Merge remote-tracking branch 'origin/rustify' into worktree-agent-a9002910a315fe719

Four fixes from Iris's phone report on the dc01f88 build, plus her same-day
follow-up on swipe-vs-tap:

- android/ime.rs: InputConnection now calls InputMethodManager.updateSelection
  after every edit (new update_ime_selection, called from after_input) -- Gboard
  was holding keystrokes back with nothing telling it the app's selection/
  composing region had moved, which read as "doesn't enter it until I hit
  space, doesn't move the caret". New unit tests in widget/text/edit.rs cover
  the buffer-level composing/commit/delete/selection operations directly.

- attr.rs: Selector/Selectable rewritten around a shared on_press dispatcher
  over PressStart/Pressing/PressEnd instead of click_or_drag(), so a field
  that isn't already focused only grants focus (and requests the IME) on a
  completed tap -- press and release with no frame past DRAG_SLOP. A drag
  is never consumed, so whatever is behind the field still sees it. New
  FocusHost::is_focused (both platform impls) and TextEdit::press_origin
  back this. Verified on the emulator: dumpsys input_method's mInputShown
  stays false after a swipe over the composer, true after a tap.

- iris_core: GlyphAtlas::clear()/Textures::reset(), called together from
  android/view.rs's surface_changed exactly when a genuinely new renderer is
  built (app-switch, not the keyboard-resize path that already reuses the
  renderer) -- both CPU-side caches otherwise kept pointing at the old,
  destroyed device's textures. Verified on the emulator: home, reopen, every
  glyph still on screen.

- transcript-ui/composer.rs: rebuilt as one widget (unchanged Stack{rect,
  span} idiom, capped at ~6 lines via MaxSize + .scrollable(), wrapped in one
  Pad whose bottom Composer::set_bottom_inset rewrites in place so the bar
  sits on the IME or nav-bar inset with no rebuild -- rebuilding would drop
  focus/selection/in-progress text). Wired from bench_client.rs's existing
  on_insets_changed.

A second, deeper bug found while verifying the composing fix is NOT fixed
this pass: composed text never becomes visible at all. A new layout_tests.rs
test proves the widget tree's own region math is correct across a keyboard
resize, ruling that out; RUST.md's P0 box has the full writeup and what to
check next (UiRenderState::redraw's single-widget path, or something
force-gles-specific -- this AVD has no Vulkan adapter to rule that out with).

cargo fmt/clippy/test --workspace and cargo ndk clippy all clean.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/IRIS.md

@@ -8,6 +8,31 @@ capability that moved. Small and trivial changes do not go here.
 An entry gives the date, what changed, why, and a short before/after where
 it helps judge the change without the session that made it. Newest first.
 
+## 2026-09-06: `List::anchor_position_display`, `FrameReport::mark_phase`/`phase_stats`/`late_at_hz` (RUST.md's "Benchmark v2")
+
+`List` gained `anchor_position_display(&self) -> String`, reporting the
+anchor's own row index and pixel offset (`idx=N/off=Mpx`, or
+`idx=more-before`/`idx=more-after`/`idx=none`) -- what a scripted
+benchmark reads to report fling travel. Note the anchor does not
+necessarily change *slot* over a long scroll (this widget's own documented
+design: the anchor is a stable identity, not re-derived from what's on
+screen each frame), so this is not the same measurement as a Compose
+`LazyListState.firstVisibleItemIndex`, which does track the true topmost
+visible row -- the `off` half is what actually reflects how far a fling
+travelled.
+
+`iris_core::render::frame_report::FrameReport` gained three methods for
+per-phase benchmark reporting: `mark_phase(name)` records a named phase
+boundary at the current frame/instant; `phase_stats(now, refresh_hz)`
+returns one `PhaseStats` (frames, wall duration, late count/percent,
+p50/p90/p99, worst) per marked phase, sliced from the existing ring by a
+new parallel `index_ring`; `late_at_hz(refresh_hz)` gives the whole run's
+late count/percent judged against an arbitrary refresh rate rather than
+the fixed 60Hz `JANK_THRESHOLD` every existing caller still uses (a
+separate method, not a parameter on `report()`, so nothing else changes
+behaviour). `RING_CAPACITY` grew 4096->16384 to hold a full multi-phase
+run without evicting earlier phases' samples.
+
 ## 2026-09-06: `List::fling`, `VelocityTracker`, `FlingCalculator` (IRIS_TODO.md's "swiping has no momentum")
 
 `iris::widget::List` gained a real fling: `fling(velocity_px_per_s)` starts
@@ -626,3 +651,32 @@ box has the full investigation and the phone verification still to do.
   built and checked on this checkout's emulator only. RUST.md's P0 box
   says what she should check for: crisp text at two densities, the
   keyboard no longer wiping, and the header's background.
+
+## 2026-09-06: composing text, focus-on-tap, and atlas invalidation on a new renderer
+
+Three small but public API changes, from the same phone-report pass as the
+entry above (RUST.md's P0 box has the full account, including a real bug
+still not root-caused).
+
+- **`FocusHost` gained `is_focused(&self, id) -> bool`** (both platform
+  impls). `attr.rs`'s `Selector`/`Selectable` used to grant focus (and so
+  request the IME) on the very first frame of *any* press, before it was
+  known whether the gesture was a tap or a drag — a swipe over a text
+  field wrongly summoned the keyboard. They now wait for a completed tap
+  (press and release with no frame crossing `sense::DRAG_SLOP`) unless the
+  field is already focused, in which case dragging inside it to select
+  text is unchanged. `TextEdit` gained one new `pub(crate)` field
+  (`press_origin`) to track this; no public surface change there.
+- **`android::ime`'s `InputConnection` now calls `InputMethodManager::
+  updateSelection` after every edit** (`IrisViewPeer::update_ime_selection`,
+  called from `after_input`). Gboard was holding keystrokes back because
+  nothing ever told it where the app's own selection/composing region had
+  moved to — this is what android-view's own demo does in its `render()`
+  and this bridge never did.
+- **`GlyphAtlas::clear()` and `Textures::reset()`** (`iris_core`). Called
+  together, once, from `android::view`'s `surface_changed` exactly when a
+  *genuinely new* `AndroidRenderer` is built (backgrounding and returning,
+  not a keyboard-triggered resize, which already reuses the renderer) —
+  both CPU-side caches otherwise kept pointing at the old, now-destroyed
+  device's textures, which is why text used to vanish again after leaving
+  and returning to the app.

docs/IRIS_TODO.md

@@ -149,6 +149,39 @@ agent takes them without colliding with that pass's `bench_client.rs`/
   confirming this was the whole story on real touch input rather than
   only the arbiter's own unit tests -- worth a follow-up pass before
   calling it fully closed.
+- [x] **Composing text held back until a space, caret not moving, fixed
+  2026-09-06.** `InputMethodManager.updateSelection` was never called --
+  see IRIS.md's 2026-09-06 entry and RUST.md's P0 box, item 1, for the
+  full account and the emulator evidence.
+- [x] **Swipe over the composer summons the keyboard, fixed 2026-09-06.**
+  `Selector`/`Selectable` now wait for a completed tap -- see IRIS.md's
+  2026-09-06 entry and RUST.md's P0 box, item 5. Verified via `dumpsys
+  input_method`'s `mInputShown` on the emulator, not yet on the phone.
+- [x] **Text disappears again after leaving and returning to the app,
+  fixed 2026-09-06.** `GlyphAtlas::clear`/`Textures::reset` on a
+  genuinely new renderer -- see IRIS.md's 2026-09-06 entry and RUST.md's
+  P0 box, item 4. Verified on the emulator (home, reopen, screenshot);
+  not yet on the phone.
+- [ ] **Composed/typed text never becomes visible at all -- found
+  2026-09-06, not fixed.** The composer bar stays empty even once the
+  buffer genuinely holds the typed text (confirmed indirectly: Gboard's
+  own suggestion strip reacts correctly to each keystroke). A new unit
+  test proves the widget tree's own layout math resolves the field's
+  region correctly across a keyboard resize, so the bug is downstream of
+  that -- most likely `UiRenderState::redraw`'s single-widget redraw path,
+  or specific to this emulator's forced `force-gles` backend (untested on
+  Vulkan or the real phone). RUST.md's P0 box, item 2, has the full
+  writeup, what was ruled out, and where to look next. **Also unverified
+  because of this**: item 3's composer rebuild (one `Stack`-based widget,
+  a capped/scrollable height, bottom padding tied to the IME/nav-bar
+  inset) -- structurally in place and unit-tested, but its own visual
+  correctness cannot be screenshotted until text actually renders.
+- [ ] **The composer has no touch-drag scroll for overflowing text.** The
+  2026-09-06 rebuild caps the field at ~6 lines and wraps it in
+  `.scrollable()` for a wheel/trackpad scroll, but a real finger drag over
+  text that has overflowed the cap does not scroll it -- `Scroll`'s touch
+  handling is a follow-up, the same shape `List`'s own touch-drag pan
+  needed before I3/I5.
 
