Painter::draw_twice(child, first, |used| second) becomes Painter::measure
plus an ordinary draw. Iris's objection was the shape it forced on the
caller rather than the two draws themselves: the arithmetic that picks
the real region had to happen inside a closure, and anything it wanted to
keep came back out through a captured &mut. LazySpan::place was the only
caller, and it now reads as the three statements it is.
DrawMode::Measure is that draw with everything it writes switched off --
no arena slot, no mask, no move slot, nothing left in `active`, nothing
marked dirty. Only the returned Size survives, and the widget is left
exactly as it was, so the real draw that follows is an ordinary first
draw or redraw. That last part is load-bearing: a measurement that left
an ActiveData behind would let the following draw hit draw_inner's
"already at this region" fast path and return having drawn nothing.
A measurement also does not consume a redraw mark, since it is not the
redraw the mark asked for, and it takes none of the fast paths, since
"already drawn here" cannot report a size.
Every Painter method that writes now returns early on the mode -- a
widget's own draw never checks, which is the point. A debug_assert at the
end of draw_inner catches one that forgot, because the failure otherwise
is a single leaked primitive per measured widget per frame, which a
screen redrawn every frame turns into an arena that grows without bound.
What this is worth, and what it is not. The amplification it applies to,
measured on a streamed frame: 1,083 Widget::draw calls over 113 distinct
widgets, with the worst drawn 11 times at nesting depth 7-8 -- it is not
two draws but two to the power of how many measuring ancestors a widget
has. Only the writes go away; the walk and the region arithmetic still
happen 11 times, and removing those needs a size answerable without a
draw, which LAYOUT.md section 5 rules out. Streamed frame p50 1.39ms ->
1.22ms, p99 4.75ms -> 3.58ms. The upload numbers do not move, because
slot recycling had already made the discarded writes free in arena terms.
Also extracts move_slot_for from draw_inner, since measuring must not
allocate one and the reuse-in-place rule wanted saying once.
Verified: run-tests.sh, iris's suite, clippy and rustfmt clean, and the
headless phone render is byte-identical to the previous commit's on the
real GPU (Venus, RX 7900 XT -- checked, not llvmpipe).
Changing any primitive re-uploaded every primitive. Measured over the
bench fixture by the new arena_churn rig: 758 MB across a fling and
1.2 GB across 401 streamed deltas, p50 3.0 MB per streamed frame.
Three separate things were wrong, and only the first is what it looked
like from the outside.
ArrBuf reallocated on every length change. A fresh Buffer's contents are
undefined, so adding one glyph -- which a streamed reply does constantly
-- forced a full rewrite, and no partial upload could have been correct
in the first place. It has a capacity now, growing geometrically and
never shrinking, and update() answers whether the Buffer identity moved
so a caller can rebuild its bind group and force the whole range dirty.
That alone took the glyph array from 95% re-uploaded to 3%, and stopped
primitive_group being rebuilt on every frame the arena changed.
A redraw freed its primitives and pushed new ones. Freed slots are not
reusable until the end of the frame -- a layer's draw order still names
them -- and Painter::draw_twice is how a container learns a child's
size, so with containers nested the arena's high-water was the transient
push count rather than the live one: 17 million pushes across 401
deltas, 127,443 slots for 11,569 live primitives, growing linearly with
the transcript. A redraw now gets its old handles back as a recycle pool
(Painter::take_recycled, Primitives::recycle) and writes into the slots
it already holds; the pool is consumed in order and whatever the draw
does not claim is freed when it ends. The arena is exactly the live
count now. The CPU frame improved with it, from p50 2.20ms to 1.39ms on
the stream run, because the freeing and the draw-order renumbering went
away.
Nothing tracked which entries changed. util::Dirty is a bitset per
uploaded array, coalesced into ranges at a 1 KiB gap. Marking is O(1)
and allocation-free; reading it back is one word per 64 entries. Both
alternatives were measured and rejected: a min..max span is nearly the
whole buffer, since a frame's changes land in 5-20 scattered runs, and a
Vec of indices would mean an allocation and a sort per frame at several
thousand marks. It replaces Primitives::updated -- one bool that covered
the instances and the per-primitive data together, so rewriting a rect's
region re-uploaded every glyph -- and TrackedArena::changed.
The trap only the rig could catch: writing an entry is not changing it.
Recycling rewrote every glyph of every moved row with identical bytes,
marking 73% of the glyph array against 0.6% genuinely changed, because
what moves is the instance's region and not the glyph. PrimitiveVec::set
and Primitives::set_instance compare before marking.
Every array now uploads within a hair of its floor: fling instances 3.4%
against 3.3%, fling glyphs 0.9% against 0.8%, stream glyphs 0.6% against
0.6%. Stream instances are at 72.7%, which *is* the floor and is a
layout question rather than an upload one -- the list is pinned to the
newest end, so a growing reply moves every row, and that should be one
move_offsets write rather than a redraw. Noted in RUST.md as the next
thing.
Also: draw_inner's four old_* parameters become one Retained struct, so
the recycle pool is a field rather than an eleventh positional argument
next to three others of the same shape; and free_primitive is the one
place a slot and its draw-order position are retired together.
The rigs move to scripts/rigs/ui-profile, a crate of their own so a
rig's dependencies stay out of the app's -- arena_churn needs bytemuck,
which nothing in ai-app does. arena_churn prints floor, uploaded and
whole side by side per array, because any two of those alone are
misleading and the 122x over-marking above was invisible until all three
were on screen together.
Iris, on the two findings from the incremental-text investigation:
"let's switch to new lines for the test, and also let's keep the single
line around for stress + could be something to try to optimize later."
The streamed tail now takes a blank line every 4-12 deltas, so it is 53
markdown blocks with a longest of 502 characters instead of one block of
14,888 -- against a measured p50 of 147 and a largest-ever 1,580 over
7,706 blocks of real assistant messages. Layer 1's streaming frame went
from p50 3.86ms / p90 8.65ms / worst 10.95ms to p50 2.20 / p90 5.90 /
worst 8.78.
