iris: on-demand message-list/image benchmarks, and two O(N) findings

IRIS_TODO.md's "Benchmarks" item: a message list of N wrapped-text rows
(first-frame cost), scrolling it, and growing an input box above which
the list must move rather than re-layout -- all as a plain, harness=false
`cargo bench` binary (iris/benches/message_list.rs) since UiRenderState
touches no GPU or window, chosen over criterion because every scenario
here reduces to a count take_counters already answers exactly, and a
new dependency wasn't worth it. Scroll (200 ticks) and the input-grow
case (40 lines) are flat across N=100/1,000/10,000: LAYOUT.md's O(1)
move chain holds.

The many-images case (d) needs a real wgpu device, so it's a headless
example (iris/examples/bench_images.rs) plus a new
GpuTextures/UiRenderNode counter, take_image_bind_group_creates,
mirroring take_counters. It found two real non-O(1) costs, recorded as
new Fix items rather than redesigned: bind-group creation takes two
frames to settle after a cold load instead of one, and appending a
single image to an already-loaded 1,000-image list rebuilds all 1,000
existing bind groups (masks/move_offsets buffer growth triggers
rebuild_image_bind_groups unconditionally).

run-bench.sh wraps both. Numbers and commands are in IRIS_TODO.md.

cargo fmt --all -- --check, cargo clippy --all-targets, and
cargo test --workspace (19 passed) all clean; benches are not run by
cargo test.

Co-Authored-By: Claude Sonnet <noreply@anthropic.com>

IRIS_TODO.md

@@ -26,18 +26,122 @@ order and what "done" looks like. Tick and date them in place.
   still reaches the button — confirmed to fail on the pre-fix code and
   pass after.
 
+- [ ] **Appending one image to an already-loaded list rebuilds every other
+  image's bind group (2026-09-05).** Found by the benchmark below, not
+  designed against: `GpuTextures::update` (`core/src/render/texture.rs`)
+  triggers `rebuild_image_bind_groups` — a loop over *every live
+  standalone image*, rebuilding its `BindGroup` — whenever the shared
+  `masks` or `move_offsets` GPU buffer is resized (`masks_resized ||
+  moves_resized` in `UiRenderNode::update`, `core/src/render/mod.rs`), and
+  a widget getting its *first* move-offset slot (LAYOUT.md section 2 —
+  every widget gets one on first draw) can be exactly what grows that
+  buffer. So one new message with one new image, appended to a transcript
+  that already has N images loaded, does not cost O(1): it costs one
+  `create_image` for the new image plus one `make_image_bind_group` per
+  *existing* image, because the new widget's own move slot pushed the
+  arena past its capacity. Measured directly in
+  `iris/examples/bench_images.rs`: appending a 1,001st image to 1,000
+  already-settled ones reports **1,001** bind-group creates for that one
+  frame, not 1 (`./run-bench.sh images`, frame 5 in the transcript below).
+  This is the same class of cost LAYOUT.md's move chain exists to avoid
+  elsewhere in the codebase, just not yet closed off here — the fix is
+  presumably to size `masks`/`move_offsets` with headroom (the array
+  texture already grows by doubling, `grow_array`, for the same reason) so
+  an ordinary append does not cross a capacity boundary, or to stop tying
+  the *image* bind group's contents to a buffer that changes on every new
+  widget in the whole tree, image or not. Not designed further here per
+  the "do not redesign, record it" instruction this benchmark was built
+  under.
+- [ ] **Bind-group creation takes two frames to reach the steady state, not
+  one (2026-09-05).** Same benchmark: loading 1,000 images cold reports
+  1,000 creates on frame 1 (expected — this is `create_image`, one per
+  new image) *and again* 1,000 on frame 2, with nothing between the two
+  frames marked dirty, before settling to 0 from frame 3. The second
+  frame's 1,000 is `rebuild_image_bind_groups` again, for the same
+  masks/move-offsets buffer-growth reason as the item above — the arena
+  apparently does not finish growing to its steady size within the first
+  frame the tree is drawn. Not chased further; recorded so whoever fixes
+  the item above checks whether the fix also closes this one, since they
+  look like the same root cause measured two different ways.
+
 ## Build
 