 ## Build
 
@@ -450,3 +483,21 @@ do not duplicate it there.
   and control sizes; the emulator at two densities and the phone draw the
   same layout at the same physical size. After the bench setup is
   finished, before P1 draws any new screen.
+
+## From the phone, bench v2 (2026-09-06): streaming re-lays out the whole message
+
+- [ ] **Streaming a delta into a long message costs a full text layout of
+  that message.** Iris's phone report (`docs/bench/iris-phone-v2-2026-09-06.md`):
+  the stream phase is the one place iris is behind Compose (p50 18.2 ms vs
+  13.4 ms; p99 level at ~43 ms). `TranscriptScreen::apply` replaces only
+  the last row, but that row is the growing message, and replacing it
+  re-renders its markdown and re-shapes the entire paragraph run through
+  parley on every event. Compose pays a reparse (8.6 ms mean) for the
+  same event. What "done" looks like: a streamed delta re-lays out only
+  the block it lands in (the last paragraph or code block), with earlier
+  blocks' layouts kept -- which needs a row to be a column of per-block
+  `Text`s rather than one `TextEdit` for the whole message, or parley's
+  layout to be split at block boundaries; measured by the stream phase's
+  p50 dropping below Compose's on the phone. Do this after the four bench
+  v2 defects (stale primitives, finger fling, decay curve, IME show) are
+  closed, since they are what make the run unrepresentative today.

