,BorderParityIOSTests on an iPhone 18 Pro simulator (iOS 27.0) covers A repeating gradient shades each stripe only where it falls on Apple hosts #3's hatch gradients, A sheet takes its route's own color-scheme on iOS, so its bar, title and close button match the surface it paints #4's assertions and A touch on an authored header control under the iOS bar reaches it #5's touch test. All pass except NavigationBarIOSTests.testAHeadingsAuthoredTypographyIsTheBarsTitleFont. It also fails with makeNavigation reverted to a plain UINavigationController(), so A touch on an authored header control under the iOS bar reaches it #5's bar class doesn't cause it.

  • bun test host/apple/presence.test.mjs: host.ios.iCloudKeyValue signs the iCloud key-value store entitlement #6's test passes. One upstream test (Mac extension refusals name the ad hoc escape…) fails on this machine because its keychain holds Apple Development identities from several teams. That is local signing state, not this merge.

  • The full bun test run was stopped partway because the machine was nearly out of disk, so it isn't evidence either way. The failures it showed before stopping were 5 s timeouts in bake and new-app scaffolding tests. Those suites don't touch the fork's files.

  • An app built against this branch with A scroll container zooms its one child by pinch on iOS #11 on top passes all 44 of its web contract tests in Chrome. That took four changes on the app's side, all from upstream: app.tamber renamed to app.forma with a forma header, a re-resolved Cargo.lock, contract::rust_entry taking &compat.inputs, and an ibex2-sqlite dependency in the Apple crate. On iOS, upstream's route clearance (expo/exact@53d567920) moves content that sat under a transparent bar down by the bar's height. An app that lays itself out under the bar needs -exact-safe-area="top" on the route, and its own top padding.

  • CI: both the macOS and Linux jobs pass.

amandeepmittal and others added 30 commits November 14, 2026 02:10
tools/gen.ts writes a route for each section page of the docs into app.contract on every
build, and main's docs have gained five since it was last committed (Apple bundle metadata,
Apple privacy manifests, Contract evaluation limits, Grouped section shadows, and the
pitfalls page on component calculations), so each build left app.contract changed. These
are the lines the generator writes at this commit.

Lands expo#452, regenerated on current main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e cache (expo#573)

The HTTP cache stored a reply when its exchange completed, whatever had been asked since. A read asked first and still in flight when an equal refresh completed stored its older reply over the refresh's, and the next equal read was answered with it (LLP 1162 D4: a forced ask replaces the entry).

An exchange now does not store when a replacing request for the same key was asked after its own request and before it completed, whether it was running, queued or behind the admission fence when the refresh was asked. Its reply still answers its own caller, a 304 included. An equal launch preload still in flight has its cache copy withheld the same way. A withheld store collects what it retired, and a waiting request carries a flag, not a copy of its key.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ays either side, else the date with its weekday (LLP 1168.000 D5)

Today, Yesterday, Tomorrow; the weekday two to six days either side; else CLDR's MMMEd (Thu, Oct 15), with the year
when it is not now's (yMMMEd: Sun, May 4, 2025). The compiler bakes both forms per format locale, their month and
weekday names, and the weekdays alone, each beginning with a capital as a line does: the locale's first letter
upper-cased at extraction (Mercredi, Jeu. 15 oct.), so no host maps case. The runner, the JS target and Lean agree;
the oracle rows (29 locales, tr among them), an en-US boundary table, the JS Intl sweep, the corpus and the difftest
generator cover it. The plan format digest moves: a runner before this does not print the mode.

Co-authored-by: ccheever <56719+ccheever@users.noreply.github.com>
…he dates and chat-header recipes

Co-authored-by: ccheever <56719+ccheever@users.noreply.github.com>
…written in September's digits (dz); dz in the oracle

Co-authored-by: ccheever <56719+ccheever@users.noreply.github.com>
A transcript that loads older pages at its start stopped when a fling
reached its first row before the page landed. No anchor holds a port
at offset 0 (CSS's ScrollAnchor), so the reader stayed at the top,
inside the edge band, and reachstart stayed disarmed until they left.