-- [ ] **Benchmarks**, not unit tests, run on demand (a `benches/` or a
+- [x] **Benchmarks**, not unit tests, run on demand (2026-09-05; a
-  script under `iris/`, never in `cargo test`). The scenario that matters
+  `benches/` or a script under `iris/`, never in `cargo test`). The
-  most is a **message list** — chat apps and this app's transcript alike —
+  scenario that matters most is a **message list** — chat apps and this
-  stressed with many messages and many images. One case in particular:
+  app's transcript alike — stressed with many messages and many images.
-  **resizing an input box** (typing enough text to grow it) that pushes a
+  One case in particular: **resizing an input box** (typing enough text to
-  long list of messages above it must stay very fast and recalculate
+  grow it) that pushes a long list of messages above it must stay very
-  almost nothing — a move of everything above, not a re-layout. That is
+  fast and recalculate almost nothing — a move of everything above, not a
-  exactly the O(1) move chain in LAYOUT.md; the benchmark is what proves
+  re-layout. That is exactly the O(1) move chain in LAYOUT.md; the
-  it. Done when the numbers are in this file with the command, and the
+  benchmark is what proves it. Done when the numbers are in this file with
-  input-box case reports draws re-run, not just frame time.
+  the command, and the input-box case reports draws re-run, not just frame
+  time.
+
+  **Built as two rigs**, chosen per scenario by whether a real `wgpu`
+  device is needed (`UiRenderState`/`Widgets` touch no GPU or window, so
+  most of this runs as an ordinary binary — the same property
+  `layout_tests.rs` relies on):
+
+  - `iris/benches/message_list.rs` — a plain `Instant`-timed binary
+    (`[[bench]] harness = false` in `iris/Cargo.toml`), not criterion: see
+    the file's own header for why (short version — every scenario here
+    reduces to a *count* `UiRenderState::take_counters` already produces,
+    which criterion's statistical machinery adds nothing to and which a
+    new dependency is not worth pulling in for). Covers (a) first-frame
+    cost of a message list of N wrapped-text rows (one in 20 also carrying
+    a small in-memory image) for N = 100/1,000/10,000; (b) per-frame cost
+    of scrolling that list, 200 ticks; (c) the input-box case — a
+    fixed-height field at the bottom of the screen growing by a line 40
+    times, with the message list above it filling the rest of the screen.
+    Run: `cd iris && cargo bench --bench message_list` (always release —
+    `cargo bench` builds the `bench` profile, which is optimized).
+  - `iris/examples/bench_images.rs` — needs a real device, so it runs
+    through `iris/run-headless.sh bench_images`, printing
+    `UiRenderNode::take_image_bind_group_creates()` (a new counter, added
+    in `core/src/render/texture.rs` and `core/src/render/mod.rs`,
+    mirroring `UiRenderState::take_counters`) each frame. Covers (d): 1,000
+    image rows, checked both cold (does bind-group creation reach zero
+    once loaded) and after appending one more image once settled (does
+    *that* stay cheap) — the second question is what actually matters for
+    a live transcript and is what turned up the two Fix items above.
+  - `iris/run-bench.sh [list|images]` runs either or both and is what to
+    run before/after touching `Scroll`, `Span`, `Sized`, the move-offset
+    chain, or `GpuTextures`.
+
+  **Numbers (2026-09-05, release, `cargo bench`/`run-headless.sh`, this
+  VM: AMD Ryzen 7 3800X, 8 cores, rustc 1.98.0 nightly-2026-09-03):**
+
+      cd iris && cargo bench --bench message_list
+      (a) first frame, N=100:    30.30ms  draws=227   rewrites=15   moves=0
+      (a) first frame, N=1000:  186.04ms  draws=2252  rewrites=150  moves=0
+      (a) first frame, N=10000:1770.36ms  draws=22502 rewrites=1500 moves=0
+      (b) scroll, N=100/1000/10000, 200 ticks each:
+          draws=200 rewrites=0 moves=200 (identical at every N)
+          per-tick average: 0.0002ms (identical at every N)
+      (c) input grows 40 lines, N=100/1000/10000 rows above it:
+          draws=320 rewrites=40 moves=160 (identical at every N)
+          per-line average: 0.0012-0.0013ms (identical at every N)
+
+      cd iris && ./run-bench.sh images
+      frame=1 bind_group_creates=1000   (cold load)
+      frame=2 bind_group_creates=1000   (see Fix item above)
+      frame=3 bind_group_creates=0
+      frame=4 bind_group_creates=0
+      (append one image here)
+      frame=5 bind_group_creates=1001   (see Fix item above)
+      frame=6 bind_group_creates=0
+
+  **Reading it**: (a) is real, necessary work — shaping and laying out N
+  never-before-seen text rows — and scales with N as it must, ~10x cost
+  per 10x N. (b) and (c) are the pass conditions that matter: both are
+  **exactly flat across N = 100 to 10,000**, confirming LAYOUT.md's O(1)
+  move chain holds for both scrolling and for a growing input box pushing
+  the message list — draws/moves per tick or per line do not grow with
+  list size, and the per-operation cost (a fraction of a microsecond) is
+  nowhere near a frame budget. (d)'s cold-load and steady-state halves
+  behave as designed; its *append* half did not, which is the two Fix
+  items above.
 - [ ] **Masks defined relative to each other.** Wanted: mask A multiplies
   by something *and also* applies mask B — a mask can reference a parent
   mask, the way the move chain references a parent offset. Today masks