The run-on message is kept as the first two backlog events, 14,824
characters in one block, just under text_cap's 16 KiB so it draws in
full. The *streaming* pathology stays in frame_profile.rs rather than the
fixture: it needs a growing block, and iterating on it there costs a
second instead of a two-minute phone run.
Adding it is purely additive -- the random state is saved and restored
around those two events, so every other backlog event is byte-identical.
That is not tidiness: the first attempt shifted the backlog and broke
`a_long_press_and_drag_selects_text`, which replays a real recording at
(300, 1000) and needs the content it was recorded against to still be
there. BACKLOG_COUNT is 3202 now, in generate.py, fixture.rs and
BenchFixture.kt, which split the file by line index.
And the answer to Iris's question, which the code already had: the newest
message does *not* cap. `build_row`'s `cap` is false for the live tail
because a row that grew while capped would appear to stop growing, and a
reply growing past the cap is never caught either since it grows through
apply_delta. So a streamed block's shaping cost has no ceiling -- ~29ms
per delta at 50k characters, ~58ms at 100k.
Recorded but not chased: the emulator's `stream: build p50` did not move
(10.4 -> 10.5ms) while layer 1's frame nearly halved, so most of a
streaming frame on a GPU path is the whole-arena primitive re-upload
layer 1 never performs -- 11,568 primitives rewritten per delta, with the
fling phase as the control at 0.4ms for the same primitives moved
through move_offsets.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris asked to look into incremental text rendering, hoping parley
supported it. It does not, by design: a `Layout` re-linebreaks and
re-aligns freely but "if the text content or the styles applied to that
content change then a new `Layout` must be created", its LRU cache holds
harfrust's per-font shaper data rather than shaped runs, and its own
`PlainEditor` rebuilds the whole layout from the whole buffer on every
keystroke.
The app already does what incremental layout would buy: `RowBlocks::
apply_delta` keeps one `TextEdit` per markdown block and re-shapes only
the one a delta landed in. Re-splitting the markdown to find it is 18µs
at 18,000 characters; comparing the blocks is 470ns.
What is left is one `TextBuffer::shape` of that block, linear in its
length at ~0.23ms per 1,000 characters here -- and the bench fixture's
streamed message is 14,888 characters in a *single* block, a run-on
paragraph with no blank line in it, so every delta reshapes all of it.
That is 3.5ms of the measured 3.86ms frame.
Real replies are not that: across 7,706 top-level blocks from 3,675 real
assistant messages on this machine (lengths only, no content copied
anywhere), p50 147 characters, p90 449, p99 836, largest 1,580, nothing
above 4,000; code fences p50 126, largest 589. At those sizes a reshape
is 48µs to 372µs here, roughly 0.12-0.93ms on the phone -- inside a
120Hz budget with no incremental anything.
So the recommendation is not to build it, and to give the fixture's
streamed message the paragraph structure a real reply has instead. Three
runs added to `frame_profile.rs` so none of this is re-derived: what
reshaping a growing message costs (including at the sizes real replies
reach), where a delta's cost is, and what the fixture actually streams.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris's phone came back "now THAT is smooth", and reading that run against
the bench's own timings found three things the report was getting wrong
-- two of them shipped yesterday in the fix for the last three.
`missed vsyncs` counted idleness. Every gap between frames was treated as
cadence, so the bench's own pauses read as stutter: 276 for sixteen 300ms
rests between flings, 2410 for twelve hundred 50ms keystroke gaps, 821
for four hundred 50ms stream gaps -- each within a few percent of the
arithmetic. A gap now measures anything only if the frame before it had
asked for another one.
`late` counted the swapchain wait as cost. A well-paced loop spends each
frame blocked in the acquire, so its total sits at exactly one refresh
period and every frame lands on the budget boundary -- 0.4ms of work and
5.7ms of waiting is not a late frame. It is judged on `FrameParts::work`.
And the refresh rate is the larger of what the platform claims and what
the run sustained, because each can only be wrong one way.
`Display.getRefreshRate()` answered 60 for a run that drew 3405 frames in
33.1s, since a phone that varies its rate answers with whatever mode it
is in when asked. The first attempt at measuring it instead took the
fastest tenth of the gaps and reported 88Hz for this repo's 60Hz
emulator, whose app manages 54 -- a budget no frame there could meet,
invented out of the app's best moments, and caught only by running the
corrected report on the emulator before shipping it. A sustained rate is
a floor and cannot do that. Both are printed when they disagree.
Also corrected in the docs: "103fps on a 120Hz screen" divided the fling
phase by its whole duration, rests included. Both runs sustained ~120.3fps
through the motion, so the callback ordering was never costing frames --
what changed is the clock, which moves no frame count at all, which is
exactly why nothing in a report could show it.
`fling_profile.rs` is `frame_profile.rs` and gained a stream run, which
says where the frame time now is: folding an arriving event is 0.35ms and
applying the diff 0.41ms, while the frame is 3.86ms here and 9.5ms on the
phone. 401 events move the item count 652 -> 654, so nearly every one is
a delta into the same row -- the cost is re-shaping one growing message,
not `fold_event`'s per-event clone, which was the hypothesis and is what
measuring it ruled out.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris, from her phone: "some stuttering when flinging in particular.
Harder to notice with my finger directly moving the scroll." Her fling
phase was 103fps on a 120Hz screen at p50 6.3ms.
Two of the four things found are corrections to the instrument, not the
renderer. The swapchain acquire -- `get_current_texture`, which *blocks*
until the compositor frees an image -- was inside the span the report
called iris's CPU work, so a fling comfortably ahead of the display read
as milliseconds of being slow. A frame is now three measured parts
(`FrameParts`: build, acquire, submit), per phase as well as per run. And
nothing could say a frame was never *produced*: `late` counts frames that
cost too much, which a reader does not see, while a frame that never
happens leaves the last one up for two refreshes, which is the stutter.