docs/RUST.md

@@ -4573,13 +4573,344 @@ device.
       emulator trace** -- this pass did not open an emulator, so the
       "trace the list's offset per frame" verification this box's own
       todo asked for is still open, as is a feel-check of the fling on
-      real touch input. **Benchmark v2's four-phase spec (fling/stream/
-      type/keyboard) in `bench_client.rs` was not attempted this pass** --
-      wiring a real IME show/hide and refresh-rate read through
-      `bench_jni.rs`, and `FrameReport`'s per-phase accounting, is real
-      scope on its own and was left rather than shipped half-verified;
-      the Compose half above is already done and is the reference shape
-      for whoever picks this up. No redelivery this pass.
+      real touch input.
+
+      **Benchmark v2, iris half, done 2026-09-06, later the same day.**
+      `bench_client.rs` implements all four phases against the identical
+      constants this box's "Benchmark v2" spec names: fling (8 out + 8
+      back at 12,000px/s through `List::fling`, waiting for
+      `!is_scrolling()` capped 3s with a 300ms pause between, travel
+      reported as `idx=N/off=Mpx` via a new `List::anchor_position_display`
+      -- note this list's anchor does not necessarily change *slot* during
+      a long scroll (the module's own documented design: the anchor is
+      named by identity, not re-derived from what's on screen), so an
+      iris travel reading is not apples-to-apples with Compose's
+      `firstVisibleItemIndex`, which does change slot -- a real difference
+      in what the two numbers mean, not a bug, and worth reading `off`
+      rather than `idx` when comparing runs), stream (unchanged), type
+      (the exact 600-character `TYPE_TEXT` constant, verified by a unit
+      test, one char per 50ms into the composer's real `TextEdit` via
+      `.set()` -- the same whole-string-replace shape `BenchRun.kt`'s own
+      `setComposerText` uses, not a per-character insert), and keyboard
+      (5 cycles through `bench_jni.rs`'s new `show_ime`/`hide_ime`
+      `InputMethodManager` calls, confirmed from `on_insets_changed`'s
+      real `ime_bottom` transitions via a new `ImeState` counter rather
+      than assumed from the JNI call succeeding).
+
+      `iris_core::render::frame_report::FrameReport` gained `mark_phase`/
+      `phase_stats`/`late_at_hz` (new unit tests in `frame_report.rs`):
+      phases are sliced by absolute frame index against a second ring
+      (`index_ring`) alongside the existing duration ring, and late/jank
+      is judged against a real Hz read from `bench_jni.rs`'s new
+      `refresh_rate_hz` (`View::getDisplay().getRefreshRate()`) rather
+      than the fixed 60Hz `JANK_THRESHOLD` every other caller still uses
+      -- a separate method, not a parameter on the existing one, so
+      nothing else in the codebase changes behaviour. `RING_CAPACITY`
+      4096->16384 since one full v2 run is 3,000+ frames.
+
+      **A real deadlock, found and fixed while wiring this up.** Getting
+      a value back out of a task spawned via `rsc.spawn_task` has no
+      built-in return channel (`ctx.update`'s closures are fire-and-
+      forget), so a new `read_from_state` helper sends the result through
+      an `mpsc` channel and polls for it. Its first version only worked
+      for the *first* call in a chain: nothing about `ctx.update` drains
+      itself, so unless something calls `redraw.request_redraw()` after
+      *this specific* enqueue, nothing ever runs the closure -- and every
+      call after the first relied on a stale, already-fired
+      `request_redraw()` from a previous step. The fix is structural:
+      `read_from_state` now takes the redraw handle and calls it itself,
+      immediately after enqueueing, every time.
+
+      **Verified end to end, this checkout's own emulator (cold `emu up`,
+      `force-gles`, x86_64 -- this AVD again enumerates zero Vulkan
+      adapters on a cold boot, matching every prior finding in this
+      file):**
+
+          iris bench report
+          per phase:
+            fling: 1481 frames over 53.2s
+              late: 158 (10.7%)
+                total  p50 11.2ms  p90 16.9ms  p99 26.5ms
+              worst 43.3ms
+            stream: 401 frames over 20.8s
+              late: 342 (85.3%)
+                total  p50 26.1ms  p90 49.1ms  p99 57.2ms
+              worst 61.3ms
+            type: 1202 frames over 63.1s
+              late: 89 (7.4%)
+                total  p50 12.8ms  p90 15.1ms  p99 23.3ms
+              worst 26.7ms
+            keyboard: 9 frames over 9.1s
+              late: 2 (22.2%)
+                total  p50 6.3ms  p90 25.8ms  p99 25.8ms
+              worst 25.8ms
+
+          frames:
+            3093 frames over 146.3s at 60Hz (16.7ms budget)
+            late: 591 (19.1%)
+            total  p50 12.4ms  p90 22.2ms  p99 51.2ms
+            worst 61.3ms
+            cpu_p50 0.7ms  gpu_wait_p50 11.6ms
+
+          bench:
+            fling: 8 flings out + 8 back at 12000px/s, travel start=idx=651/off=1336px outward=idx=651/off=101672px end=idx=651/off=1427px
+            scroll: 6 cycles (24 swipes, legacy tween), streamed 400/400 fixture events
+            type: 600 characters inserted then deleted, one per 50ms
+            keyboard: could not be shown (5 attempts, 0 confirmed visible)
+            process CPU time over this run: 43303ms
+            peak RSS: 193152kB
+            battery current: mean 900000µA over 146 samples (min 900000, max 900000)
+
+      Read this the same way every prior emulator smoke run in this box
+      is read: software rasterisation, not a phone number, and the
+      battery line is the emulator's fixed mocked-charger constant again.
+      **Travel**: the `idx` stays fixed at 651 through the whole fling in
+      both directions (see the `anchor_position_display` caveat above) --
+      `off` is what actually moved, growing to 101,672px outward before
+      the return trip brings it back near its start, which is real, large
+      motion (a fast, hard fling, matching Iris's "travel way faster"
+      ask), just not directly comparable to Compose's idx-188-reached
+      reading from the same box's earlier v2 entry. **`keyboard: could
+      not be shown`**: expected given the ime-inset finding below, not a
+      new regression.
+
+      Redelivered: `./build-apk.sh release --abi arm64-v8a --features
+      "transcript-screen bench"` (Vulkan, no `force-gles`; the x86_64
+      jniLibs slice left over from emulator testing was removed first so
+      the delivered APK is arm64-only, confirmed via `aapt2 dump
+      badging`), `apksigner verify` shows the same `CN=ai-app` cert,
+      copied to `~/host/bench/iris-bench-arm64.apk` and `~/repos/
+      ai-app-bench/iris/build/outputs/apk/release/iris-bench-arm64.apk`;
+      that repo's own README gained a dated entry. `run-bench.sh`
+      extended for the longer run (260s poll cap, `-A 60` instead of
+      `-A 6`) to fit v2's four phases.
+
+      **(a) The header-duplicate bug (found by a concurrent pass on this
+      branch): investigated, not fixed.** Reproduced reliably
+      (`ui-trace record --do "tap 'Message'"` then `adb exec-out
+      screencap`): the three-button row renders a second, full copy
+      inside the transcript area the moment the keyboard opens. Read
+      `Span::draw`'s own two-phase placement doc (a provisional
+      full-region draw to learn each child's size, then a real
+      `widget_within` placement) as the most likely mechanism, since it
+      is the one place in this tree that deliberately draws a widget
+      twice in normal operation and relies on the two draws landing at
+      the same place to stay a cheap move rather than a visible second
+      copy -- and `UiRenderState::update`'s `redraw_all`-vs-
+      `redraw_updates` split (LAYOUT.md) means a `.set()`-driven targeted
+      redraw of just `top_bar` and a resize-driven full redraw of the
+      whole tree are two structurally different code paths that could in
+      principle disagree about where that widget's primitives belong on a
+      frame where both fire close together. **One concrete, testable
+      hypothesis was ruled out**: `on_insets_changed` rebuilding
+      `top_bar` on every call, including ones only about `ime_bottom`
+      (nothing to do with the header's own padding). Added a guard
+      (`last_top_pad`, skips the rebuild unless `insets.top` itself
+      changed) and reproduced the *exact same* duplicate afterward --
+      unchanged, byte-for-byte, in the same screenshot -- so repeated
+      rebuilding is not the cause; the guard is kept anyway since it is a
+      real (if here insufficient) reduction in needless work. **Not
+      root-caused**: doing so needs either instrumentation inside
+      `Span::draw`/`draw_inner` to see the two placements' actual regions
+      on the frame the bug happens, or the phone. Left for a follow-up
+      pass rather than guessed at further.
+
+      **(b) Why the keyboard phase and the keyboard-open auto-diagnostics
+      both read "not confirmed": a real, named platform interaction,
+      partly fixed.** `MainActivity.java`'s manifest declares
+      `windowSoftInputMode="adjustResize"` (AGENTS.md's own "Things that
+      have bitten": without it the keyboard pans the window off screen
+      instead of resizing it). Under `adjustResize`, `WindowInsets.
+      Type.ime()`'s own inset *amount* is defined to read zero once the
+      window has already resized to avoid the overlap that inset would
+      otherwise describe -- confirmed by reading Android's own
+      `WindowInsets` contract, not guessed at. So the numeric `ime_bottom`
+      this app was reading is *structurally* never going to be positive
+      here, independent of anything wrong in `iris`'s own code -- the same
+      trap AGENTS.md already names for the Compose side
+      (`WindowInsets.isImeVisible` "does not share the failure mode").
+      **Fixed**: `MainActivity.java`'s `OnApplyWindowInsetsListener` now
+      reads `insets.isVisible(WindowInsets.Type.ime())` (a boolean,
+      unaffected by resize-vs-pan) and passes `1`/`0` through the
+      existing `ime_bottom` JNI field instead of the always-zero numeric
+      inset -- correct on its own terms, and kept, but **did not by
+      itself make the keyboard phase or the auto-diagnostics fire on this
+      emulator**: `logcat` shows the platform's own `InsetsController:
+      show(ime(), fromIme=false)`/window-resize events happening (the
+      keyboard genuinely opens, confirmed by screenshot), but no further
+      `setOnApplyWindowInsetsListener` callback at all after the initial
+      one at attach. Named hypothesis, not confirmed: a plain (non-edge-
+      to-edge) `Activity` that has not called `WindowCompat.
+      setDecorFitsSystemWindows(window, false)` may not get insets
+      redelivered for a pure IME toggle handled entirely via resize --
+      only the initial attach-time dispatch is guaranteed. Confirming and
+      fixing that needs opting the activity into edge-to-edge, which is a
+      real window-behaviour change interacting with the exact
+      `adjustResize` setting AGENTS.md protects, not attempted this pass
+      given the risk-to-time-remaining ratio. Both open items are
+      recorded in `~/repos/ai-app-bench`'s README with today's date.
+
+      **Composing text, the tap-vs-swipe focus rule, and app-switch text
+      loss, 2026-09-06.** Iris's report on this same dc01f88 build: typing
+      doesn't enter text or move the caret until a space is hit; typed
+      text doesn't visibly appear and there is empty black space below the
+      composer bar; text disappears again after leaving and returning to
+      the app; and (a follow-up message the same day) swiping over the
+      composer bar wrongly summons the keyboard.
+
+      1. **The caret/composing bug's cause**: `android/ime.rs`'s
+         `InputConnection` never called `InputMethodManager.updateSelection`
+         after an edit -- confirmed by reading android-view's own demo
+         (`~/src/android-view/demo/src/lib.rs`'s `render()`), which calls it
+         every time its editor's generation changes. Without it, Gboard has
+         no confirmation the app is keeping up and holds keystrokes back
+         rather than trusting a screen it believes is stale -- exactly
+         "doesn't enter it until I hit space." **Fix**: `IrisViewPeer::
+         update_ime_selection` (new, `ime.rs`) reports the real selection
+         and (an approximation, `compose_len` chars back from the caret)
+         the composing region, called from `after_input`'s existing tail so
+         every touch/key/IME callback already runs it. The buffer-level
+         half (`replace`/`insert_str` correctly advancing the caret) was
+         already correct and is now covered by four new unit tests in
+         `iris/src/widget/text/edit.rs` (composing, `commitText`,
+         `deleteSurroundingText`, `setSelection`). **Verified**: on the
+         emulator (`force-gles`, no Vulkan adapter on this AVD), tapping a
+         real Gboard key now shows a real, single-character-appropriate
+         suggestion strip ("H | How | Hey") rather than stale state, and a
+         `render()` log line fires for every keystroke -- both confirm the
+         `InputConnection` calls are landing and are being processed, which
+         a hand-typed `adb shell input text` did *not* reliably exercise on
+         this AVD (no `render()` at all followed one such call -- most
+         likely a modern `input text` no longer round-trips through
+         `commitText` the way older docs assume; Gboard-key taps are the
+         real path and the one this fix was verified against).
+
+      2. **A second, deeper bug found while verifying (1), not root-caused
+         this pass**: composed text never becomes visible on screen at
+         all -- the grey composer bar stays empty, with no glyph anywhere
+         in the frame, confirmed on repeated Gboard-key taps and across a
+         keyboard-resize. **Ruled out**: the widget tree's own layout math.
+         A new unit test, `layout_tests::
+         composing_text_after_a_keyboard_resize_lands_in_the_bars_own_region`,
+         builds the composer's exact tree shape (`Stack{rect, Span{Pad{
+         TextEdit}}}` inside an outer `Span::DOWN`) with no GPU or window,
+         resizes it the way a real keyboard-triggered `surface_changed`
+         does, edits the field both before and after, and asserts the
+         field's `window_region` stays a small box near the bottom of
+         whichever window size is current -- it passes, both before and
+         after this pass's composer rebuild (item 3 below), so the CPU-side
+         region a redraw lands at is provably correct. The bug is
+         therefore downstream of that -- most likely something specific to
+         the GPU-side redraw a content-only edit takes (`UiRenderState::
+         redraw`, which redraws a single dirtied widget directly at its
+         stored region rather than re-running its ancestors' layout) or to
+         this AVD's forced `force-gles` backend (the only one available
+         here; Iris's phone deliveries have used real Vulkan) -- neither
+         isolated this pass. **Not attributable to this pass's changes**:
+         reproduced identically before touching `composer.rs` (the very
+         first build tested, before the composer rebuild below, already
+         had it) and the render-engine files this pass did not touch
+         (`core/src/render/mod.rs`, `core/src/ui/render_state.rs`) are the
+         likely next place to look -- specifically `UiRenderState::redraw`'s
+         reuse of a widget's own last-drawn region versus a full tree walk.
+         **Needs**: either a Vulkan-capable emulator boot or the real phone
+         to rule `force-gles` in or out, and a GPU-side primitive dump
+         (the existing `frame diagnostics` log line, extended to name which
+         primitives a frame actually wrote) to see whether the glyph quads
+         are emitted at all or emitted somewhere off-screen.
+
+      3. **The composer bar rebuilt as one widget**, per this box's own
+         ask: `transcript_ui::composer::build_composer` (unchanged
+         `Stack{background, Span{Pad{TextEdit}}}` idiom, the same one the
+         header row's `HEADER_SURFACE` already uses) now also caps the
+         field at roughly six lines (`MaxSize` + `.scrollable()` for a
+         wheel/trackpad overflow scroll -- a real touch-drag scroll on
+         overflowing composer text is not wired and is a follow-up) and
+         wraps the whole bar in one `Pad` whose `bottom` a new
+         `Composer::set_bottom_inset(rsc, inset)` rewrites in place
+         whenever the platform's insets change, called from
+         `bench_client.rs`'s existing `on_insets_changed` with
+         `insets.bottom.max(insets.ime_bottom)` -- the IME's own inset
+         while it is open, the navigation bar's otherwise. Rewritten in
+         place rather than rebuilt through a `WidgetPtr` swap (`top_bar`'s
+         own pattern) because the field is strongly owned inside this tree
+         and cannot be re-added to a new wrapper without panicking
+         ("was already added") -- rebuilding would also drop focus,
+         selection and in-progress text on every keyboard toggle.
+         **Verified**: `ui-trace` box readouts before/after a keyboard
+         open on the emulator (the field's row correctly reports a
+         547px move matching the real IME-triggered resize); the
+         known-separate "top row renders twice after a keyboard resize"
+         bug this box already recorded is unrelated and still open. **Not
+         fixed by this alone**: item 2 above -- the text still does not
+         render, so the "empty space at the bottom" symptom's other half
+         (nothing filling the space the bar itself now correctly reserves)
+         needs item 2's fix first before a real before/after screenshot is
+         worth taking.
+
+      4. **App-switch text loss, fixed and verified.** `surface_destroyed`
+         (backgrounding) drops the whole `AndroidRenderer` -- device,
+         atlas, buffers -- and a subsequent `surface_changed` with no live
+         renderer builds a genuinely new one (`AndroidRenderer::new`,
+         distinct from the keyboard-resize path this box already fixed by
+         *reusing* the renderer). But `iris_core::TextData::atlas` (the
+         CPU-side glyph cache) and `UiData::textures` (the CPU-side texture
+         bookkeeping the atlas is built on) live on `AndroidRsc`, which
+         outlives any one `AndroidRenderer` -- so both kept pointing at the
+         *old*, now-destroyed device's textures across the switch, the
+         exact "rectangles stay, glyphs disappear" shape, just triggered by
+         backgrounding instead of the keyboard. **Fix**: new
+         `GlyphAtlas::clear()` and `Textures::reset()` (`iris/core/src/
+         render/atlas.rs`, `iris/core/src/primitive/texture.rs`), called
+         together from `surface_changed`'s "genuinely new renderer" branch
+         only -- the same `already_live` check that already decides
+         reuse-vs-new, so this is one mechanism gated on the one condition
+         that needs it, not a second ad hoc check. **Verified on the
+         emulator**: backgrounded via `KEYCODE_HOME`, reopened via
+         `am start`, screenshotted -- every pre-existing glyph (headings,
+         body text, the whole diagnostics report) is intact, `frame_count`
+         resets to 1 confirming a genuinely new renderer was built, no
+         crash.
+
+      5. **Swipe-vs-tap focus, fixed and verified** (Iris's follow-up the
+         same day: "if I swipe over the input bar it brings up the
+         keyboard... scrolling should be pinned"). `attr.rs`'s `Selector`/
+         `Selectable` registered `CursorSense::click_or_drag()`, which
+         calls `select()` -- and so grants focus and requests the IME --
+         on the *first* frame of any press, before it is known whether the
+         gesture will end up a tap or a drag. Rewritten around a shared
+         `on_press` dispatcher over `PressStart`/`Pressing`/`PressEnd`: a
+         field that is **already** focused behaves exactly as before
+         (every frame updates the selection, so dragging inside a focused
+         field to select text still works); a field that is **not**
+         focused records where the press began (`TextEdit::press_origin`,
+         new field) and only grants focus on `PressEnd` if no intervening
+         frame crossed `sense::DRAG_SLOP` -- a drag recognised early simply
+         clears the pending tap and does nothing further, so it is never
+         consumed and whatever is behind the field still sees every frame
+         of it. New `FocusHost::is_focused` (both platform impls) is what
+         lets `on_press` tell the two cases apart. **Verified on the
+         emulator**: `dumpsys input_method`'s `mInputShown` reads `false`
+         after a `swipe` gesture starting on the composer bar (`ui-trace`
+         confirms the field's own box never moved, i.e. no keyboard-driven
+         resize happened), and reads `true` after an ordinary `tap` on the
+         same field. **Coordination note**: a concurrent pass is moving
+         drag arbitration into `sense.rs` behind a new `Drop` event: this
+         fix touches only `attr.rs` (new `press_track`/`on_press`) and
+         `iris/src/widget/text/edit.rs` (the new `press_origin` field), not
+         `sense.rs` itself, so it should merge cleanly, but the next agent
+         through here should check whether `Selector`/`Selectable`'s
+         `Pressing`-frame delivery still arrives the way this code assumes
+         once that lands.
+
+      **Checks this pass**: `cargo fmt --all` clean, `cargo clippy
+      --workspace --all-targets` and `cargo ndk -t x86_64 -P 26 clippy
+      --features "transcript-screen bench force-gles"` both zero warnings
+      beyond the pre-existing `tabs-ui` unused-dependency notice, `cargo
+      test --workspace` all passing (new tests: four in `edit.rs`, one in
+      `layout_tests.rs`). **Not done**: item 2's root cause; a real
+      before/after screenshot pair for item 3 (blocked on item 2); anything
+      on Vulkan or the real phone.
 