iris/Cargo.toml

@@ -20,6 +20,14 @@ tokio = { workspace = true, features = ["sync", "rt", "rt-multi-thread"] }
 [dev-dependencies]
 tokio = { workspace = true, features = ["sync", "rt", "rt-multi-thread", "time"] }
 
+# Plain Instant-timed binaries, not criterion -- see benches/message_list.rs's
+# header for why. `harness = false` opts out of the unstable `#[bench]`
+# test-crate harness cargo would otherwise want, in favour of an ordinary
+# `fn main()`.
+[[bench]]
+name = "message_list"
+harness = false
+
 [workspace]
 members = ["core", "macro"]
 

iris/benches/message_list.rs

@@ -0,0 +1,258 @@
+//! On-demand benchmarks for iris's message-list scenario -- IRIS_TODO.md's
+//! "Benchmarks" item. Never run by `cargo test`; run explicitly with
+//! `cargo bench --bench message_list --release` or `./run-bench.sh`.
+//!
+//! **Why a plain `Instant`-timed binary, not criterion.** Every scenario
+//! here is really "how many `Widget::draw` calls and primitive rewrites did
+//! this frame cost," which `UiRenderState::take_counters` already answers
+//! exactly (see `iris/src/layout_tests.rs`, which this file's harness
+//! mirrors). A short loop that times itself and prints the counters
+//! alongside the wall time says everything criterion's warm-up/sampling/
+//! outlier-removal machinery would add on top, for scenarios that are
+//! fundamentally about a *count*, not a noisy microbenchmark distribution
+//! -- and it avoids a new dependency this crate does not otherwise need.
+//! Per the code rules, the plain option is also the one shorter to explain.
+//!
+//! Scenarios (LAYOUT.md's O(1) move chain, and IRIS_TODO.md's "Benchmarks"
+//! wording):
+//!
+//! - (a) first-frame cost of a message list of N wrapped-text rows, some
+//!   with an image, for N = 100 / 1,000 / 10,000.
+//! - (b) per-frame cost of scrolling that list -- must be O(1) moves, not
+//!   re-layout.
+//! - (c) the input-box case: growing a fixed-height field at the bottom of
+//!   the screen must move the message list above it, not re-lay its rows.
+//!   Reports frame time *and* the draw/rewrite/move counters LAYOUT.md
+//!   section 8 defines.
+//!
+//! (d), many images with zero steady-state bind-group creation, needs a
+//! real `wgpu` device and lives in `iris/examples/bench_images.rs` instead,
+//! driven through `run-headless.sh` -- see that file's header.
+//!
+//! `UiRenderState`/`Widgets` touch no GPU or window (as `layout_tests.rs`
+//! notes), so everything here runs as an ordinary `--release` binary with
+//! no compositor. Numbers are recorded in IRIS_TODO.md, not here -- this
+//! file is the rig, not the result.
+
+use iris::prelude::*;
+use std::time::Instant;
+
+/// The minimal `UiRsc` a benchmark needs -- identical in shape to
+/// `layout_tests.rs`'s `TestRsc`.
+struct BenchRsc {
+    ui: UiData,
+}
+
+impl UiRsc for BenchRsc {
+    fn ui(&self) -> &UiData {
+        &self.ui
+    }
+    fn ui_mut(&mut self) -> &mut UiData {
+        &mut self.ui
+    }
+}
+
+/// Long enough to force real wrapping at a phone-plausible column width, and
+/// varied enough (no two rows byte-identical) that nothing can special-case
+/// on repeated content.
+const BODY: &str = "The quick brown fox jumps over the lazy dog. Iris lays \
+out wrapped text by shaping once per width and caching the result, so a \
+row that is offered the same width twice does not reshape. This sentence \
+exists only to give a row enough text to wrap across several lines at a \
+typical phone column width.";