`PhaseStats::missed` counts vsyncs nothing was drawn for. It closes on
the emulator: 1548 frames + 452 missed over 33.0s at 60Hz is 1980
vsyncs.
The other two are the frame loop. `Choreographer.postFrameCallback`
schedules for the next vsync after the call, and iris asked at the *end*
of the callback -- so any frame whose work ran past the boundary
registered too late and got the vsync after, one frame over budget
silently costing a second. It is asked for immediately after
`tick_animations` now, on both backends. And the fling was advanced on
`Instant::now()` rather than the vsync `do_frame` carries: frames are
presented on an even cadence whatever clock computes them, so sampling
the spline at "whenever the callback ran" moves the content unevenly with
no frame late enough to appear in any report -- and a drag never had it,
which is the asymmetry Iris described. `PointerClock` is `DeviceClock`
and the view keeps one, anchored by whichever of a touch or a frame comes
first, so a fling is advanced on the clock its velocity was measured on.
`opt-level` for the Android release build goes from "s" to 3. The table
in RUST.md picked "s" on bytes alone; over the same warm fling eight
times iris's own per-frame work is p90 0.15ms/p99 0.42ms at "s" against
p90 0.09ms/p99 0.26ms at 3, for 1.8 MB of arm64 APK.
`app-rust/tests/fling_profile.rs` is the rig that established what a
fling frame actually costs and is kept for next time (Iris: "please keep
the profiling rig around for future use"): only one frame in six lays
anything out, and the multi-millisecond spikes are all first-pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris: "remove both decisions and iris.md. I've decided to instead make
decisions when planning with agents rather than after they do things, and
they're both too long for me to wanna read, + don't cover all the
decisions I'll wanna make about the code anyways. I'll just naturally run
into things for now. Todo is important though."
So docs/DECISIONS.md (850 lines) and docs/IRIS.md (1,986) are gone, and
AGENTS.md now says not to start another: raise a choice while planning it
with her, otherwise decide it and put the reasoning at the code it
governs. The TODO lists stay. docs/SUBAGENTS_DECISIONS.md went with them
-- same artefact, same reasoning, and she did not name it, so its six
decisions were folded into docs/SUBAGENTS.md rather than deleted.
Deleting the logs left ~30 citations dangling in code comments and docs.
Each states its reason inline and cited the file only for provenance, so
they now read "decided 2026-09-07" or name the module doc that carries
the reasoning.
The root had six things that were not a program or a document. Moved,
per "I only meant top level sh files":
run-tests.sh, test-wg-tunnel.sh, wg-setup-host.sh -> scripts/
rigs/ -> scripts/rigs/
xtask/ -> scripts/xtask/
A project's own scripts stayed with the project: app/*.sh, app-rust/*.sh,
iris/*.sh and server/enroll-link.sh did not move.
`target/` at the root is deleted and cannot come back: there was never a
workspace there, and the 29 MB was only xtask's scratch space, now in
scripts/xtask/target/. `cargo xtask apk` still runs from the repo root
and now publishes to scripts/build/outputs/apk/<mode>/ -- one directory
deep, because that is what Dev Updater's `*/build/outputs/apk/*/*.apk`
discovery pattern needs, and scripts/xtask/build would have been two.
Verified: ./scripts/run-tests.sh and `cd iris && cargo test` green, clippy
and fmt clean everywhere, `cargo xtask apk debug --abi x86_64` builds and
signs an APK carrying lib/x86_64/libai_app.so at the new publish path, and
the repo root is now eleven entries with no build output among them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris: "the documentation is also pretty crazy too. Can you go through it
and remove everything that's already done and decided? There's entire md
files iirc for projects already complete. And many with checkboxes already
ticked off that just fill up context."
docs/RUST.md 8503 -> 905 the framework bake-off (options,
recommendation, twelve closed
experiment boxes) and two superseded
"where things stand" sections, out;
what the experiments settled kept as
one line each
docs/IRIS_TODO.md 1383 -> 229 fifty closed items and six
phone-report sections whose defects
are all fixed
docs/LAYOUT.md 1116 -> 829 the pre-implementation framing: the
old trait, the checklist, the
migration list, the pass conditions
docs/TEXTURES.md 496 -> 240 the prior-art survey, the proposal
and its review, all implemented
docs/REVIEW-*.md 673 -> 0 two completed review passes; the two
findings left open on purpose (mask
hit-testing, the phone's font set)
moved into RUST.md
What survives a prune is what cannot be cheaply re-derived: measurements
(the APK-size table, the phone bench reports), dead ends, invariants and
their reasons, and the design of what exists now rather than the route to
it. AGENTS.md now says that, so the next session prunes as it goes rather
than appending; docs/IRIS_TODO.md's header says items are deleted when
they land rather than ticked.
Deleting the two review files left eighteen citations dangling in code
comments that state their reason inline and cited the file for provenance
only — those now read "(review, 2026-09-06)" and carry no dead pointer.
The emulator's measured GPU capabilities moved to the this-machine-android
skill, where machine facts belong. IRIS.md and DECISIONS.md are dated
records and were not rewritten; each gained one note that paths in older
entries predate the 2026-09-08 crate merge, pointing at the mapping.
Not touched, deliberately: docs/DECISIONS.md's entries (that file *is* the
queue of things for Iris to review, so deleting decided items would remove
what it exists for) and iris/readme.md and iris/TODO, which are hers.
Verified: ./run-tests.sh and `cd iris && cargo test` green, clippy and fmt
clean in every workspace, and every remaining docs/*.md cross-reference
resolves.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris: "the organization of the rust rewrite is a mess right now... there
shouldn't be anything related to the app inside of iris. Iris is supposed
to be the UI framework alone." And, on the crate count: "I'm confused why
the app only code needs more than one crate though."