A page at the start is new rows before the kept first row, or new rows
between a leading row that stays first and the kept old second row, or
the first rows given to a leading row that was alone: the mirror of the
end's rule for rows put past the last. The start then owes a re-arm,
paid at the first report whose window is measured, so estimates cannot
fire a page, and spent when the start fires or the list empties. An
arrival at the end, a reorder, an identical or empty answer, a replaced
first row or a window that slid re-arm nothing. Only a list with a
reachstart handler checks, and only the new rows are hashed. Both
runners keep it; the JS target's paging-start plan pins it end to end.

LLP 1168.000 D1.1; LLP 1010 (the 2026-09-29 ruling's mirror).

Co-authored-by: ccheever <56719+ccheever@users.noreply.github.com>
Restart owned servers between tests, across files and verify passes. Replay
exactly the chosen test with its resolved launch facts before relaunch. Keep
original capture evidence; authored replay screenshots have separate paths.
Capture first screens sequentially, with an empty server before each width.
Retain every unexpected server exit across restarts, including between
large-text tour tests. Direct tests and verify read the record after cleanup,
report every death, and fail even when all tests passed on a replacement.

Retain storage, ownership and the port turn until shutdown is verified.
A failed database deletion leaves stop retryable and its exit handler attached.
Refuse a nonempty POSIX group with no surviving observed member. On Windows,
track ownership and kill attempts by PID and creation time. Adopt replacements
through live owned parents, rejecting descendants older than their parent.
Revalidate each identity before killing it alone, leaves first; observe again
between kills for new children. Bound each shutdown phase, including a loop of
replacement workers, by its deadline; failed cleanup retains ownership.
Skip CIM rows without creation dates, wait for
a new root to appear, and kill the child if identification fails. Signal exits
fail startup and the current test. Signal handlers exit even when cleanup
throws, and the exit listener catches cleanup errors.

The guides and help describe per-test isolation. Only test and agent --test
share an ancestor’s server; verify refuses an answering local server.
The issue stays open for an owned process supervisor on POSIX, a Job Object
on Windows for unobserved descendants, respawning launchers and PID reuse
between validation and taskkill, and native Windows validation. Native
capture ordering and Windows process operations have substituted-host tests.

Part of issues/20261113-test-server-cleanup-can-miss-unobserved-descendants.md: not run on a device.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Carry the composed side permutation through Apple’s measurement ABI and host blocks. Apply it in the kernel’s shared block-style path so trial measurement and final layout agree.

Include the tactic in Apple’s measurement cache and selected placement. Cover asymmetric dimensions, constraints, percentage margins and combined flips in both document directions, and remove the guide’s equal-dimensions workaround.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A `console` line went to the runner's journal, which the agent's `logs`
reads, and nowhere else: nothing in Console.app, `log stream` or Xcode,
and nothing a tool watching the system's log could find (LLP 1027.009).

Each line the app's code logs now also goes to the platform's log as
the journal takes it:

- Apple: the unified log, subsystem the app's bundle identifier,
  category `console`, public; `console.debug` at debug, `console.error`
  at error, the rest at the default level.
- Linux and Windows: standard error, as `console: …`.
- Android: logcat under the tag `exact-app` at the method's priority,
  in a development build only.

Apple, Linux and Windows write them in every build.

The method that wrote a line is kept from `console.js` to the host. A
source hands the runner `console.<method>: …`; the runner journals
`console: …` as before and hands the line and its method to the sink
its host installed (`Runner::set_console`). No batch carries the lines,
so one the presenter skips or drops loses none.

A host installs the sink when its runner becomes the one presented: on
Apple through `exact_set_console`, in `presenter_hooks` and the
viewless boot; on Linux and Windows when the presenter takes a runner,
past its last refusal. Lines journaled before that are held and written
first, in order. A candidate that is never committed, and a module that
is dropped before a session adopts it, write nothing. A host that
installs no sink (the web, the terminal) keeps the lines in its journal
only.

What the runtime itself says through `console` (a refused call's
message, an option it does not support, an error a callback did not
catch) is journaled with the text it had and reaches no sink: the
author did not write it. A widget's refresh takes the module's lines as
each answer and reply ends, where it took none.

Android's lines come through the same sink. Main's b087ac3 wrote a
development build's lines by reading the journal back with one count
for the process, which skipped a reloaded or updated plan's first
lines; the journal's echo is `EXACT_LOG_JOURNAL`'s alone again. An
Android release build writes none, and no Android device ran the path
(QUEUE.md).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ion, and verify reads the accessibility tree

Implements LLP 1166.000 (stages 1 and 2), LLP 1166.001 (stage 1) and LLP 1166.002. LLP 1166
is the research behind them: what holding Exact against Vercel's UI SDK found.