+
+/// One message row: a wrapped `Text`, and every `image_every`th row also an
+/// `Image` beneath it -- a small in-memory RGBA square rather than a file,
+/// so N=10,000 rows costs no disk I/O.
+fn build_row(rsc: &mut BenchRsc, i: usize, image_every: usize) -> StrongWidget {
+    let mut text = Text::new(format!("Message {i}: {BODY}"));
+    text.wrap = true;
+    let text = rsc.ui.widgets.add_strong(text).any();
+
+    if image_every > 0 && i.is_multiple_of(image_every) {
+        let img = image::DynamicImage::new_rgba8(64, 64);
+        let image_widget = image::<BenchRsc>(img)(rsc);
+        let image_widget = rsc.ui.widgets.add_strong(image_widget).any();
+        let mut row = Span::empty(Dir::DOWN);
+        row.push(text);
+        row.push(image_widget);
+        rsc.ui.widgets.add_strong(row).any()
+    } else {
+        text
+    }
+}
+
+/// A `Scroll` over `n` message rows, one in `image_every` of them carrying
+/// an image (0 disables images entirely). Returns the scroll widget (weak,
+/// so the caller can drive it) and the erased root to render.
+fn build_list(
+    rsc: &mut BenchRsc,
+    n: usize,
+    image_every: usize,
+) -> (WeakWidget<Scroll>, StrongWidget) {
+    let mut span = Span::empty(Dir::DOWN);
+    for i in 0..n {
+        span.push(build_row(rsc, i, image_every));
+    }
+    let span = rsc.ui.widgets.add_strong(span);
+    let scroll = rsc.ui.widgets.add_strong(Scroll::new(span.any(), Axis::Y));
+    (scroll.weak(), scroll.any())
+}
+
+fn report(label: &str, elapsed: std::time::Duration, draws: u64, rewrites: u64, moves: u64) {
+    println!(
+        "{label}: {:.2}ms  draws={draws} rewrites={rewrites} moves={moves}",
+        elapsed.as_secs_f64() * 1000.0
+    );
+}
+
+/// (a) First-frame cost of a message list of N rows.
+fn bench_first_frame(n: usize) {
+    let mut rsc = BenchRsc {
+        ui: UiData::default(),
+    };
+    let (_scroll, root) = build_list(&mut rsc, n, 20);
+    let mut render = UiRenderState::new();
+    render.resize((1080.0, 2000.0));
+
+    let start = Instant::now();
+    render.update(&root, &mut rsc);
+    let elapsed = start.elapsed();
+    let (draws, rewrites, moves) = render.take_counters();
+    report(
+        &format!("(a) first frame, N={n}"),
+        elapsed,
+        draws,
+        rewrites,
+        moves,
+    );
+}
+
+/// (b) Per-frame cost of scrolling an already-laid-out list of N rows.
+/// Warms up (as `layout_tests.rs`'s scrolling test documents: `Scroll`
+/// needs one no-op tick before a real scroll becomes a same-size move
+/// rather than a resize), then times a run of individual scroll ticks.
+fn bench_scroll(n: usize, ticks: usize) {
+    let mut rsc = BenchRsc {
+        ui: UiData::default(),
+    };
+    let (scroll, root) = build_list(&mut rsc, n, 20);
+    let mut render = UiRenderState::new();
+    render.resize((1080.0, 2000.0));
+    render.update(&root, &mut rsc);
+    rsc.ui.widgets.get_mut(&scroll).unwrap().scroll(0.0);
+    render.update(&root, &mut rsc);
+    render.take_counters();
+
+    let mut total = std::time::Duration::ZERO;
+    let mut total_draws = 0u64;
+    let mut total_rewrites = 0u64;
+    let mut total_moves = 0u64;
+    for _ in 0..ticks {
+        rsc.ui.widgets.get_mut(&scroll).unwrap().scroll(-8.0);
+        let start = Instant::now();
+        render.update(&root, &mut rsc);
+        total += start.elapsed();
+        let (draws, rewrites, moves) = render.take_counters();
+        total_draws += draws;