Nine cargo workspaces become three, and the port's project code -- which
sat in five places, four of them inside the framework -- becomes one crate,
`ai-app`, in `app-rust/`:
client-core -> app-rust/src/client
iris/transcript-ui -> app-rust/src/ui
iris/transcript-fixture -> app-rust/src/ui/fixture.rs + tests/ + touch/
iris/desktop-app -> app-rust/src/desktop + src/bin_desktop.rs
iris/android-app -> app-rust/src/android + android-project/
android-shell -> app-rust/src/shell
iris/ keeps core, macro, the iris crate, tabs-ui and rig-input, and now
mentions no session, transcript, setup or server anywhere.
Only two of the old splits had a reason that survived reading. event-model
stays a crate at the repo root because server/ depends on it too, so a
crate is what makes the backend and the app agree by construction. The two
Android .so names looked like a hard constraint -- a package produces one
library artifact -- until P2 turned out to already plan merging those two
Android apps into one; both faces now come out of libai_app.so, picked
apart by features so `--no-default-features --features shell` keeps wgpu,
parley and iris out of the Compose app's APK. docs/RUST.md's "One app
crate" has the rest, including what each remaining feature is for.
DECISIONS.md and SUBAGENTS.md move into docs/ with everything else.
Verified: ./run-tests.sh and `cd iris && cargo test` green, clippy and fmt
clean in all five workspaces, `cargo ndk -t x86_64` links libai_app.so,
build-apk.sh produces an APK that installs and launches on this checkout's
emulator (Gl ... virgl, as expected), and the phone-sized headless
screenshot renders the transcript unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
First of three steps agreed with Iris for getting scrolling out of the
list and into `Scroll`, so that `.scrollable()` is the one way anything
in iris scrolls. docs/IRIS_TODO.md's "In progress" block carries the
whole plan and the decisions behind it; this step is the rename and the
direction.
`List` -> `LazySpan`, and it moves in beside `Span` under
`widget/position/`. It is what `Span` is -- a sequence of children along
an axis -- laid out lazily from an anchor instead of eagerly from the
start, and the name says the one thing that matters about it. It also
stops colliding with `BlockKind::List` in the markdown code.
`ListRow` -> `LazyItem`; `RowKey` keeps its name, since rows are the
vocabulary in transcript-ui.
`Axis` -> `Dir`, with the sign meaning what it means in `Span`: which
end of the box item 0 sits at. **That is a different question from which
end the view is pinned to**, and conflating them would stand a
transcript on its head -- its oldest message is item 0 and sits at the
top (`Dir::DOWN`) while the view clings to the bottom. So the pin is its
own constructor argument, `LazySpan::new(dir, at_end)`, spelled the same
way as `Scroll::new`'s.
Making `Dir::UP` real rather than nominal is most of the diff. The walk
now works entirely in direction-relative pixels from the leading edge --
`Edge::Top`/`Bottom` are `Leading`/`Trailing`, `Placement` likewise, and
`RowExtent`'s fields and every local are `lead`/`trail` -- with two
places converting: `abs_region`, which flips the box for `Sign::Neg`,
and `flip_pos`, which converts the screen-space positions the public
helpers speak in (`note_tap`, `key_at`, `extent`, all fed by pointer
events) into the walk's space. Without the second, a reversed span would
hit-test at the mirror of where it drew.
`a_dir_up_span_grows_upward_from_item_zero` asserts on where each row was
**actually drawn** (`UiRenderState::active`), not on `extents`: the first
version of it read `extent()` and passed with `abs_region`'s flip deleted
-- checking the bookkeeping against itself while every row painted at the
mirror of where it belonged. It now fails with the flip removed (row 2 at
80..100 instead of 0..20), which is the check that matters.
Verified: cargo fmt --check, clippy --workspace --all-targets clean,
cargo test --workspace green (21 suites), including the phone-shaped
fixture tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The last place in iris that corrected itself on a later frame, and the
item docs/IRIS_TODO.md carried from the Scroll change. Iris's rule:
"nothing in the framework should ever self heal because it should not be
drawn incorrectly in the first place. If you need 2 draws to get
something into the correct position then that should happen within the
same frame."
`clamp_to_content` measured the gap past the end of the content from the
edges the walk had just placed, wrote it to the anchor and asked for
another frame -- so one frame was drawn with the list past its own end,
and a fling that had already stopped was not going to ask for the frame
that fixed it. Now the walk outward from the anchor is `List::lay_out`,
`overscroll_gap` is a pure measurement of the same gap (no painter, no
redraw handle), and `draw` moves the anchor and walks a second time
inside the same frame.
One further pass always settles it: the gap comes from the edges the
first walk placed, so moving the anchor by it puts that edge exactly on
the viewport's, and the opposite end can only open a new gap when the
content is shorter than the viewport, which `overscroll_gap` declines to
touch. The second walk is paid only on an overscrolled frame and re-offers
every row the same cached-height box at a new offset, which `draw_inner`
dispatches as an O(1) move.
`Painter::draw_again` had no other caller and is removed with it, so the
framework no longer offers a way to ask for a corrective frame at all.
Simplification in the same change: a placement is one pinned edge plus a
height, so `Placement::edges(height)` gives the box and `place`'s
top-known and bottom-known cases stop being two copies of the same
arithmetic -- four match arms down to two.
Four tests draw no settling frame on purpose and fail without the change:
`fling_toward_the_start_stops_at_the_first_row` and the new
`scrolling_past_the_start_is_given_back_in_the_same_frame` (list.rs), and
`scrolling_past_the_first_row_settles_on_it` /
`scrolling_past_the_last_row_settles_on_it` (layer 1, top_edge.rs).
Verified: cargo fmt --check, clippy --workspace --all-targets clean,
cargo test --workspace and ./run-tests.sh green, the phone-shaped
headless window replaying flick-120hz.touch draws the transcript
correctly, and the arm64 release APK builds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`cargo upgrade --incompatible` in each of the nine workspaces here, then
`cargo update`. Most of it is version numbers only -- log, winit,
bytemuck, image, tokio, libc, android_logger, proc-macro2/quote, and syn
2 -> 3 with no source change. The wg-app-link submodule's twelve
dependencies were already at their latest majors, so that shared
repository needs no commit.
wgpu 28 -> 30 (and pollster 0.4 -> 1.0) is the part with API in it:
bind-group and vertex-buffer slots are optional now, `Instance::new`
takes an owned `InstanceDescriptor` carrying the platform's display
handle (the desktop passes winit's, since wgpu wants it for a GLES
surface presented on Wayland -- which is what this machine's fallback
produces; Android passes none), `RequestAdapterOptions` and
`SurfaceConfiguration` each gained a field kept at its historical value,
`get_current_texture` answers with an enum instead of a Result, and
`present` moved onto the queue.