A painted group's keys are decided once (LLP 1166.000). The arrow keys of a `radiogroup`, a
`tablist` and a `menu` were written separately in the web glue, the presenter and the Apple
host, and the hosts disagreed: the Linux host gave a radio group of buttons no keys, the Mac
gave a tablist it paints none.

- `kernel/tables/schema.json` gains `roles`. `Kernel::role_step` and `role_tab_stop` decide
  a group's members, eligibility, where an arrow, Home or End goes, and the Tab stop.
- The presenter and the Apple host call them (Apple through one ABI entry,
  `exact_role_keys`). The web keeps `input-glue.js`.
- One file of 193 cases (`kernel/tables/role-key-cases.json`) runs against the kernel and
  against the real web glue in Chrome.
- `lower-role-keys` warns for a composite role Exact supplies no keys for (`listbox`,
  `tree`, `grid`, `toolbar`, `menubar`) when nothing in it handles keys and no element names
  it with `aria-controls`.

A field names its active option (LLP 1166.001, stage 1). `aria-activedescendant` is carried
to the DOM and reported by `tree --ax`. `docs/recipes/combobox.md` is the composition; the
docs site's search and the harness's completion list use it. Native hosts carry the prop and
read nothing from it yet.

Names and roles are checked (LLP 1166.002). `verify` reads each judged screen's
accessibility tree once on the web and iOS, and counts an unnamed control, an element exposed
while hidden and one reachable outside a modal as `check`'s findings are. A tree that was not
read whole is said on the last line and in `--json`'s `axUnread`. A literal `role` ARIA does
not have is an error, `lower-unknown-role`, with the nearest role named.

Behaviour that changes for apps:

- Linux, Windows, Android TV: a painted radio group of buttons moves and presses by the
  arrows and is one Tab stop. A menu painted in the page moves by Up, Down, Home and End.
- Mac, iOS: a painted tablist and a page menu move by the arrows. An `aria-disabled` radio
  is skipped. A member with a written `tabindex` is a Tab stop only as its group's stop.
- Web: a member that is inert or not shown is skipped. Home and End from an ineligible
  member go to the first and last eligible one.
- Compiler: `role="tabpanle"` no longer compiles. A `listbox`, `tree`, `grid`, `toolbar` or
  `menubar` with no key handler warns.
- `verify`: one more group in the report, and a longer last line when a tree was not read.
  Exit codes are unchanged.

Brought onto main as one commit. Main had moved the presenter to `host/painted` (its
`tabs.rs` and the two new role test files go there, and `note_press` there lost its key
argument), and changed `verify`'s help, imports and `tree --ax`'s modal rule, which already
holds what this branch changed in `agent-ax.mjs`. Two files went over the 1,500-line cap:
the ARIA tests of `host/web-js/src/emit_tests.rs` move to `aria_tests.rs`, and
`kernel/build.rs` watches its `build/` directory in place of naming each file in it (which
also covers `cold.rs` and `rare.rs`, which it did not name). `SCHEMA_DIGEST` is recomputed.

The branch's own checks are in expo#404. On this commit the default members build
and the workspace compiles; the tests run on the commits that follow, which fix what review
found.

Lands expo#404.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`Kernel::role_step_order` decided by its own eligibility alone whether Home and End could
go past the focused member: from an eligible member it offered only the members before it
(Home) or after it (End). A host that cannot focus one of those, or the focused member
itself, was left with nothing to choose. On the painted hosts a tab with no box can hold
the focus (an action focused it) but is no destination for a key, so Home from such a
first tab moved nothing though the second tab could take the focus; End from such a last
tab likewise. LLP 1166.000 §8 listed this as not built.

Home's order is now every eligible member from the first, End's from the last, with the
focused member in its place. The host takes the first it can focus; when that is the
focused member, it is that end and the key moves nothing, as before (the presenter and the
Apple host already leave the focus where it is for a destination equal to the focused
member). Kernel answers with an accepting predicate are unchanged, so the 193 shared cases
hold.