+        total_rewrites += rewrites;
+        total_moves += moves;
+    }
+    report(
+        &format!("(b) scroll, N={n}, {ticks} ticks (totals; expect draws/moves independent of N)"),
+        total,
+        total_draws,
+        total_rewrites,
+        total_moves,
+    );
+    println!(
+        "    per-tick average: {:.4}ms",
+        total.as_secs_f64() * 1000.0 / ticks as f64
+    );
+}
+
+/// (c) The input-box case: a fixed-height field at the bottom of the screen
+/// growing by a line at a time, with a message list of N rows filling the
+/// rest of the screen above it. Growing the input shrinks the *offered*
+/// height of the scroll container (a single widget, from the outer
+/// `Span`'s point of view) without changing the width it offers its
+/// content -- so the rows underneath, which only care about width, must
+/// not redraw; the scroll's own re-registration of where its content sits
+/// is the one O(1) move this is checking for. See LAYOUT.md's `Scroll`
+/// design note on offering the child last frame's content length, which is
+/// exactly what keeps this a move instead of a reflow.
+fn bench_input_grows(n: usize, lines: usize) {
+    let mut rsc = BenchRsc {
+        ui: UiData::default(),
+    };
+    let (scroll, list_root) = build_list(&mut rsc, n, 20);
+    let list_area = rsc.ui.widgets.add_strong(Sized {
+        inner: list_root,
+        x: None,
+        y: Some(rest(1.0)),
+    });
+
+    let line_height = 24.0;
+    let input_rect = rsc.ui.widgets.add_strong(Rect::new(UiColor::WHITE));
+    let input_area = rsc.ui.widgets.add_strong(Sized {
+        inner: input_rect.any(),
+        x: None,
+        y: Some(abs(line_height)),
+    });
+
+    let input_area_weak = input_area.weak();
+    let mut root_span = Span::empty(Dir::DOWN);
+    root_span.push(list_area.any());
+    root_span.push(input_area.any());
+    let root = rsc.ui.widgets.add_strong(root_span).any();
+
+    let mut render = UiRenderState::new();
+    render.resize((1080.0, 2000.0));
+    render.update(&root, &mut rsc);
+    rsc.ui.widgets.get_mut(&scroll).unwrap().scroll(0.0);
+    render.update(&root, &mut rsc);
+    render.take_counters();
+
+    let mut total = std::time::Duration::ZERO;
+    let mut total_draws = 0u64;
+    let mut total_rewrites = 0u64;
+    let mut total_moves = 0u64;
+    for line in 1..=lines {
+        rsc.ui.widgets.get_mut(&input_area_weak).unwrap().y =
+            Some(abs(line_height * (line + 1) as f32));
+        let start = Instant::now();
+        render.update(&root, &mut rsc);
+        total += start.elapsed();
+        let (draws, rewrites, moves) = render.take_counters();
+        total_draws += draws;
+        total_rewrites += rewrites;
+        total_moves += moves;
+    }
+    report(
+        &format!(
+            "(c) input grows by {lines} lines above N={n} rows (totals; \
+             draws/rewrites must not scale with N)"
+        ),
+        total,
+        total_draws,
+        total_rewrites,
+        total_moves,
+    );
+    println!(
+        "    per-line average: {:.4}ms",
+        total.as_secs_f64() * 1000.0 / lines as f64
+    );
+}
+
+fn main() {
+    println!("iris message-list benchmark -- release build, this machine's CPU");
+    for &n in &[100usize, 1_000, 10_000] {
+        bench_first_frame(n);
+    }
+    for &n in &[100usize, 1_000, 10_000] {
+        bench_scroll(n, 200);
+    }
+    for &n in &[100usize, 1_000, 10_000] {
+        bench_input_grows(n, 40);
+    }
+}