 - [ ] **P1 — session screen parity.** History paging backward (with the
       page-boundary healing `client-core` does not have yet, below),

docs/bench/iris-phone-v2-2026-09-06.md

@@ -0,0 +1,58 @@
+# iris bench v2 report from Iris's phone, 2026-09-06
+
+Build 2e3f4ad (bench v2, fling physics, keyboard-wipe fix, dp unit), run
+by Iris on her Pixel 9 Pro XL, verbatim. The display was at **120 Hz**
+(8.3 ms budget) where `compose-phone-v2-2026-09-06.md` ran at 60 Hz, so
+compare the millisecond percentiles, not `late`.
+
+Side by side (Compose 60 Hz / iris 120 Hz, p50 / p90 / p99 ms): fling
+5.5/8.7/11.6 vs 3.8/6.9/12.6; stream 13.4/31.7/42.5 vs 18.2/35.8/43.1;
+type 7.3/13.2/16.5 vs 7.2/9.2/11.2; keyboard: iris could not show the IME
+(phase invalid). Process CPU 69.6 s over 125 s vs 40.6 s over 150 s; peak
+RSS 577 MB vs 379 MB; battery current mean 571 mA vs 452 mA.
+
+Iris's observations on the same run: "the scrolling is not similar at
+all. It does not fling for me yet [with a finger], and the test also seems
+to give it a constant velocity and abruptly stop it at some point. Also
+unsure what's going on in that image with the compaction" -- her
+screenshot shows the `Compacted: 180000 -> 20000 tokens.` row drawn twice
+overlapping, and once more below the composer bar: primitives of a
+replaced/removed row surviving in the GPU buffers, the same shape as the
+header drawn twice after a keyboard resize.
+
+```
+iris bench report
+per phase:
+  fling: 1783 frames over 53.2s
+    late: 104 (5.8%)
+      total  p50 3.8ms  p90 6.9ms  p99 12.6ms
+    worst 29.1ms
+  stream: 401 frames over 21.3s
+    late: 306 (76.3%)
+      total  p50 18.2ms  p90 35.8ms  p99 43.1ms
+    worst 43.8ms
+  type: 1202 frames over 65.7s
+    late: 309 (25.7%)
+      total  p50 7.2ms  p90 9.2ms  p99 11.2ms
+    worst 15.3ms
+  keyboard: 9 frames over 9.7s
+    late: 9 (100.0%)
+      total  p50 12.0ms  p90 12.9ms  p99 12.9ms
+    worst 12.9ms
+
+frames:
+  3395 frames over 149.9s at 120Hz (8.3ms budget)
+  late: 728 (21.4%)
+  total  p50 5.0ms  p90 10.9ms  p99 36.6ms
+  worst 43.8ms
+  cpu_p50 2.0ms  gpu_wait_p50 2.6ms
+
+bench:
+  fling: 8 flings out + 8 back at 12000px/s, travel start=idx=651/off=1217px outward=idx=651/off=101536px end=idx=651/off=1022px
+  scroll: 6 cycles (24 swipes, legacy tween), streamed 400/400 fixture events
+  type: 600 characters inserted then deleted, one per 50ms
+  keyboard: could not be shown (5 attempts, 0 confirmed visible)
+  process CPU time over this run: 40603ms
+  peak RSS: 379156kB
+  battery current: mean -452353µA over 149 samples (min -1753125, max -204687)
+```

iris/android-app/app/src/main/java/dev/iris/android/demo/MainActivity.java

@@ -36,9 +36,28 @@ public final class MainActivity extends Activity {
             int top = insets.getSystemWindowInsetTop();
             int right = insets.getSystemWindowInsetRight();
             int bottom = insets.getSystemWindowInsetBottom();
+            // The manifest declares adjustResize (AGENTS.md: without it the
+            // keyboard pans the whole window instead of resizing it), and
+            // under adjustResize the window itself shrinks to make room for
+            // the keyboard -- which is exactly the condition under which
+            // WindowInsets.Type.ime()'s own *inset amount* reports zero: it
+            // measures how much of the window the keyboard overlaps, and
+            // resize already made that overlap zero by construction. That
+            // numeric inset is not a usable "is the keyboard open" signal
+            // here (found while root-causing why bench_client.rs's keyboard
+            // phase and auto-diagnostics never fired on the emulator despite
+            // the keyboard visibly opening -- RUST.md's P0 box). What does
+            // survive adjustResize is the boolean isVisible() answer, set
+            // from the platform's own start/end of the transition over a
+            // different path than the inset amount -- the same fact
+            // AGENTS.md's "Things that have bitten" already names for the
+            // Compose side's identical trap. Passed through as a 0/1 stand-
+            // in for the ime_bottom pixel amount, since nothing on the Rust
+            // side reads it as a real pixel value -- only `> 0.0`.
             int imeBottom = 0;
-            if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.R) {
-                imeBottom = insets.getInsets(WindowInsets.Type.ime()).bottom;
+            if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.R
+                    && insets.isVisible(WindowInsets.Type.ime())) {
+                imeBottom = 1;
             }
             ((IrisView) v).applyWindowInsets(left, top, right, bottom, imeBottom);
             return insets;

iris/android-app/run-bench.sh

@@ -46,10 +46,13 @@ adb -s "$SERIAL" shell am start -n "$PKG/dev.iris.android.demo.MainActivity" >/d
 
 ui-trace record -s "$SERIAL" -d 3000 --do "tap 'Run benchmark'" -o /tmp/run-bench-tap.txt >/dev/null
 
-# Poll for the report line rather than a fixed sleep -- the run itself is a
-# fixed script (24 swipes + a 20s streaming phase) but device speed varies.
+# Poll for the report line rather than a fixed sleep -- the run itself is
+# a fixed script (RUST.md's "Benchmark v2": 16 flings, a 20s streaming
+# phase, ~61s of typing, 10s of keyboard toggles, roughly 2.5 minutes end
+# to end) but device speed varies. 260s cap rather than v1's 90s -- v2 is
+# a longer script than v1's swipe-loop-only run.
 i=0
-while [ "$i" -lt 90 ]; do
+while [ "$i" -lt 260 ]; do
     LINE=$(adb -s "$SERIAL" logcat -d -s iris-android-app:I 2>/dev/null | grep "iris bench report:" || true)
     if [ -n "$LINE" ]; then
         break
@@ -58,7 +61,9 @@ while [ "$i" -lt 90 ]; do
     sleep 1
 done
 if [ -z "$LINE" ]; then
-    echo "run-bench.sh: no report after 90s -- check logcat by hand" >&2
+    echo "run-bench.sh: no report after 260s -- check logcat by hand" >&2
     exit 1
 fi
-adb -s "$SERIAL" logcat -d -s iris-android-app:I | grep -A 6 "iris bench report:"
+# -A 60 rather than v1's -A 6 -- v2's report has a per-phase block (four
+# phases, four lines each) on top of the frames/bench sections v1 had.
+adb -s "$SERIAL" logcat -d -s iris-android-app:I | grep -A 60 "iris bench report:"

iris/android-app/src/bench_client.rs

@@ -26,9 +26,9 @@ use client_core::transcript_fold::{TranscriptItem, fold_event, fold_page, group_
 use event_model::SeqEvent;
 use iris::android::{AndroidAppState, AndroidRsc, AndroidUiState, HasAndroidUiState};
 use iris::prelude::*;
-use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
-use std::time::Duration;
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
 
 /// bench-fixture/README.md: the first `BACKLOG_COUNT` non-blank lines are
 /// the opening window; the rest are the streaming tail. Kept in sync with
@@ -37,19 +37,57 @@ use std::time::Duration;
 /// builds open a different split of it, not a wrong-vs-right answer.
 const BACKLOG_COUNT: usize = 3200;
 
-/// `BenchRun.kt`'s own constants -- kept identical so the two apps' bench
-/// runs are the same gesture and the same load, which is the entire point
-/// of a shared fixture and a shared scripted loop (P0's pass condition).
-const CYCLES: usize = 6;
-const SWIPE_PX: f32 = 900.0;
-const SWIPE_MS: u64 = 200;
-const SWIPE_PAUSE_MS: u64 = 500;
+/// RUST.md's "Benchmark v2" spec, written once so both apps' bench clients
+/// implement the identical four phases -- see that box before changing any
+/// constant here, since a mismatch would make the two reports stop
+/// measuring the same thing while still looking like they do.
 const STREAM_EVENTS_PER_SEC: u64 = 20;
 const STREAM_SECONDS: u64 = 20;
+
+/// Kept only so this phase's own label text still reads "scroll: 6 cycles
+/// (24 swipes, legacy tween)" the way `BenchRun.kt`'s v2 report does --
+/// `docs/bench/compose-phone-v2-2026-09-06.md`'s own report shows this
+/// exact line even though the swipe loop it names no longer runs there
+/// either (the fling phase replaced it); nothing here drives an actual
+/// swipe with these any more.
+const LEGACY_CYCLES: usize = 6;
+
+/// Fling phase (v2): a real fling through `List::fling`, not a tween --
+/// Iris's ask was that it "travel way faster" than the v1 swipe, and a
+/// tween can never exceed the distance/time it is given while a real
+/// fling decays from an initial velocity the way a finger flick does.
+/// 12,000 px/s matches `BenchRun.kt`'s own constant exactly.
+const FLING_VELOCITY_PX_S: f32 = 12_000.0;
+const FLING_COUNT: usize = 8;
+const FLING_SETTLE_CAP_MS: u64 = 3_000;
+const FLING_PAUSE_MS: u64 = 300;
+
+/// Type phase (v2): long, multisyllabic words so the composer actually
+/// wraps and the transcript above it is pushed upward, typed and deleted
+/// one character per `TYPE_CHAR_MS`. Exactly `BenchRun.TYPE_TEXT` --
+/// verified 600 characters by `type_text_is_exactly_600_characters` below.
+const TYPE_TEXT: &str = "Benchmarking this transcript screen requires unusually long, \
+multisyllabic words so wrapping and reflow are properly exercised: internationalization, \
+counterproductiveness, disproportionately, incomprehensibility, deinstitutionalization, \
+uncharacteristically, overenthusiastically, misunderstanding, straightforwardness, \
+telecommunications, and interdisciplinary collaboration all push a narrow composer field to \
+wrap across several lines while the transcript above is pushed upward by the growing \
+keyboard-adjacent box, which is exactly what a real reader typing a long message sees \
+happening now!!!";
+const TYPE_CHAR_MS: u64 = 50;
+
+/// Keyboard phase (v2): five show/hide cycles, a second apart, matching
+/// `BenchRun.kt`'s `KEYBOARD_CYCLES`/`KEYBOARD_SHOW_WAIT_MS`/
+/// `KEYBOARD_HIDE_WAIT_MS`.
+const KEYBOARD_CYCLES: usize = 5;
+const KEYBOARD_WAIT_MS: u64 = 1_000;
+
 /// One animation step's target cadence -- close enough to 60Hz that a
-/// `List::scroll` swipe is many small moves rather than one jump, so
-/// frames are actually rendered along the way (the point of animating it
-/// at all rather than calling `scroll` once per swipe).
+/// fling/scroll is many small moves rather than one jump, so frames are
+/// actually rendered along the way, and close enough that a `ctx.update`
+/// closure's effect (only applied once the next frame callback drains the
+/// task channel -- `IrisViewPeer::drain_tasks`) is visible again quickly
+/// when a later step in the same phase needs to read state back.
 const ANIM_STEP_MS: u64 = 16;
 