Tests: `home_and_end_reach_the_first_member_the_host_can_focus` (kernel, a predicate that
refuses the end tab) and `home_and_end_pass_a_tab_this_host_cannot_focus_by_key` (the
presenter, zero-sized first and last tabs focused by `set_focus`). Both fail on the
previous commit. `exact_role_keys`'s ABI test now expects Home's and End's longer orders.
`cargo test -p exact-kernel --tests`, `exact-apple`'s lib tests and `cargo test -p
exact-painted --lib --tests` (2,413) pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… Tab stop on the Mac

The kernel counts every `role="tab"` under a tablist as a member, a run of inline text
included, as the web counts an `a role="tab"`. On the Mac an arrow from a button tab could
focus such a link, but the link's own key handler took a group's keys only inside a menu,
so no arrow could leave it. Its responder was also a Tab stop of its own with Keyboard
Navigation on, while each other member is a stop only as its group's one stop.

`Presenter.inlineRoleKeys` asks the kernel for an inline run that can be a member by the
same rule a node uses (`mayJoinGroup`: a `tab` or `radio` role anywhere; a link or a
`menuitem` inside a menu), and `roleKey(inline:)` takes whatever group's step it answers.
`InlineKeyView.canBecomeKeyView` refuses a run that is a member of a painted group and not
its stop, and the key-view loop lists every such run so that the stop can move to it with
the focus, as it lists a node member. `isStop` is the one place a stop is chosen from the
kernel's order. A step whose first focusable member is the focused one moves nothing.

Test: `RoleKeysMacTests.testATablistsInlineLinkTabTakesTheGroupsKeysAndItsOneStop`, a
tablist of a button, an inline link and a button: the arrows reach the link from both sides
and leave it, Home and End go on from it, and the group is one Tab stop. On the previous
commit it fails five assertions (the link is a stop of its own; Right, Left, Home and End
from it do nothing). `bun host/apple/build.mjs --test --only RoleKeysMacTests`: 11 tests,
0 failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'s loop too

From a focused member of a painted group that can never be a stop (`tabindex=-1`), the Mac
goes on by the member's place in tree order, as Chrome does, not by the loop that sorts
positive `tabindex`es first. That correction walked only the page's tree, and a guard
skipped every other window, so in a popover AppKit shows (a window of its own, whose
group's members are in its own sorted loop) Tab followed the sorted loop: from such a tab
between a `tabindex=1` and a `tabindex=2` control, Tab skipped the `tabindex=2` one and
Shift-Tab reached it in place of the `tabindex=1` one.

The loops are now named once (`keyViewScopes`: the page's or the open dialog's, and each
popover AppKit shows), and the correction walks the loop whose entry view shares the
member's window, going past either end to that loop's first or last stop. The tree walk the
loop and the correction share (`inTabOrder`) is one function, so the two cannot visit
different nodes.

Test: `RoleKeysMacTests.testTabFromAMemberOutsideTheOrderGoesOnByItsPlaceInAPopoverAndADialog`,
Tab and Shift-Tab from such a tab in a popover and in a dialog's sheet. On the previous
commit the popover's two assertions fail (Tab reached `tail`, Shift-Tab `after`); the
dialog's pass there, its sheet's Tab having its own path, and still pass. 21 Mac test
classes that touch focus, key views, popovers, sheets, radios and segments pass (302 tests).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each Canvas recording needs its own matrix. Invalidate the recorder's
cached matrix at scroll-group boundaries so the first background and
the drawing after the group receive their transforms.

Part of expo#558.

Co-authored-by: Chris Laupama <42254688+chrislaupama@users.noreply.github.com>
Co-Authored-By: GPT-6 Astra <noreply@openai.com>
The README said a development build writes the data module's `console` lines to logcat
under `exact-app`, and not that each goes at its method's priority, which an
`adb logcat '*:W'` filter depends on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Expose frozen through exactPage and prefer, produce web and Android lifecycle
transitions, and release gated streams in the transition commit. Desktop hosts
keep false. Document the feed policy and leave Apple's execution-eligibility
producer for the Mac stage.

Part of issues/20261113-no-fact-for-a-suspending-host-or-a-phone.md: not run on a device.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Apply stream cancellation and sequence stopping after successful settlement,
before a shared-element transition can defer the tree. Retain each commit's
sequence inputs and freeze edge, so resume cannot erase the stop and older
tree callbacks cannot restart the sequence.