iris/core/src/render/mod.rs

@@ -441,4 +441,11 @@ impl UiRenderNode {
     pub fn view_count(&self) -> usize {
         self.textures.view_count()
     }
+
+    /// Standalone-image bind groups built since the last call -- see
+    /// `GpuTextures::take_bind_group_creates`. Call once per frame before
+    /// `update()` to measure exactly that frame.
+    pub fn take_image_bind_group_creates(&mut self) -> u64 {
+        self.textures.take_bind_group_creates()
+    }
 }

iris/core/src/render/texture.rs

@@ -61,6 +61,14 @@ pub struct GpuTextures {
     /// nothing of its own to put there: rects and glyphs never sample it,
     /// but the layout requires something bound regardless.
     null_view: TextureView,
+
+    /// Standalone-image bind groups actually built (`create_image`'s own
+    /// build, or one per slot touched by `rebuild_image_bind_groups`) since
+    /// the last `take_bind_group_creates`. IRIS_TODO.md's "many images"
+    /// benchmark reads this to prove the steady-state cost of an
+    /// unchanging image list is zero, the same way `UiRenderState`'s
+    /// `draw_count`/`region_mut_count` prove the layout side.
+    bind_group_creates: u64,
 }
 
 impl GpuTextures {
@@ -297,12 +305,13 @@ impl GpuTextures {
                     masks,
                     move_offsets,
                 );
+                self.bind_group_creates += 1;
             }
         }
     }
 
     fn create_image(
-        &self,
+        &mut self,
         image: &DynamicImage,
         rsc_layout: &BindGroupLayout,
         masks: &ArrBuf<Mask>,
@@ -339,6 +348,7 @@ impl GpuTextures {
             masks,
             move_offsets,
         );
+        self.bind_group_creates += 1;
         ImageGpu {
             texture,
             view,
@@ -424,9 +434,17 @@ impl GpuTextures {
             page_count: 0,
             sampler,
             null_view,
+            bind_group_creates: 0,
         }
     }
 
+    /// Reads and zeroes the standalone-image bind-group creation counter --
+    /// call once per frame before `update()`, mirroring
+    /// `UiRenderState::take_counters`.
+    pub fn take_bind_group_creates(&mut self) -> u64 {
+        std::mem::take(&mut self.bind_group_creates)
+    }
+
     pub fn array_view(&self) -> &TextureView {
         &self.array_view
     }