The one that would not have failed at compile time: naga now requires
`@interpolate(flat)` on integer varyings, so `shader.wgsl`'s three u32
outputs were rejected at `create_shader_module` -- an abort on the device
rather than a build error. Flat is the only interpolation an integer can
have, so this states what the hardware already did.
Checked: build, clippy, fmt and tests in all nine workspaces (iris 196,
server 160); layer 2 screenshots on Vulkan and on force-gles, identical;
the arm64 release APK builds and the x86_64 bench ran a full
fling/stream/type/keyboard cycle on the emulator's GLES adapter.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
RUST.md gets the report verbatim with what each of the four defects
actually was, the tests that pin them, and two traps worth not
re-finding (a fixed-coordinate tap that "failed" by 544px because it had
toggled a tool group, and a layer-1 repro that only reproduces inside a
`List`). IRIS.md and DECISIONS.md get the design half: one `Flinger`
whose seam puts the sign convention and the content's end with the
caller, a cancel as a first-class end to a gesture, and why a row is
drawn twice on the frame its height changes.
LAYOUT.md gains the two rules those turned on, since both govern the
layout rather than this pass: padding works in any container and is an
inset or an outset depending on how tight the parent's region is (Iris's
own words), and a widget offered a box it does not fit is drawn again at
its true box in the same frame rather than the next one.
IRIS_TODO.md's "worked around in tool.rs rather than fixed here" is gone
-- Iris, 2026-09-08: "There should never be workaround code." Two of the
four entries are ticked; the two that remain are missing capabilities
rather than defects being dodged, and each now carries a diagnosis of
what building it costs instead of a workaround: an overflow ellipsis
needs `TextBuffer` to have a displayed string distinct from its source
(parley has none of its own, and every byte-offset consumer -- spans,
`byte_at`, `Selection`, `apply_delta` -- moves if the buffer is
truncated), and selectable tool-card text needs a register/unregister
lifecycle across the three routes that rebuild a card, which is where a
stale `Selection` handle panics.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris asked why `mark` existed at all -- "the font should be working if
it's working for compose and nerd fonts are bundled". It was not: the
Compose app draws its icons from its own committed Nerd Fonts subset,
while iris, which bundles no font since 2026-09-07, was setting the
disclosure mark with bare geometric codepoints (U+25B8/25BE/25B4) out of
whatever face the platform resolved -- an empty box on her phone, a dot
on this VM. The 2026-09-07 note that "iris had no equivalent icon font to
keep" is the gap: it had none because it had never had one.
So iris ships the same kind of subset. iris/core/build-icon-font.sh is
the Compose script with its own GLYPHS list, writing a 992-byte
nerd_icons.ttf with three Material Design glyphs from the Mono face;
iris::icon names the codepoints; Family::Icons is how text asks for them.
The variant names an intention rather than a font name -- only TextData
knows what the file registered as, and it resolves it during shaping --
and it is a named family, never a generic one, so nothing falls back into
it for text and an icon cannot fall back out of it onto a system face
that happens to have the codepoint.
every_icon_is_in_the_bundled_font maps each constant through the shipped
font's charmap, which is the guard the script's "the two lists have to
agree" comment asks for. FontDiagnostics gains icon_family, so a build
whose font failed to register says so instead of drawing tofu; the
emulator reports icons=Some("Symbols Nerd Font Mono").
widget/mark.rs is deleted. It drew one correct triangle, but every
further icon would have been another rasteriser, and an icon as text
takes the size, colour and baseline of the line it sits in for free.
Looked at rather than only compiled: closed and open marks in
run-headless.sh phone --phone either side of a tap, and the collapse
bar's up mark under IRIS_TOOLS_EXPANDED=1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The bench APK panicked on frame 1 on the emulator:
iris panic at iris/core/src/render/texture.rs:461:22:
texture slot 89 is not a live standalone image: None
widget::mark called Textures::add per widget, so a folded card per tool
call meant a standalone image, a bind group and a draw call each --
hundreds of copies of three pictures. Textures::reset, which the Android
surface-rebuild path calls for a genuinely new renderer, then threw the
slot numbering away with the pixels, leaving every one of those live
handles naming a slot nothing recognised. Its doc had said the only
standalone image in the workspace was tabs-ui's, "confirmed by grep" --
true when written, false the moment mark existed.
Textures::reupload replaces reset: queue every slot for upload again in
slot order, empty slots included, so the new device gets the same slot
numbering and a handle a widget has been holding still names its own
texture. The glyph atlas is no longer cleared on that path either, so an
app switch stops re-rasterising every glyph on screen.
Textures::shared(key, make) is one texture per description, keyed by a
SharedTextureKey the caller packs exactly rather than hashes. mark keys on
direction and colour: three mark textures for the screen, not one a card.
And the devlog can finally show a panic. After a crash, Dev Updater's
query starts the app process for the provider alone, so no activity ran,
so set_crash_dir never replayed the panic hook's file -- the Runtime tab
held one line, the provider announcing itself. DevLogProvider.nativeReady
takes the files directory and does the replay from onCreate; the hook also
saves the dying run's last 80 lines beside the panic, read through a new
non-blocking LogRing::try_tail_text so a panic holding the ring's lock
cannot deadlock the hook.