Cover pending and skipped transitions, resume before the callback, repeated
freeze edges and refusal rollback. Check freeze/resume page listeners and
install the freeze tempo test's own browser stubs.

Part of issues/20261113-no-fact-for-a-suspending-host-or-a-phone.md: not run on a device.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…unning

Read running audio owners and successful activation for iOS and tvOS execution eligibility. Forward frozen through the Apple bridge, preserve synchronous sequence stopping and gated stream release, and support its agent substitute. macOS keeps ordinary execution when hidden.

Part of issues/20261113-no-fact-for-a-suspending-host-or-a-phone.md: not run on a device.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…arget

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ferences

Determine authored surface presence before class expansion inserts absent rows. Bind defaultSurface to the active class arm while preserving explicit surfaces.

Refresh marked iOS popovers before the batch flush when context-menu references change. Include inline runs and references retired with an exiting subtree. Reapply material geometry and invalidate paint so the surface updates together.

Regression captures wait for display transactions and presentation completion.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s it

Derive presence usage from secret grants and carry it through the bake, generated Rust entry, and Swift composition. Move the keychain implementation into ExactPresence and specialize store endowment so other apps retain no vault implementation.

Select production Swift products from the baked linked set shared with Rust. Unknown grants retain every capability, while development continues to select costly Swift frameworks by use.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ccheever and others added 24 commits November 14, 2026 17:04
Each redraw of a 2D canvas made a full copy of its bitmap for the
painters, and Ink, which keys its textures by the snapshot's address,
made a new mipmapped texture for every snapshot and generated its mips.
A canvas redrawn on each pointer move created and freed a full-size
texture each time, which fits the transient peak expo#508 measures on
Windows with the default painter and not with the CPU one.

Now a texture whose pixmap no one holds any more is kept for one frame
and takes the next pixmap of its size in place, so a canvas redrawn at an
unchanged size keeps one texture; a new size makes a new texture and
retires the old one at the next frame. A canvas texture has a mip chain
only when drawn at less than half its size. The replayer shares its
bitmap with the snapshot and copies it only when it draws while a painter
still holds the last one, so redraws between two paints copy it once.

On this Mac the Linux host under EXACT_PAINTER=ink, stepping the canvas
gallery's page 3 three times, uploaded 24 redraws into existing textures
and created none for them; before, each made a texture. What D3D11
retains on Windows is not measured here.

Part of expo#508

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'s first frame

setLocale's choice was kept across launches and read before the first
frame; setScheme's lived for the session, so an app re-applied a kept
choice from a mount task one millisecond after the first frame, which
showed the system's scheme first.

Now the runner keeps setScheme's argument beside persisted state under
exact.state.#scheme once its commit stands: "light" or "dark", while
"system" forgets it. At boot, a plan that runs setScheme asks for the kept
scheme in the boot's commands, which every host applies with the first
frame; the painted presenter also applies it as it boots, since its hosts
run commands after a paint. The web JS target keeps the same key in
localStorage and applies it when the command is installed, before the
plan's first render. Like a chosen language, a chosen scheme keeps an
undeclared route's rendered page from waiting for interaction.

The theme recipe, the guide and `contract vocab color-scheme` drop the
mount-task advice, and the docs site drops its mount task.

Fixes expo#517

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…expo#625)

Both renderers now choose the same hero by the same rule:
- A symbol (`data-symbol-source`) is an icon, never the hero. Rust read
  its `symbol:` source, which props_of has already rewritten to its file.
- A subtree with `display: none` holds no hero. In the JS render an
  inline value wins over the class's, as with `visibility`.
- A popover or dialog holds no hero: Exact writes none open, and the
  browser shows no closed one.
- The render DOM reflects `src`, so an icon symbols.js draws from its
  file is written with its file, as element.rs writes it.

The JS lookup reads an element's one class (tree.js) and no longer
splits the attribute. A walk of 2,000 elements takes 32-37 us whatever
the shell's classes, against 16-87 us before; writing the root takes
about 820 us.

LLP 1048.007 D5 (amended 2026-11-14).