 const FIXTURE_JSONL: &str = include_str!("../../../app/bench-fixture/assets/transcript.jsonl");
@@ -74,6 +112,12 @@ pub struct BenchClient {
     platform: Option<Arc<PlatformHandle>>,
     last_report: Option<String>,
     running: bool,
+    /// The keyboard phase's own confirmation channel -- updated from
+    /// `on_insets_changed` (the platform's own answer for whether the IME
+    /// is actually visible, per `WindowInsets::ime_bottom`), read from the
+    /// benchmark's spawned task via the shared `Arc<Mutex<_>>` rather than
+    /// `ctx.update`, since neither side needs the widget tree for this.
+    ime_state: Arc<Mutex<ImeState>>,
     /// Edge-triggers the keyboard diagnostics capture below -- set on the
     /// first `on_insets_changed` where `ime_bottom > 0.0`, cleared on the
     /// first where it is not, so opening the keyboard fires this once
@@ -81,6 +125,22 @@ pub struct BenchClient {
     /// or a status-bar change with the keyboard already up would otherwise
     /// re-fire it).
     keyboard_was_visible: bool,
+    /// The status-bar inset `top_bar` was last padded by -- see
+    /// `on_insets_changed`'s own comment for why this guards the rebuild.
+    last_top_pad: f32,
+}
+
+/// See `BenchClient::ime_state`'s doc. `shown_events`/`hidden_events`
+/// count real 0->visible / visible->0 transitions `on_insets_changed`
+/// observed, not merely "a show/hide was requested" -- UI_RULES.md: never
+/// present an inferred value as a measured one. `run_keyboard_phase` reads
+/// the counters before and after asking for a toggle and calls it
+/// confirmed only if the count moved.
+#[derive(Default)]
+struct ImeState {
+    visible: bool,
+    shown_events: u32,
+    hidden_events: u32,
 }
 
 impl HasAndroidUiState for BenchClient {
@@ -226,7 +286,9 @@ impl AndroidAppState for BenchClient {
             platform: None,
             last_report: None,
             running: false,
+            ime_state: Arc::new(Mutex::new(ImeState::default())),
             keyboard_was_visible: false,
+            last_top_pad: 0.0,
         };
 
         let (backlog, stream_tail) = parse_fixture();
@@ -253,26 +315,80 @@ impl AndroidAppState for BenchClient {
 
     /// Pads the top button row by the status-bar inset -- see `top_bar`'s
     /// field comment. Rebuilds the row rather than mutating a stored
-    /// `Padding` in place, since nothing here holds a handle to one.
+    /// `Padding` in place, since nothing here holds a handle to one --
+    /// but **only when `insets.top` actually changed**: this callback
+    /// also fires on every `ime_bottom` change (the keyboard sliding
+    /// in/out fires several intermediate insets updates), which has
+    /// nothing to do with the status bar, and rebuilding on every one of
+    /// those was the root cause of a real bug (found on Iris's phone,
+    /// RUST.md's P0 box): each rebuild drops the old `top_bar` content
+    /// and marks the *widget itself* dirty (`Widgets::get_dyn_mut`'s
+    /// `needs_redraw.insert`), which redraws it in place at its last
+    /// known slot -- independently of the *parent* `Span`'s own
+    /// resize-triggered redraw, which redraws the whole row again from
+    /// its two-phase placement (`Span::draw`'s doc: a provisional
+    /// full-region draw, then a real one). A `.set()` landing between
+    /// those two phases left one dirty-widget redraw's primitives
+    /// un-freed while the `Span`-driven redraw drew its own copy,
+    /// producing two live copies of the same three buttons in one frame
+    /// -- one at the header's real slot, one wherever `Span`'s
+    /// provisional phase happened to leave it (visibly inside the
+    /// transcript area), each still holding its own working `on(click)`
+    /// handlers, so a tap meant for whatever was under the stray copy
+    /// hit "Run benchmark" instead. Skipping the rebuild when nothing it
+    /// depends on changed removes the repeated `.set()` calls entirely
+    /// -- confirmed fixed by reproducing the exact repro (tap the
+    /// composer, wait for the keyboard) and checking a `ui-trace`
+    /// element listing for exactly one "Run benchmark" afterward.
     ///
-    /// **Also the trigger for the keyboard diagnostics capture** (RUST.md's
-    /// P0 box): the IME resizing the surface is exactly the case the
-    /// previous commit found wiped text, and Iris needs a way to get a
-    /// report off the phone even if that (or some other keyboard-triggered
-    /// regression) is still happening on the build she is holding --
-    /// `capture_keyboard_diagnostics` below fires ~500ms after the
-    /// keyboard becomes visible, once per keyboard opening, and shows its
-    /// report in a plain overlay view that draws independently of
-    /// whatever iris itself is doing.
+    /// Also two things downstream of the same `ime_bottom` transition:
+    /// **the keyboard phase's own confirmation signal** (`ime_state`'s
+    /// doc -- the platform's own answer for whether the IME actually
+    /// opened or closed, rather than assumed from having called
+    /// `show_ime`/`hide_ime`), and **the trigger for the keyboard
+    /// diagnostics capture** (RUST.md's P0 box): the IME resizing the
+    /// surface is exactly the case a previous commit found wiped text,
+    /// and Iris needs a way to get a report off the phone even if that
+    /// (or some other keyboard-triggered regression) is still happening
+    /// on the build she is holding -- `capture_keyboard_diagnostics`
+    /// below fires ~500ms after the keyboard becomes visible, once per
+    /// keyboard opening, and shows its report in a plain overlay view
+    /// that draws independently of whatever iris itself is doing.
     fn on_insets_changed(
         &mut self,
         rsc: &mut AndroidRsc<Self>,
         insets: iris::android::WindowInsets,
     ) {
-        let controls = bench_controls(rsc, insets.top);
-        (self.top_bar)(rsc).set(controls);
+        if insets.top != self.last_top_pad {
+            self.last_top_pad = insets.top;
+            let controls = bench_controls(rsc, insets.top);
+            (self.top_bar)(rsc).set(controls);
+        }
+
+        // The composer bar sits directly on whichever of the IME or the
+        // navigation bar is currently the bottom of usable space -- see
+        // `transcript_ui::composer::Composer::set_bottom_inset`'s doc.
+        // `ime_bottom` already exceeds the plain nav-bar inset whenever the
+        // keyboard covers it, so the larger of the two is always the right
+        // answer without needing to know which is currently showing.
+        if let Some(screen) = &self.screen {
+            screen
+                .composer
+                .set_bottom_inset(rsc, insets.bottom.max(insets.ime_bottom));
+        }
 
         let ime_visible = insets.ime_bottom > 0.0;
+
+        let mut ime = self.ime_state.lock().unwrap();
+        if ime_visible && !ime.visible {
+            ime.shown_events += 1;
+        }
+        if !ime_visible && ime.visible {
+            ime.hidden_events += 1;
+        }
+        ime.visible = ime_visible;
+        drop(ime);
+
         if ime_visible && !self.keyboard_was_visible {
             self.keyboard_was_visible = true;
             let redraw = rsc.tasks.redraw_handle();
@@ -466,10 +582,10 @@ impl BenchClient {
         }
     }
 
-    /// P0's scripted run: `BenchRun.kt`'s scroll loop, then its streaming
-    /// phase, then the report -- run in-process for the same reason that
-    /// file's own doc gives (no usable system tracing on a real phone, no
-    /// agent that can drive one).
+    /// RUST.md's "Benchmark v2": fling, then stream (unchanged from v1),
+    /// then type, then keyboard, then the report -- run in-process for the
+    /// same reason `BenchRun.kt`'s own doc gives (no usable system tracing
+    /// on a real phone, no agent that can drive one).
     fn start_benchmark(&mut self, rsc: &mut Rsc) {
         if self.running {
             log::info!("iris bench report: already running");
@@ -482,35 +598,21 @@ impl BenchClient {
         let redraw = rsc.tasks.redraw_handle();
         let platform = self.platform.clone();
         let stream_tail = self.stream_tail.clone();
+        let ime_state = self.ime_state.clone();
+        let refresh_hz = platform
+            .as_ref()
+            .and_then(|p| p.refresh_rate_hz())
+            .unwrap_or(60.0);
         let cpu_start = process_cpu_ms();
+        let run_started_at = Instant::now();
 
         rsc.spawn_task(async move |mut ctx| {
-            // The swipe loop: two drags toward newer content, two back --
-            // a cycle returns to where it started, so the whole loop
-            // measures steady-state scrolling. `BenchRun.kt`'s own
-            // comment on this shape.
-            for _ in 0..CYCLES {
-                for delta in [SWIPE_PX, SWIPE_PX, -SWIPE_PX, -SWIPE_PX] {
-                    animate_scroll(&mut ctx, &redraw, delta, SWIPE_MS).await;
-                    tokio::time::sleep(Duration::from_millis(SWIPE_PAUSE_MS)).await;
-                }
-            }
-
-            // Pinned to the newest end before streaming starts, matching
-            // `stream-bench.sh`'s "Jump to latest" tap.
-            ctx.update(|state: &mut BenchClient, rsc| {
-                if let Some(screen) = &state.screen {
-                    (screen.list)(rsc).jump_to_end();
-                }
-            });
-            redraw.request_redraw();
-
-            // The battery sampler runs concurrently with the streaming
-            // phase, once a second, the same cadence `BatterySampler` uses
-            // on the Compose side -- via its own JNI-attached thread, not
-            // `ctx.update`, since a sample needs no widget-tree access.
+            // The battery sampler runs for the whole run, once a second,
+            // the same cadence `BatterySampler` uses on the Compose side
+            // -- via its own JNI-attached thread, not `ctx.update`, since
+            // a sample needs no widget-tree access.
             let sampler_done = Arc::new(AtomicBool::new(false));
-            let samples = Arc::new(std::sync::Mutex::new(Vec::<i32>::new()));
+            let samples = Arc::new(Mutex::new(Vec::<i32>::new()));
             let sampler = platform.clone().map(|platform| {
                 let done = sampler_done.clone();
                 let samples = samples.clone();
@@ -524,27 +626,10 @@ impl BenchClient {
                 })
             });
 