Verified on this checkout's emulator: opens clean, survives 33 full-screen
scrolls back through the fixture, image_bind_group_creates_prev=1; a real
panic replays into the next launch, and a hand-written last-panic.txt
replays in a process started by a provider query with no activity.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The fling phase called List::tick_fling itself every 16ms, so on a 120Hz
phone every second frame redrew a position already drawn -- Iris saw the
benchmark scroll visibly less smoothly than her own finger, and it was
the rig rather than the renderer. A real fling is advanced once per frame
by UiData::tick_animations from the frame callback, so the phase now
starts one the way a gesture does (fling + animate) and polls
is_scrolling to know when it settled. ANIM_STEP_MS becomes POLL_MS,
which is what it always was here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris's call, after the guest measurement: the emulator is a GLES rig and
nothing chases hardware Vulkan in it; the Vulkan path is covered by the
desktop build and by her phone.
Nothing had to be forced. The emulator has no hardware Vulkan at all --
its only Vulkan is SwiftShader in software -- and its GLES is the host's
real RX 7900 XT through virgl at ES 3.1, so iris's existing runtime
fallback lands there by itself. Verified end to end with an ordinary
debug APK: "no Backends(VULKAN|...) adapter on this device, falling back
to GLES", then "Android Emulator OpenGL ES Translator (virgl (AMD Radeon
RX 7900 XT ...)) (Gl, OpenGL ES 3.1 ...) on Backends(GL)". So the
emulator and the phone run the same binary, differing only in what it
finds -- which is the point, and `force-gles` must not be reintroduced to
arrange the emulator's backend.
What changed:
- The Android renderer logs the full adapter line at startup, as the
desktop already did. Only the backend enum was logged, which cannot
tell `Gl` on the host's GPU from `Gl` on SwiftShader; the same rule was
written on one member of the pair and not the other.
- `run-bench.sh` prints that line before any number.
- build-apk.sh, Cargo.toml and RUST.md's "Vulkan in the emulator" carried
the stale premise that the emulator defaults to software Vulkan and has
to be steered off it. The recipes are marked superseded rather than
deleted, since the record of why host Vulkan is unavailable is still
worth having.
- Drive-by: an `#[allow]`-free clippy warning in android/platform.rs
(useless JObject conversion) that only appears on the android target.
No Vulkan requirement was found in iris itself to remove: neither backend
asks for a feature, `device_limits()` stays at wgpu's defaults with the
compute fields zeroed, and both probe rather than expect an adapter.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`gpu-probe` cross-compiled with cargo-ndk and run inside a default
`emu up`: the guest's GL adapter is the host's real RX 7900 XT through
virgl, reporting OpenGL ES 3.1 with compute shaders, 1024 invocations per
workgroup and 64 KB of workgroup storage -- the same numbers the desktop
gets. `Backends::PRIMARY` still finds nothing, because the guest's only
Vulkan is SwiftShader.
So GPU acceleration in the emulator is not a thing to get working; it is
the default, and it is GLES. What is missing is GPU-accelerated Vulkan,
and the Venus retry on mesa 26.2.2 fails exactly as it did on 26.1.7 with
no newer emulator package to try.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Compute is a downlevel capability rather than a feature -- unconditional
on any Vulkan 1.0 device, GLES 3.1 and up -- so the question shadows and
blur raise is what the *weakest* adapter iris can fall back to offers.
Both here answer yes: Venus and virgl each report COMPUTE_SHADERS, 1024
invocations per workgroup and 64 KB of workgroup storage, virgl because
it is ES 3.2. The only no-compute machine in this project is the
emulator's SwiftShader software GL at ES 3.0.
RUST.md gains that table, why DRM native context is unrelated to it, and
what each of shadows/blur/paths actually needs -- only vello proper turns
the compute question on.
`mask_sdf` SIGSEGVd after printing `test result: ok`, and the workaround
was to hand the device to the process with `mem::forget` on the reading
that "dropping a wgpu device on Venus segfaults". Every part of that
except the symptom was wrong.
`rigs/gpu-probe`'s new `teardown` bin is the experiment, one variable per
mode: the same open-and-close exits 0 on the main thread and SIGSEGVs on
a spawned one; it needs no GPU work and no device, only an instance; raw
`ash` does it with no wgpu involved at all; and keeping the instance
alive fixes it. Destroying the last VkInstance makes the loader dlclose
the ICD, and Mesa's ICD here registers a pthread_key_create destructor
into its own text without `-z nodelete`, so glibc calls it through
unmapped memory when the thread exits. libtest runs every #[test] on a
spawned thread, which is the whole reason this looked like a drop bug.
`VK_LOADER_DISABLE_DYNAMIC_LIBRARY_UNLOADING=1` confirms the mechanism.
So the fix is one `wgpu::Instance` for the process -- what wgpu asks for
anyway -- and the device, queue and everything else drop normally again.
The escape and its paragraph of reasons are gone.
Also: the machine-level graphics notes duplicated in docs/RUST.md,
run-headless.sh and two source comments now point at the
`this-machine-graphics` skill, which is the only copy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The GPU findings from 2026-09-08 would bite any project on this machine,
not just this one, so they are now a skill (AGENTS.md's own rule about
where a machine-wide lesson belongs). This section keeps the iris- and
port-specific half and names the skill for the rest.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three findings from one morning, all of them things that were invisible
rather than wrong. docs/RUST.md's two new sections have the full account.
**`cargo test --workspace` was taking half an hour, and it was debug
info.** rustc's default `debug = true`, times eight test binaries each
statically linking the whole wgpu + naga + winit + parley graph, means
every one of them gets a private copy of that graph's DWARF written into
it: the linkers for one run had written ~54 GB between them and were
still going at thirty minutes -- the worst single one 16.9 GB for one
test binary -- leaving an 88 GB target/. It was not CPU: the machine was
87% idle, and rust-lld's threads were in D state in btrfs
`handle_reserve_ticket`, blocked on space reservation at 83% full. So
`debug = "line-tables-only"` on both `profile.dev` and `profile.test` --
both, because `cargo test` builds dependencies under one and the test
targets under the other. Cold, with all 19 suites run: 69 s and a 3.7 GB
target. Backtraces keep file and line; `RUSTFLAGS="-C debuginfo=2"` per
run buys back variable inspection when a debugger actually needs it.