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The web agent keeps storage with --keep-storage (scripts/agent.mjs); Apple
builds set no HOST_CXXFLAGS; Apple answers `perf frames live`
(FrameSampler.liveFrames); and Excord's three entries went with
apps/excord in b1e8f1e.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…her app of it compiled

Co-authored-by: ccheever <56719+ccheever@users.noreply.github.com>
The dev compiler is `cargo run` of the app's dev binary, whose build script
bakes the Contract, so a Contract error makes it exit. The loop treated that
exit as its own and stopped. Now it reports the refused generation on the
page and the console, keeps serving the last good one, and starts the
compiler again at the next change to the Contract sources.

Fixes expo#512

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dev page boots through `Bridge::boot_plan`, which booted with an empty
store and ignored the one the page handed in, so every `persist` state read
its initial value after a reload. A first `boot_plan` now reads that store,
as `boot` does; one beside a running host still carries the host's.

Fixes expo#522

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A trap leaves the module as it stopped, a Rust borrow held, so every later
timer wake and input trapped again and flooded the console. The page now
calls the module through guarded exports: the first `RuntimeError` stops the
clock and the text flow, marks the root, and reports once; later calls throw
without entering the module and are not reported. A stage still imports the
module's own exports. In the dev loop the overlay says what happened, and the
next generation reloads the page.

A game's dev loop, which is always the wasm loop, now builds apart from
`target/web-dist` for an outside app, as `--wasm` already did, so `agent web`
and `test web` no longer replace the files it serves.

Fixes expo#500

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A browser starts audio output only inside a user gesture, and a game's audio
unlocked only on a press that reached its canvas, so a game whose first
controls are Contract buttons stayed silent until the player touched the
canvas. The page now tells every live surface of a trusted pointerdown,
keydown or touchend anywhere in it (`Lifecycle::Activated`, code 4), and the
game's audio output resumes inside that gesture, as the Contract sound player
resumes its own. It changes no game state.

The surface's lifecycle tests move to their own file, which keeps surface.rs
under the 1,500-line cap.

Fixes expo#520

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`bindValues` read the canvas's bounding box on every bind, in the middle of
a batch, which forced the layout of everything the batch had written so far.
The size now reaches the module at creation, before the first bind, and
after that from the ResizeObserver, after layout. In the driver's Chromium,
pausing Beacons with a 300-row sheet mounted had two layouts forced from
`bindValues` inside the click (1.6 to 4.2 ms); now it has none. The batch's
per-element cost, the larger part of the click, is in QUEUE.md.

The surface fixture now declares `heldOutputs`, which the glue it slices
reads; without it 74 of the 86 cases in surface-record.test.mjs failed.

Part of expo#533

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both web targets held every pointer move for the next animation frame, so a
drag's actions landed a frame or two behind the pointer. A move now goes out
in its own event unless one already went out since the last frame; then the
latest goes at the next frame, so moves stay at most one a frame (LLP 1056).
Chromium and Firefox already deliver moves aligned to frames, so there each
goes out as it comes.

Part of expo#543

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dev page fetched and kept every file of a generation, each texture
family included, where the built site fetches what the renderer asks for.
Files under assets/ and deck/ are now read when the page asks for them, and
verified against their cards; a file read once is kept for the next
generation that has it unchanged. A surface's request reads its file, an
asset swap reads only the files surfaces have asked for, media loads from
the file's generation URL, and a font is read before the 100 ms font
barrier. The plan, its tables, shaders and modules are still read with the
generation. An asset a table edit leaves unchanged keeps the URL of the
generation on disk that wrote it, since the server holds only the last three
table generations.

In the driver's Chromium, Beacons with two unused 16 MiB files went from
32.5 MiB of backing storage and both files fetched to 0.54 MiB and neither.

From shreyam1008's prototype on the issue: the lazy card with its memoized
read, the font read before the barrier, and the table generations' URLs.

Fixes expo#499

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
a89e725 added a field whose type tripped clippy's type_complexity
lint in exact-painted, which is outside default-members.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep the blocking loop to format, caps, boot, default-member checking and
small unit suites. Move full checks and platform qualification to the hourly
lane, and bisect failures through affected commits with a passing parent.

LLP 1172.010 E13, with Charlie's 2026-11-14 30-second gate instruction.

Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
Restore tracked inputs edited by lane commands before probes and final checkout. Attribute each completed test independently, follow the failing suffix after repairs, and clear test failures covered by default-member passes.