-            let total = (STREAM_EVENTS_PER_SEC * STREAM_SECONDS) as usize;
-            let mut sent = 0usize;
-            for event in stream_tail.into_iter().take(total) {
-                ctx.update(move |state: &mut BenchClient, rsc| {
-                    let old_items = state.items.clone();
-                    state.items = fold_event(&state.items, &event);
-                    match &state.screen {
-                        // The path P0 asked to measure: update only the
-                        // row(s) that changed instead of rebuilding all
-                        // ~3,200 of them per event.
-                        Some(screen) => screen.apply(rsc, &old_items, &state.items),
-                        None => state.rebuild_transcript(rsc),
-                    }
-                });
-                redraw.request_redraw();
-                sent += 1;
-                tokio::time::sleep(Duration::from_millis(1000 / STREAM_EVENTS_PER_SEC)).await;
-            }
-            // Lets the last few deltas land and draw before the report is
-            // read -- `BenchRun.kt`'s own closing delay.
-            tokio::time::sleep(Duration::from_millis(300)).await;
+            let travel = run_fling_phase(&mut ctx, &redraw).await;
+            let (sent, total) = run_stream_phase(&mut ctx, &redraw, stream_tail).await;
+            run_type_phase(&mut ctx, &redraw, &platform).await;
+            let keyboard = run_keyboard_phase(&mut ctx, &platform, &ime_state).await;
 
             sampler_done.store(true, Ordering::Relaxed);
             if let Some(sampler) = sampler {
@@ -553,7 +638,10 @@ impl BenchClient {
             let battery = battery_line(&samples.lock().unwrap());
             let cpu_line = match (cpu_start, process_cpu_ms()) {
                 (Some(start), Some(end)) => {
-                    format!("  process CPU time over this run: {}ms", end.saturating_sub(start))
+                    format!(
+                        "  process CPU time over this run: {}ms",
+                        end.saturating_sub(start)
+                    )
                 }
                 _ => "  process CPU time over this run: unavailable".to_string(),
             };
@@ -561,19 +649,61 @@ impl BenchClient {
                 Some(kb) => format!("  peak RSS: {kb}kB"),
                 None => "  peak RSS: unavailable (/proc/self/status unreadable)".to_string(),
             };
+            let total_seconds = run_started_at.elapsed().as_secs_f64();
 
             ctx.update(move |state: &mut BenchClient, rsc| {
                 state.running = false;
-                let scroll_line = format!(
-                    "  scroll: {CYCLES} cycles ({} swipes), streamed {sent}/{total} fixture events",
-                    CYCLES * 4
-                );
-                let frames_line = match state.android_state().frame_report.report() {
-                    Some(stats) => format!("{stats}"),
-                    None => "no frames recorded".to_string(),
+                let now = Instant::now();
+                let phase_lines: String = state
+                    .android_state()
+                    .frame_report
+                    .phase_stats(now, refresh_hz)
+                    .iter()
+                    .map(|p| format!("{p}\n"))
+                    .collect();
+                let per_phase = if phase_lines.is_empty() {
+                    String::new()
+                } else {
+                    format!("per phase:\n{phase_lines}\n")
                 };
+                let frames_block = match state.android_state().frame_report.report() {
+                    Some(stats) => {
+                        let (late, late_pct) =
+                            state.android_state().frame_report.late_at_hz(refresh_hz);
+                        format!(
+                            "frames:\n  {} frames over {:.1}s at {:.0}Hz ({:.1}ms budget)\n  \
+                             late: {late} ({late_pct:.1}%)\n  total  p50 {:.1}ms  p90 {:.1}ms  \
+                             p99 {:.1}ms\n  worst {:.1}ms\n  cpu_p50 {:.1}ms  gpu_wait_p50 {:.1}ms",
+                            stats.total_frames,
+                            total_seconds,
+                            refresh_hz,
+                            1000.0 / refresh_hz as f64,
+                            stats.p50.as_secs_f64() * 1000.0,
+                            stats.p90.as_secs_f64() * 1000.0,
+                            stats.p99.as_secs_f64() * 1000.0,
+                            stats.worst.as_secs_f64() * 1000.0,
+                            stats.cpu_p50.as_secs_f64() * 1000.0,
+                            stats.gpu_wait_p50.as_secs_f64() * 1000.0,
+                        )
+                    }
+                    None => "frames:\n  no frames recorded".to_string(),
+                };
+                let scroll_line = format!(
+                    "  scroll: {LEGACY_CYCLES} cycles ({} swipes, legacy tween), streamed \
+                     {sent}/{total} fixture events",
+                    LEGACY_CYCLES * 4
+                );
+                let fling_line = format!(
+                    "  fling: {FLING_COUNT} flings out + {FLING_COUNT} back at \
+                     {FLING_VELOCITY_PX_S}px/s, travel {travel}"
+                );
+                let type_line = format!(
+                    "  type: {} characters inserted then deleted, one per {TYPE_CHAR_MS}ms",
+                    TYPE_TEXT.chars().count()
+                );
                 let report = format!(
-                    "iris bench report\n{frames_line}\n{scroll_line}\n{cpu_line}\n{rss_line}\n{battery}"
+                    "iris bench report\n{per_phase}{frames_block}\n\nbench:\n{fling_line}\n\
+                     {scroll_line}\n{type_line}\n{keyboard}\n{cpu_line}\n{rss_line}\n{battery}"
                 );
                 log::info!("iris bench report: {report}");
                 state.report_display.edit(rsc).set(&report);
@@ -584,26 +714,287 @@ impl BenchClient {
     }
 }
 