**The desktop had no logger at all**, so every `log::` call on that side
went to `log`'s no-op default -- including the GLES fallback warning
added hours earlier. `DefaultApp::run` installs a stderr logger
(`src/default/logging.rs`, no new dependency: a level and a line is a
page of code against env_logger plus its filter dialect), and the
renderer now says which adapter won at `info`. That line is the point:
with a silent fallback, a layer-2 screenshot rendered by llvmpipe and one
rendered by the host's GPU are the same PNG, and which one it was is
exactly what the screenshot is being taken to judge.
**`tests/mask_sdf.rs` is a render pass now, not a compute pass.** It
asked for `adapter.limits()` because `iris_core::device_limits()`
deliberately zeroes the six `max_compute_*` fields -- a decision on
record since 2026-09-05, which this quietly worked around instead of
following. It now asks for what iris asks for and calls the function from
the fragment stage, where the renderer calls it. The compute pass was
*not* why it crashed, and the record should not say it was: the rewrite
crashes identically. What the crash is: dropping a wgpu device on this
VM's Venus adapter segfaults, after the test has produced its answer
(worst CPU/shader disagreement 5.8e-6). Narrowed -- plain Vulkan creating
and destroying five VkDevices on the same adapter is clean, and the same
binary with Vulkan hidden falls back to GL and exits clean. Worked around
at `Gpu::leak`, with the reason and the delete-me condition written
there.
`rigs/virtgpu-probe` is the new rig behind the Venus half: which capsets
the host offers (0x16 -- VIRGL, VIRGL2, VENUS; no capset 6, so no DRM
native context without host-side work), whether the device has compute
(it does: 1024 invocations/workgroup -- the "no compute" finding on
record is about the Android emulator's SwiftShader, a different machine),
and whether plain Vulkan teardown is clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Iris, on the code fence: "the code block scrolling currently masks in an
inner rectangle. Ideally masks should have a shape associated with them,
rounded rectangle being one of them ... so that the mask becomes the
parent container with rounded edges. Make sure alpha works properly with
it, eg. on the corners where alpha should be decreased / multiplied."
`Mask` is now `{ primitive, parent }` -- the slot of a primitive already
written, plus the mask this one nests inside. The fragment stage
evaluates that primitive's own coverage at the masked pixel, through the
same `rounded_rect_coverage` a drawn rect goes through, and multiplies it
into the alpha along the whole `parent` chain. Nothing about the shape is
copied, so a rounded container's corner and its children's clipped corner
are one piece of arithmetic and cannot drift; two nested feathers dim a
pixel twice, which is the multiply she asked for.
`.masked()` is unchanged for callers: it writes an undrawn rect
(`Drawn::No`/`NOT_DRAWN` -- owned, moved, resized and freed like any
other primitive, simply never rasterized) and points at that, so square
clipping is the same mechanism rather than a special case. New
`.masked_by(shape)` draws `shape` behind the content in its own layer and
clips to the first primitive it drew, with no radius written twice; it
replaces `.masked().background(w)`, which drew both and clipped to the
box. `transcript-ui`'s `BlockFrame::Verbatim` is the first caller.
Hit-testing applies the shape (`SensorUi::run_sensors` ->
`UiRenderState::mask_admits`, coverage above one half, which is where the
drawn edge is), as well as the widget's own box -- the two ask different
questions and both have to hold. `primitive_corners` is a floor-for-floor
transliteration of the shader's `corners_of`, not `region.to_px()`: the
phone's 2.55 density puts nothing on a whole pixel, and skipping the
rounding disagrees with the pixels by up to one along each edge.
A mask's shape must be a rect, asserted by name in `set_mask_to`. A glyph
would need a CPU-side alpha plane before the hit test could agree with
the shader, and a standalone image a bind-group switch the fragment stage
cannot make. So no texture mask exists; the branch where one would go is
in both copies of `mask_coverage`. docs/LAYOUT.md's section end lists this
and the three other places the code is narrower than the design.
Tests. Layer 1, `layout_tests.rs`: the child's coverage swept across the
container's corner arc equals the container's own exactly; nested masks
multiply rather than intersect, asserted where both feathers are partial,
which is the only place the two differ; a press in a rounded-away corner
misses while one inside the curve and one on a straight edge hit; and
`a_plain_mask_still_clips_to_a_square_box`, the half this had no reason to
touch. The first version of the corner test swept the straight chord
between the arc's ends, which lies inside the circle everywhere -- it
proved nothing and said so, which is why it counts both sides now.
`iris/tests/mask_sdf.rs` is the only test here that needs a GPU: it lifts
`distance_from_rect` and `rounded_rect_coverage` out of
`iris_core::SHAPE_SHADER` by name -- lifted, not copied, since a copy
would be edited alongside the shader -- and runs them in a compute pass
over ~200k points at five radii against `iris_core::rounded_rect_coverage`.
Worst disagreement under 1e-5; the negative control (`+ 0.01` inside the
shader's smoothstep) fails it at 0.03.
Layer 2 for looking: `./run-headless.sh phone --phone --shot /tmp/mask.png
--seconds 6 -- -p transcript-fixture` draws the fixture's horizontally
scrolled code fence clipped on the curve at both top corners.
Two things found on the way and fixed here:
- The winit backend had the defect the Android one was fixed for in
85869d0 -- `Backends::PRIMARY` and an `.expect` on the adapter. This
VM's Venus device disappears when the host runs out of virgl contexts,
which happened mid-task, and layer 2 aborted with `Could not get
adapter!` while GL sat there working. It probes and rebuilds the
instance on `Backends::GL` exactly as Android does now, and the request
names the backends it tried. The rule had been written on one member of
a set of two.
- `active_primitive_count` counted mask shapes, so `iris::frame`'s
`primitives=` -- a number Iris reads off a phone report as "how much is
on screen" -- would have gained one per masked widget.