Run each gate unit crate separately and remove the hosted hourly macOS job. Use the Android and Windows builders; mark Windows steps unavailable on other hosts. Remote Windows dispatch awaits a configured worker.

Validation: gate passed in 12.88 s and in 24.73 s after a runner edit; 42 focused controls and 11 adjacent selection tests passed. Verification-only changes add no app cost.

Co-Authored-By: Codex (GPT-6) <noreply@openai.com>
Keep workspace test failures until their own check covers them. Repair lane-owned tracked inputs before checkout, refresh the game lock for probes, and schedule game Bun suites outside the dedicated steps.

Report Windows steps as not run when a Windows worker is needed, and leave the daily full-run timestamp unchanged for unavailable checks. Worker dispatch remains outside this fix pass.

Validation: five regressions failed before the fixes; 46 focused controls and 12 adjacent selection tests pass. The gate passes in 18.19 s. Verification tooling adds no app cost.

Co-Authored-By: Codex (GPT-6) <noreply@openai.com>
…only when they leave exact2's

LLP 1153.003 D1. A crate beside app.json whose Cargo.toml names
`package.workspace = "../.exact/shells"` (a GPU crate: "../.exact/gpu") is
a member of the workspace exact2 generates, so an app with a data crate
keeps its host crates generated instead of copying the whole workspace up.
A crate that names neither is refused with the line to add, and one that
takes a generated host crate's name is pointed at the authored workspace,
which stays the escape (D3). resolveApp admits exactly the two generated
workspaces and refuses any other beside them; a GPU crate alone leaves the
host crates generated.

The lock follows the game's rule (LLP 1046.011 D2), now in code both
generators call (scripts/app-shells.mjs, moved from game/app/shells.mjs):
the generated workspace resolves against exact2's lock until the app's
crates need a version it does not hold, which a build refuses with the
command to run; `bun exact.mjs lock` then captures Cargo.lock beside
app.json, which every rewrite of the shell copies in and never replaces.
For an app with crates of its own, rust-toolchain.toml and
.cargo/config.toml (target-dir "target") are written beside app.json and
gitignored, so `cargo test` in a member uses exact2's compiler and the
app's target/. The stamp names the member list. NO_WORKSPACE's advice to
copy the shell up is gone. rules/DEFERRED.md records the expansion
(waived by Charlie, 2026-11-14).

Part of expo#423
Part of expo#578

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ates link it through one shared bake

LLP 1153.003 D2. `rust.data { crate, type, source? }` names the data crate
and its DataSource; the source defaults to `<type>::default()` and is the
type itself, `::default()` or `::new(…)` with literal arguments. The
generated workspace gives such an app host crates that link the data crate
and whose build script is one line, `exact_bake::host::app(platform, app,
source)` (bake/src/host.rs, the game's bake lifted and generalized; the
game now calls it too). Without app.ts they link no exact-js, so no Hermes;
with app.ts beside the crate the two compose (`exact_data::Mixed`), Rust
answering what the crate's `RUST_SOURCES` names and TypeScript the rest.
The web's Rust module is a member of the same workspace
(`.exact/shells/module`), built `--locked --offline` over the source the
manifest names, so `bakedRustSource` and the source-copying it did are
gone. The JS dev loop also watches the module's path dependencies outside
the app, read from its dep-info (D1.6).

Decided here, by the RFC's principles:
- A data crate composed with app.ts exports `RUST_SOURCES`, the sources it
  answers; the mixed bake reads it rather than guessing from the plan.
- The generated hosts wrap the source in `exact_data_host::Storage`, as the
  authored Rust apps already did, so storage-backed answers work unchanged.
- A data crate that names no app (`DataSource::app_id`'s default, as
  `exact new --rust` writes it) takes app.json's identity: alone, through
  `Swappable::named` as before; beside app.ts, through the new
  `exact_data::Placed::named`, which gives it the module's, since `Mixed`
  requires both sides to share one. The bake's identity check likewise
  lets an unnamed Rust source pass, as the runner's gate does (LLP 1023
  D5). A crate that names its own id keeps it.
- `rustData` joins the compatibility inputs only when declared, so every
  in-repository Rust-data app's compatibility id changes once, with this
  commit, and apps without it keep theirs.
- `contract::rust_entry_naming` refuses a web entry whose source differs
  from the manifest's, which the JS target's module runs. A native host
  crate the app wrote may pass another type of the same data crate:
  markdown-stress's Apple host runs `NativeMarkdownStress`, its native
  document lane, by design, which the RFC's refusal of every differing
  authored entry would forbid. messages-stress names
  `ReusableMessagesStress`, the source all three of its entries already
  passed; a web build of its stateless control now sets rust.data too
  (its README says so).