iris/examples/bench_images.rs

@@ -0,0 +1,101 @@
+//! (d) of IRIS_TODO.md's "Benchmarks" item: 1,000 image rows, checking that
+//! standalone-image bind-group *creation* -- a real `wgpu` resource, unlike
+//! the counters in `benches/message_list.rs` -- goes to zero once every
+//! image has loaded. This needs an actual `wgpu` device (`GpuTextures`,
+//! `UiRenderNode`), so unlike the rest of the suite it cannot run as a
+//! plain binary; run it through `iris/run-headless.sh bench_images`, which
+//! gives it a real (headless, GPU-accelerated) compositor and surface. See
+//! `run-bench.sh` for the wrapper that greps its output into one line.
+//!
+//! Each `RedrawRequested` prints the frame number and
+//! `UiRenderNode::take_image_bind_group_creates()` for that frame, then
+//! requests another redraw (nothing else marks the scene dirty, so without
+//! this the app would only ever draw once). The first frame is expected to
+//! report 1,000 (one create per image, on first load); the steady state
+//! IRIS_TODO.md asks this scenario to prove is every frame after settling
+//! down to 0.
+//!
+//! After `SETTLE_FRAMES` it appends one *new* image row (a transcript
+//! receiving one more message) and keeps counting -- a chat transcript's
+//! real access pattern is "one more image arrives," not "reload the whole
+//! list," so the steady-state question that actually matters is the
+//! *incremental* cost of that one append, not just whether an untouched
+//! scene costs zero. It exits after `FRAMES`.
+
+use iris::prelude::*;
+
+const ROWS: usize = 1000;
+const SETTLE_FRAMES: usize = 4;
+const FRAMES: usize = 6;
+
+#[derive(DefaultUiState)]
+struct State {
+    ui_state: DefaultUiState,
+    span: WeakWidget<Span>,
+    frame: usize,
+    appended: bool,
+}
+
+impl DefaultAppState for State {
+    fn new(
+        mut ui_state: DefaultUiState,
+        rsc: &mut DefaultRsc<Self>,
+        _: Proxy<Self::Event>,
+    ) -> Self {
+        let mut span = Span::empty(Dir::DOWN);
+        for _ in 0..ROWS {
+            let img = image::DynamicImage::new_rgba8(32, 32);
+            let widget = image::<DefaultRsc<Self>>(img)(rsc);
+            let widget = rsc.ui.widgets.add_strong(widget);
+            span.push(widget.any());
+        }
+        let span = rsc.ui.widgets.add_strong(span);
+        let span_weak = span.weak();
+        let root = rsc.ui.widgets.add_strong(Scroll::new(span.any(), Axis::Y));
+        ui_state.set_root(root.any());
+        Self {
+            ui_state,
+            span: span_weak,
+            frame: 0,
+            appended: false,
+        }
+    }
+
+    fn window_event(
+        &mut self,
+        event: winit::event::WindowEvent,
+        rsc: &mut DefaultRsc<Self>,
+        _render: &mut UiRenderState,
+    ) {
+        if !matches!(event, winit::event::WindowEvent::RedrawRequested) {
+            return;
+        }
+        self.frame += 1;
+        let creates = self.ui_state.renderer.ui.take_image_bind_group_creates();
+        println!(
+            "BENCH_IMAGES frame={} bind_group_creates={creates}",
+            self.frame
+        );
+        if self.frame == SETTLE_FRAMES && !self.appended {
+            self.appended = true;
+            let img = image::DynamicImage::new_rgba8(32, 32);
+            let widget = image::<DefaultRsc<Self>>(img)(rsc);
+            let widget = rsc.ui.widgets.add_strong(widget);
+            rsc.ui
+                .widgets
+                .get_mut(&self.span)
+                .unwrap()
+                .push(widget.any());
+            println!("BENCH_IMAGES appended one image after settling");
+        }
+        if self.frame < FRAMES {
+            self.ui_state.window.request_redraw();
+        } else {
+            std::process::exit(0);
+        }
+    }
+}
+
+fn main() {
+    DefaultApp::<State>::run();
+}

iris/run-bench.sh

@@ -0,0 +1,24 @@
+#!/bin/sh
+# Runs iris's on-demand benchmark suite (IRIS_TODO.md's "Benchmarks" item).
+# Never run by `cargo test`; run this by hand or before/after a layout
+# change. Always release -- see AGENTS.md's own rule against reading a
+# frame time from a debug build.
+#
+#   ./run-bench.sh            # everything
+#   ./run-bench.sh list       # just the CPU-only message-list scenarios
+#   ./run-bench.sh images     # just the GPU bind-group-creation scenario
+set -eu
+here=$(cd "$(dirname "$0")" && pwd)
+cd "$here"
+
+what="${1:-all}"
+
+if [ "$what" = "all" ] || [ "$what" = "list" ]; then
+    echo "=== message_list (CPU-only, no window) ==="
+    cargo bench --bench message_list
+fi
+
+if [ "$what" = "all" ] || [ "$what" = "images" ]; then
+    echo "=== bench_images (real wgpu device, via run-headless.sh) ==="
+    timeout 60 ./run-headless.sh bench_images --seconds 4 2>&1 | grep "^BENCH_IMAGES"
+fi