`widget_trait!` now accepts a `///` doc comment on its functions, since
`masked_by` is public API and rustdoc is where a contract is read.
docs/LAYOUT.md, docs/RUST.md (both queue boxes, the commands, and where
the GPU test sits among the three layers), docs/IRIS.md and
docs/IRIS_TODO.md ("Masks defined relative to each other", now closed).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both docs/IRIS_TODO.md's night bullets and docs/RUST.md's queue items
covered by the two client-core/iris-android-app commits above.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
`iris::diagnostics::set_trace` landed with nothing to press it. It is the
bench header's fourth control now, reading `Trace off` or `Trace on` --
a toggle whose own appearance never changes is a button that looks like
it did nothing. Its accessibility label stays the fixed "Trace input and
frames", because that is what `run-bench.sh` and `ui-trace --do "tap
'...'"` find it by and a control that renames itself when pressed is one
no script can find twice. Pressing it rebuilds the header and shows the
diagnostics pane, so the state is on screen at the moment of the press.
Both reports carry `trace_line`, from the flag read at the *start* of
what is being reported as well as at the end: the switch is on screen
while a benchmark runs, so "somebody moved it half way through" is a
state that happens, and reported as either "on" or "off" it would be a
confident sentence about a log covering half the run.
The row's type size is one constant for all four labels and drops from
18 to 13: with a fourth control the labels overlapped each other on a
1080px screen. Shrinking one label to fit is what the UI rules forbid;
resizing the row is a layout decision and all four still match.
Checked on the emulator: the switch flips its own text and colour, the
pane reads "input/frame trace: on", and `iris::frame`/`iris::input`
lines appear in the ring only after it is pressed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Iris's call once the upload route was working: put it in Dev Updater
properly. So the app now exposes its own ring through a ContentProvider
at `<applicationId>.devlog` -- Dev Updater's contract, written down in
that project's README, not something invented here -- and Dev Updater's
phone app reads it on the same device and forwards it to its own build
machine. No tunnel, no token, no second enrolment, and any app that
server delivers can implement the same and get the same Runtime tab.
`DevLogProvider.java` plus `devlog.rs` are the platform glue only: a flat
`String[]` across JNI, a `MatrixCursor` on the Java side, and
`nativeReady` telling Rust the authority the provider actually
registered, so the Diagnostics pane can name somewhere a reader can
query rather than composing a guess. `LogRing::newest_seq()` is the one
addition in `client-core`: an in-memory ring starts again at zero, so it
is what lets a reader notice the process restarted instead of silently
skipping everything since.
Deleted with it, so there is one mechanism: `client_core::log_upload`,
`POST /client-log` on ai-server, the `AI_APP_LOG_*` baking (which left
`build.rs` with nothing to do), and the uploader on both Android
clients. Kept: the ring, `RingLogger`, `install_process_logger`, and the
Diagnostics line -- whose second half is now `devlog provider:
content://<authority>`.
Verified end to end on this checkout's emulator: iris's own
`iris::android::view` startup lines read out of the provider by the
shell, forwarded by Dev Updater's Runtime tab, and served back from
`GET /apps/android-app/components/app/logs?kind=runtime`. A component
whose package has no provider says so in as many words.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/REVIEW-2026-09-07.md's D5. Four places quoted 11750 px/s as the old
average estimator's answer -- for `flick-120hz.touch` *and* for the
press-plus-one-move-frame set, which are different sample sets, and one
number in both rows is the tell. `iris/benches/velocity_reference.py`,
which the same section says every number below it comes from, prints
12250 for the recording and 12500 for the two-sample set, and
`sense.rs:1406` already had the 12250.
Half of where 11750 came from is recoverable and is written down beside
the table: it is the recording's 196 px over 16.68 ms, a 60 Hz frame
rather than the 16 ms span the file itself records. That explains the
flick row; the other row was copied from it. The 1.30x ratio derived from
it becomes 1.24x.
Also settles the second disagreement about the same experiment (the
review's rule finding on the negative control): `sense.rs`'s doc comment
claimed reverting `velocity` to total-over-span fails "exactly this one,
the flick recording, and phone_screen.rs" while RUST.md said seven. Run
again today with the revert in place: seven in `-p iris` (the flick
recording, the accelerating flick, the horizon, the stopped finger, the
minimum sample count, both `drag_gesture` flick tests) plus
`phone_screen.rs`'s flick, everything else green. RUST.md was right and
the comment now says the same thing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Iris asked for a button to copy raw input events and per-frame timings
through the same report Copy report already produces. sense::log_input_event
(one line per platform pointer sample, historical samples inline on
Android) and diagnostics::log_frame (one line per frame: frame number,
frame clock, time since last input, layout/draw durations, redraw kind,
primitives on screen, animating) both land under iris::diagnostics's
trace_enabled() gate, off by default since the ring is 2000 lines/256KiB
and either target at 120Hz fills it in seconds. report_to_touch.py turns
a report's iris::input lines back into a .touch file for harness/desktop
replay, round-tripped in transcript-fixture's input_log_roundtrip test.
Folds in docs/REVIEW-2026-09-07.md's D1: four older per-frame debug!
lines (android::view's two render() lines, list.rs's fling tick,
text/mod.rs's text render) were unconditional at Debug and, with the
ring's RingLogger recording everything the app's Debug install lets
through regardless of target, filled it before Copy report ever saw
anything else. All four (and sense.rs's drag-release-samples line) are
now behind the same gate. The same test proves both directions: tracing
off leaves zero Debug lines from a replayed flick, tracing on produces
the expected iris::input/iris::frame lines with real durations.
Not wired to a Diagnostics-pane button: bench_client.rs is open under
another agent. set_trace(bool) is the whole surface a control needs.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
DECISIONS.md gets the decision with both rejected options and what the
longer link measures (89 -> 652 bytes, a 45x23 QR -> 93x47), RUST.md ticks
the enrolment queue item and marks the log-upload route superseded rather
than editing it, and IRIS.md says what changed for anyone building the
Android app.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
iris