Every in-repository Rust-data app's app.json names its rust.data.

Part of expo#423
Part of expo#578

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…no host crate

LLP 1153.003 D4. The Rust starter is `data/` (package.workspace =
"../.exact/shells", exact2's crates taken as `.workspace = true`),
app.json's `rust.data` naming its type, and no app.ts. The authored
workspace it wrote is gone: the root Cargo.toml, apple/ and web/, the
copied .cargo/ and Cargo.lock, and the writers `hostBuild`, `hostLib` and
`manifest` in scripts/app-rust.mjs. Since every host crate is generated,
`--rust --with linux` (or android, windows) is no longer refused, and the
listing keeps the linux and windows verbs. `exact new` still runs
`bun exact.mjs lock`, which now writes the generated workspace so that
cargo and an editor work in data/ before the first build (D1.7) and
captures no lock while data/ needs nothing exact2's lock lacks.

Part of expo#423

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ir own

LLP 1153.003 D5.1. A GPU module links nothing of the host's, so the crates
beside app.json that name `package.workspace = "../.exact/gpu"` are members
of `.exact/gpu/`, generated apart from the host's `.exact/shells/`. It
carries exact2's `[patch.crates-io]` and, as workspace dependencies,
`exact-gpu` and the web dependencies `module!` exports through at exact2's
pins (the wasm-bindgen CLI the web build runs refuses any other). Its lock
follows the host's rule: exact2's until the GPU crates need a version it
lacks, then `bun exact.mjs lock` captures `Cargo.gpu.lock` beside app.json.
A crate named as the app's GPU module that joins the host's workspace is
refused with the line to change.

The bake, the web's GPU build and the web build's source roots read the
host's graph and the GPU crates' as one (`joinMetadata`) and run Cargo for
a GPU crate in its own workspace (`gpuView`); an in-repository or authored
app builds both in its one workspace as before. docs/recipes/gpu-module.md
no longer copies the workspace up: it writes gpu/ with the member line.

Part of expo#578

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ow an app moves to it, and how to leave it

LLP 1153.003 D3 and D4. docs/reference.md, "Rust data sources", says what
`rust.data` names, how a data crate joins `.exact/shells/`, how it composes
with app.ts (`RUST_SOURCES`), where cargo and rust-analyzer work, when a
lock is captured, and the steps that move an app `exact new --rust` made
with host crates of its own (or one that copied the workspace up for a
data or GPU crate) onto the generated shape. Its new section "An authored
workspace" is the escape for an app that writes its host crates: copy the
generated workspaces up and rebase their paths. Both were run: the move on
an app the earlier `--rust` wrote (`test web` passes), the copy-up on an
app with a data and a GPU crate (web and macOS draw the GPU canvas).
docs/start-here.md's routing table names a GPU crate's workspace and the
Rust data section; docs/contract-for-humans.md no longer says a new app
has host crates.

Fixes expo#423
Part of expo#578

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…esystem.mjs imports

filesystem.mjs began importing tools.mjs (2303f74), which imports
kept.mjs; the fixture copies scripts/ file by file and had neither, so
29 of app-workspace.test.mjs's tests failed to load app.mjs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/patterns/chat landed (09bf6ac) without its entry in PATTERNS, so
vocab::every_recipe_and_pattern_answers_by_name failed on main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream's sheet color-scheme change (9213f3c) covers the fork's #4, so
ColorScheme.swift and ModalIOS.swift are upstream's; #4's test stays.

The fork's fadeDuration and loading props move to ids 310 and 311, after
upstream's, and the schema digest is recomputed. The image fade glue moves
from rt.js to tree.js, where upstream's PropHooks now lives. The iCloud
key-value entitlement test moves to host/apple/presence.test.mjs, so
scripts/app-apple.test.mjs stays under its line cap.
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.

10 participants