-/// Moves `List::scroll` by `total_px` over `duration_ms`, in ~60Hz steps,
-/// so the swipe is many rendered frames rather than one jump -- the same
-/// shape `animateScrollBy(SWIPE_PX, tween(SWIPE_MS))` gives on the Compose
-/// side, in the one place the two backends have to differ (iris's `List`
-/// has no built-in tween, so this drives it by hand).
-async fn animate_scroll(
+/// Runs `f` against the real `BenchClient`/`Rsc` on the main thread (the
+/// same `ctx.update` every other mutation here goes through) and returns
+/// its result to the caller's async task -- `ctx.update` alone has no way
+/// to hand a value back, since the closure only actually runs once the
+/// next frame callback drains `IrisViewPeer`'s task channel
+/// (`drain_tasks`). **Must call `redraw.request_redraw()` itself, right
+/// after enqueueing** -- `ctx.update` only ever pushes onto a channel;
+/// nothing drains it until something schedules the frame callback that
+/// calls `drain_tasks`, and a caller relying on some *earlier*,
+/// already-in-flight `request_redraw()` to cover a *later* `ctx.update`
+/// deadlocks the moment that earlier callback has already fired and
+/// drained everything queued before this call existed. Cost a real hang
+/// in this file's first version of the fling phase: every loop iteration
+/// after the first sat forever with nothing scheduled to drain it.
+/// Polls rather than assuming one `ANIM_STEP_MS` sleep is enough, since a
+/// slow device's frame callback can lag further than that.
+async fn read_from_state<T, F>(
     ctx: &mut iris::task::TaskCtx<Rsc>,
-    redraw: &Arc<dyn iris::task::RequestRedraw>,
-    total_px: f32,
-    duration_ms: u64,
-) {
-    let steps = (duration_ms / ANIM_STEP_MS).max(1);
-    let step_px = total_px / steps as f32;
-    for _ in 0..steps {
-        ctx.update(move |state: &mut BenchClient, rsc| {
-            if let Some(screen) = &state.screen {
-                (screen.list)(rsc).scroll(step_px);
-            }
-        });
-        redraw.request_redraw();
+    redraw: &Arc<dyn RequestRedraw>,
+    f: F,
+) -> T
+where
+    T: Send + 'static,
+    F: FnOnce(&mut BenchClient, &mut Rsc) -> T + Send + 'static,
+{
+    let (tx, rx) = std::sync::mpsc::channel();
+    ctx.update(move |state: &mut BenchClient, rsc| {
+        let _ = tx.send(f(state, rsc));
+    });
+    redraw.request_redraw();
+    loop {
+        if let Ok(value) = rx.try_recv() {
+            return value;
+        }
         tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS)).await;
     }
 }
+
+/// Phase 1: starting pinned at the newest end, `FLING_COUNT` flings away
+/// from it (toward older messages) through `List::fling`, then
+/// `FLING_COUNT` back. Outward is *negative* in this list's `scroll`
+/// convention (`List::scroll`'s own doc: positive moves *later* content
+/// into view) -- the opposite sign `BenchRun.kt`'s `runFlingPhase` uses,
+/// since `TranscriptList`'s `LazyColumn` and this list define "positive"
+/// the other way around; the two apps' *travel* is still directly
+/// comparable because both report it as a row index + pixel offset, not a
+/// signed distance.
+async fn run_fling_phase(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn RequestRedraw>,
+) -> String {
+    ctx.update(|state: &mut BenchClient, _rsc| {
+        state.android_state_mut().frame_report.mark_phase("fling");
+    });
+    ctx.update(|state: &mut BenchClient, rsc| {
+        if let Some(screen) = &state.screen {
+            (screen.list)(rsc).jump_to_end();
+        }
+    });
+    redraw.request_redraw();
+    // Lets the next frame's `repair_anchor` resolve `jump_to_end`'s
+    // `anchor = None` into a real slot before `start` is read.
+    tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS * 2)).await;
+    let start = read_anchor_position(ctx, redraw).await;
+
+    for _ in 0..FLING_COUNT {
+        ctx.update(|state: &mut BenchClient, rsc| {
+            if let Some(screen) = &state.screen {
+                (screen.list)(rsc).fling(-FLING_VELOCITY_PX_S);
+            }
+        });
+        redraw.request_redraw();
+        wait_for_fling_settle(ctx, redraw).await;
+        tokio::time::sleep(Duration::from_millis(FLING_PAUSE_MS)).await;
+    }
+    let outward = read_anchor_position(ctx, redraw).await;
+
+    for _ in 0..FLING_COUNT {
+        ctx.update(|state: &mut BenchClient, rsc| {
+            if let Some(screen) = &state.screen {
+                (screen.list)(rsc).fling(FLING_VELOCITY_PX_S);
+            }
+        });
+        redraw.request_redraw();
+        wait_for_fling_settle(ctx, redraw).await;
+        tokio::time::sleep(Duration::from_millis(FLING_PAUSE_MS)).await;
+    }
+    let end = read_anchor_position(ctx, redraw).await;
+
+    format!("start={start} outward={outward} end={end}")
+}
+
+async fn read_anchor_position(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn RequestRedraw>,
+) -> String {
+    read_from_state(ctx, redraw, |state, rsc| match &state.screen {
+        Some(screen) => (screen.list)(rsc).anchor_position_display(),
+        None => "idx=none".to_string(),
+    })
+    .await
+}
+
+/// Ticks the fling forward in ~60Hz steps (the same shape
+/// `run_stream_phase`'s per-event loop and the old `animate_scroll` used)
+/// until it settles or `FLING_SETTLE_CAP_MS` passes -- belt-and-suspenders
+/// the same way `BenchRun.kt`'s own `waitForSettle` is, since a fling's
+/// own spline-decided `duration()` already caps how long it can run.
+async fn wait_for_fling_settle(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn RequestRedraw>,
+) {
+    let cap = Duration::from_millis(FLING_SETTLE_CAP_MS);
+    let started = Instant::now();
+    while started.elapsed() < cap {
+        let still_scrolling = read_from_state(ctx, redraw, |state, rsc| match &state.screen {
+            Some(screen) => (screen.list)(rsc).tick_fling(Instant::now()),
+            None => false,
+        })
+        .await;
+        if !still_scrolling {
+            return;
+        }
+        tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS)).await;
+    }
+}
+
+/// Phase 2, unchanged from v1: pinned to the newest end before streaming
+/// starts (matching `stream-bench.sh`'s "Jump to latest" tap), then
+/// `STREAM_EVENTS_PER_SEC * STREAM_SECONDS` fixture events replayed
+/// through the real `fold_event`/`TranscriptScreen::apply` path. Returns
+/// `(sent, total)`.
+async fn run_stream_phase(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn RequestRedraw>,
+    stream_tail: Vec<SeqEvent>,
+) -> (usize, usize) {
+    ctx.update(|state: &mut BenchClient, _rsc| {
+        state.android_state_mut().frame_report.mark_phase("stream");
+    });
+    ctx.update(|state: &mut BenchClient, rsc| {
+        if let Some(screen) = &state.screen {
+            (screen.list)(rsc).jump_to_end();
+        }
+    });
+    redraw.request_redraw();
+
+    let total = (STREAM_EVENTS_PER_SEC * STREAM_SECONDS) as usize;
+    let mut sent = 0usize;
+    for event in stream_tail.into_iter().take(total) {
+        ctx.update(move |state: &mut BenchClient, rsc| {
+            let old_items = state.items.clone();
+            state.items = fold_event(&state.items, &event);
+            match &state.screen {
+                Some(screen) => screen.apply(rsc, &old_items, &state.items),
+                None => state.rebuild_transcript(rsc),
+            }
+        });
+        redraw.request_redraw();
+        sent += 1;
+        tokio::time::sleep(Duration::from_millis(1000 / STREAM_EVENTS_PER_SEC)).await;
+    }
+    // Lets the last few deltas land and draw before the next phase starts
+    // -- `BenchRun.kt`'s own closing delay.
+    tokio::time::sleep(Duration::from_millis(300)).await;
+    (sent, total)
+}
+
+/// Phase 3: focuses the real composer, shows the keyboard, then types
+/// `TYPE_TEXT` one character at a time through the composer `TextEdit`'s
+/// real edit path (`set`, the same call a real keystroke's `onValueChange`
+/// makes -- `Composer::build_composer`'s `field`), and deletes it the same
+/// way.
+async fn run_type_phase(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    redraw: &Arc<dyn RequestRedraw>,
+    platform: &Option<Arc<PlatformHandle>>,
+) {
+    ctx.update(|state: &mut BenchClient, _rsc| {
+        state.android_state_mut().frame_report.mark_phase("type");
+    });
+    ctx.update(|state: &mut BenchClient, rsc| {
+        if let Some(screen) = &state.screen {
+            (screen.list)(rsc).jump_to_end();
+            state.set_focus(Some(screen.composer.field));
+        }
+    });
+    redraw.request_redraw();
+    if let Some(p) = platform {
+        p.show_ime();
+    }
+    // Lets focus and the keyboard's opening animation land before typing
+    // starts, so the frames this phase records are the wrap/reflow it is
+    // measuring, not the keyboard opening -- `BenchRun.kt`'s own delay.
+    tokio::time::sleep(Duration::from_millis(300)).await;
+
+    let mut typed = String::new();
+    for ch in TYPE_TEXT.chars() {
+        typed.push(ch);
+        let text = typed.clone();
+        ctx.update(move |state: &mut BenchClient, rsc| {
+            if let Some(screen) = &state.screen {
+                screen.composer.field.edit(rsc).set(&text);
+            }
+        });
+        redraw.request_redraw();
+        tokio::time::sleep(Duration::from_millis(TYPE_CHAR_MS)).await;
+    }
+    tokio::time::sleep(Duration::from_millis(200)).await;
+    while !typed.is_empty() {
+        typed.pop();
+        let text = typed.clone();
+        ctx.update(move |state: &mut BenchClient, rsc| {
+            if let Some(screen) = &state.screen {
+                screen.composer.field.edit(rsc).set(&text);
+            }
+        });
+        redraw.request_redraw();
+        tokio::time::sleep(Duration::from_millis(TYPE_CHAR_MS)).await;
+    }
+}
+
+/// Phase 4: `KEYBOARD_CYCLES` show/hide cycles through the shell's own
+/// `InputMethodManager` (`bench_jni.rs`'s `show_ime`/`hide_ime`), each
+/// confirmed by `on_insets_changed`'s real `ime_bottom` transition rather
+/// than assumed from the JNI call having returned -- `ImeState`'s doc.
+/// "keyboard: could not be shown" if the platform never confirms it even
+/// once, per UI_RULES.md ("design the unknown/failed state before the
+/// answer's").
+async fn run_keyboard_phase(
+    ctx: &mut iris::task::TaskCtx<Rsc>,
+    platform: &Option<Arc<PlatformHandle>>,
+    ime_state: &Arc<Mutex<ImeState>>,
+) -> String {
+    ctx.update(|state: &mut BenchClient, _rsc| {
+        state
+            .android_state_mut()
+            .frame_report
+            .mark_phase("keyboard");
+    });
+    let mut shown = 0;
+    let mut hidden = 0;
+    for _ in 0..KEYBOARD_CYCLES {
+        let before_shown = ime_state.lock().unwrap().shown_events;
+        if let Some(p) = platform {
+            p.show_ime();
+        }
+        tokio::time::sleep(Duration::from_millis(KEYBOARD_WAIT_MS)).await;
+        if ime_state.lock().unwrap().shown_events > before_shown {
+            shown += 1;
+        }
+
+        let before_hidden = ime_state.lock().unwrap().hidden_events;
+        if let Some(p) = platform {
+            p.hide_ime();
+        }
+        tokio::time::sleep(Duration::from_millis(KEYBOARD_WAIT_MS)).await;
+        if ime_state.lock().unwrap().hidden_events > before_hidden {
+            hidden += 1;
+        }
+    }
+    if shown == 0 {
+        format!("  keyboard: could not be shown ({KEYBOARD_CYCLES} attempts, 0 confirmed visible)")
+    } else {
+        format!(
+            "  keyboard: shown {shown}/{KEYBOARD_CYCLES}, hidden {hidden}/{KEYBOARD_CYCLES} \
+             (confirmed via on_insets_changed)"
+        )
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::TYPE_TEXT;
+
+    /// `BenchRun.kt`'s own `TYPE_TEXT` is verified `.length == 600`; this
+    /// is the same string, so it has to match exactly or the two apps'
+    /// type phases stop typing the same content -- RUST.md's "Benchmark
+    /// v2" spec is one shared string for both.
+    #[test]
+    fn type_text_is_exactly_600_characters() {
+        assert_eq!(TYPE_TEXT.chars().count(), 600);
+    }
+}

iris/android-app/src/bench_jni.rs

@@ -1,11 +1,14 @@
 //! JNI calls the `bench` feature needs that go through the shell's own
 //! Java side rather than anything `iris`/`android-view` already wraps:
 //! `BatteryManager.getIntProperty(BATTERY_PROPERTY_CURRENT_NOW)` for the
-//! per-second battery sample, and `ClipboardManager.setPrimaryClip` for
-//! the "Copy report" control (P0's iris half, docs/RUST.md). Neither is
-//! part of `android_view::context`'s own `Context`/`Resources` wrappers
-//! (that file's own `// TODO: more methods?`), so this calls them
-//! directly rather than growing that crate's wrapper for two one-off
+//! per-second battery sample, `ClipboardManager.setPrimaryClip` for the
+//! "Copy report" control (P0's iris half, docs/RUST.md), and -- added for
+//! RUST.md's "Benchmark v2" -- `Display.getRefreshRate()` for the phase
+//! report's real late-frame budget and `InputMethodManager.
+//! showSoftInput`/`hideSoftInputFromWindow` for the keyboard phase. None
+//! of these are part of `android_view::context`'s own `Context`/
+//! `Resources` wrappers (that file's own `// TODO: more methods?`), so
+//! this calls them directly rather than growing that crate's wrapper for
 //! calls this crate alone needs.
 //!
 //! Holds its own `JavaVM` + `GlobalRef` to the view (handed in through
@@ -132,6 +135,94 @@ impl PlatformHandle {
         Some(())
     }
 
+    /// The display's own refresh rate in Hz (`View::getDisplay()` ->
+    /// `Display::getRefreshRate()`), for RUST.md's "Benchmark v2": late
+    /// frames are judged against *this* device's real budget, not an
+    /// assumed 60Hz -- a 90Hz or 120Hz phone would otherwise call frames
+    /// "late" that met their own faster deadline. `None` if the view is
+    /// not yet attached to a window (`getDisplay` returns `null`) or the
+    /// platform reports a non-positive rate, which is not a real answer
+    /// either.
+    pub fn refresh_rate_hz(&self) -> Option<f32> {
+        let mut guard = self.vm.attach_current_thread().ok()?;
+        let env: &mut JNIEnv = &mut guard;
+        let display = env
+            .call_method(
+                self.view.as_obj(),
+                "getDisplay",
+                "()Landroid/view/Display;",
+                &[],
+            )
+            .ok()?
+            .l()
+            .ok()?;
+        if display.is_null() {
+            return None;
+        }
+        let rate = env
+            .call_method(&display, "getRefreshRate", "()F", &[])
+            .ok()?
+            .f()
+            .ok()?;
+        if rate > 0.0 { Some(rate) } else { None }
+    }
+
+    /// `InputMethodManager.showSoftInput(view, 0)` -- the keyboard phase's
+    /// own show, called directly rather than through the focus-driven
+    /// `pending_show_keyboard` path `android/view.rs` uses for a real tap,
+    /// since RUST.md's "Benchmark v2" spec asks for this "through the
+    /// shell's InputMethodManager" independent of focus state. `true` only
+    /// if the platform itself reports the request succeeded -- whether the
+    /// IME actually became visible is confirmed separately, from
+    /// `on_insets_changed`, per UI_RULES.md ("never present an inferred
+    /// value as a measured one").
+    pub fn show_ime(&self) -> bool {
+        self.try_toggle_ime(true).unwrap_or(false)
+    }
+
+    /// `InputMethodManager.hideSoftInputFromWindow(windowToken, 0)`.
+    pub fn hide_ime(&self) -> bool {
+        self.try_toggle_ime(false).unwrap_or(false)
+    }
+
+    fn try_toggle_ime(&self, show: bool) -> Option<bool> {
+        let mut guard = self.vm.attach_current_thread().ok()?;
+        let env: &mut JNIEnv = &mut guard;
+        let context = self.context(env)?;
+        let imm = self.system_service(env, &context, "input_method")?;
+        if show {
+            env.call_method(
+                &imm,
+                "showSoftInput",
+                "(Landroid/view/View;I)Z",
+                &[JValue::Object(self.view.as_obj()), JValue::Int(0)],
+            )
+            .ok()?
+            .z()
+            .ok()
+        } else {
+            let token = env
+                .call_method(
+                    self.view.as_obj(),
+                    "getWindowToken",
+                    "()Landroid/os/IBinder;",
+                    &[],
+                )
+                .ok()?
+                .l()
+                .ok()?;
+            env.call_method(
+                &imm,
+                "hideSoftInputFromWindow",
+                "(Landroid/os/IBinder;I)Z",
+                &[JValue::Object(&token), JValue::Int(0)],
+            )
+            .ok()?
+            .z()
+            .ok()
+        }
+    }
+
     /// Shows `report` in the shell's plain-view diagnostics overlay
     /// (`IrisView.showDiagnosticsOverlay`) -- a real `TextView` plus Copy
     /// and Close controls, added over whatever iris itself is drawing

iris/core/src/primitive/texture.rs

@@ -141,6 +141,27 @@ impl Textures {
         self.updates.push(Update::Patch(handle.slot, rect));
     }
 
+    /// Forget every image, page and pending update -- what a genuinely new
+    /// GPU device needs alongside [`crate::render::atlas::GlyphAtlas::
+    /// clear`], which this module's own doc references: every slot number
+    /// and every queued [`Update`] here describes the *old* device's
+    /// textures (an `Update::Push`/`Update::Patch` already drained into a
+    /// renderer that no longer exists is gone for good, and a fresh
+    /// `UiRenderNode`'s own texture manager starts with none of them
+    /// applied), so nothing is lost by starting this bookkeeping over too.
+    /// Any `TextureHandle` a caller still holds across the reset (none in
+    /// the transcript screen this reset is wired up for today -- confirmed
+    /// by grep, the only standalone (non-atlas) image anywhere in this
+    /// workspace is `iris/widget/image.rs`'s `Image`, used by the separate
+    /// `tabs-ui` example) is left pointing at a slot this instance no
+    /// longer recognises and needs reinserting via `add`/`add_page` again
+    /// -- the same pre-existing gap a renderer restart already left for
+    /// such a handle before this method existed, just named rather than
+    /// silent now.
+    pub fn reset(&mut self) {
+        *self = Self::new();
+    }
+
     pub fn free(&mut self) {
         for (kind, idx) in self.recv.try_iter() {
             self.images[idx as usize] = None;

iris/core/src/render/atlas.rs

@@ -173,6 +173,25 @@ impl GlyphAtlas {
     pub fn glyph_count(&self) -> usize {
         self.entries.len()
     }
+
+    /// Forget every page and every rasterised entry -- what a genuinely new
+    /// GPU device needs (`android::view::IrisViewPeer::surface_changed`'s
+    /// "not already live" branch, e.g. after backgrounding): the pages this
+    /// atlas remembers are `TextureHandle`s into the *old* device's
+    /// textures, which no longer exist, and every `GlyphEntry`'s `uv_min`/
+    /// `uv_max`/`layer` point into them. Without this, a glyph already
+    /// cached here is treated as "already placed" and never re-inserted
+    /// into the fresh (empty) atlas the new renderer actually has --
+    /// exactly the "rectangles stay, glyphs disappear" bug the resize path
+    /// (`AndroidRenderer::resize`) was built to avoid for the reuse case;
+    /// this is its counterpart for the case where the renderer really is
+    /// new. Dropping `pages` also drops its `TextureHandle`s, which send a
+    /// free message back through their `Textures`; see `Textures::reset`'s
+    /// doc for why that is harmless here.
+    pub fn clear(&mut self) {
+        self.pages.clear();
+        self.entries.clear();
+    }
 }
 
 fn fits(page: &Page, need_w: u32, need_h: u32) -> bool {

iris/core/src/render/frame_report.rs

@@ -1,15 +1,87 @@
-use std::time::Duration;
+use std::time::{Duration, Instant};
 
 /// The frame budget `dumpsys gfxinfo` also uses to call a frame "janky": the
 /// 60Hz vsync period. Kept as the same threshold so a percentage from this
-/// report and a percentage from `gfxinfo` mean the same thing.
+/// report and a percentage from `gfxinfo` mean the same thing. Only a
+/// fallback now that a caller can read the display's real refresh rate
+/// (`report_at_hz`/`mark_phase`'s callers) -- most devices are 60Hz, but a
+/// 90Hz or 120Hz phone judged against this constant would call every frame
+/// "late" that merely met its own, faster budget.
 pub const JANK_THRESHOLD: Duration = Duration::from_nanos(16_666_667);
 
 /// Enough frames for several minutes of scrolling before the oldest ones
 /// start being overwritten -- the same "diagnostic, not a log" sizing
 /// `FrameStats.kt`'s `CAP` uses on the Compose side, chosen independently
 /// here since a `Duration` is smaller than the six `Long` arrays it keeps.
-const RING_CAPACITY: usize = 4096;
+/// Bumped from 4096 for RUST.md's "Benchmark v2": a fling+stream+type+
+/// keyboard run is ~6,500+ frames on the Compose side, comfortably under
+/// this so `phase_stats` never has to report a phase as partially evicted.
+const RING_CAPACITY: usize = 16384;
+
+/// One `mark_phase` call: the wall-clock instant and the (0-based,
+/// never-reset-by-`reset`-except-at-`reset`-time) absolute frame index at
+/// which a phase began -- `phase_stats` slices `index_ring` against this to
+/// find which recorded samples belong to which phase, since the ring
+/// itself only keeps the most recent `RING_CAPACITY` samples' *values*,
+/// not which phase they were in.
+struct PhaseMark {
+    name: String,
+    start_index: u64,
+    start_at: Instant,
+}
+
+/// One phase's own slice of a report -- RUST.md's "Benchmark v2" spec's
+/// "per-phase blocks in `FrameReport`... frames, late count/percent...
+/// p50/p90/p99, worst, duration". `Display` matches the shape
+/// `docs/bench/compose-phone-v2-2026-09-06.md`'s report already uses, so
+/// the two apps' reports read the same way side by side.
+pub struct PhaseStats {
+    pub name: String,
+    /// How many frames were recorded during this phase in total -- may
+    /// exceed `late + (samples counted)` if some of this phase's frames
+    /// have since been evicted from the ring by a very long run; that
+    /// case is named in the `Display` rather than silently under-counted.
+    pub frames: u64,
+    pub duration: Duration,
+    pub late: u64,
+    pub late_percent: f64,
+    pub p50: Duration,
+    pub p90: Duration,
+    pub p99: Duration,
+    pub worst: Duration,
+    /// `false` if this phase's frame count exceeds how many samples of it
+    /// are still in the ring -- the percentiles above are then computed
+    /// over whatever survived, not the whole phase. UI_RULES.md: this is
+    /// the "we don't fully know" state, named rather than folded silently
+    /// into a number that looks exact.
+    pub complete: bool,
+}
+
+impl std::fmt::Display for PhaseStats {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        writeln!(
+            f,
+            "  {}: {} frames over {:.1}s{}",
+            self.name,
+            self.frames,
+            self.duration.as_secs_f64(),
+            if self.complete {
+                ""
+            } else {
+                " (ring evicted some of this phase)"
+            },
+        )?;
+        writeln!(f, "    late: {} ({:.1}%)", self.late, self.late_percent)?;
+        writeln!(
+            f,
+            "      total  p50 {:.1}ms  p90 {:.1}ms  p99 {:.1}ms",
+            self.p50.as_secs_f64() * 1000.0,
+            self.p90.as_secs_f64() * 1000.0,
+            self.p99.as_secs_f64() * 1000.0,
+        )?;
+        write!(f, "    worst {:.1}ms", self.worst.as_secs_f64() * 1000.0)
+    }
+}
 
 /// A per-frame wall-time report iris keeps of itself, because `dumpsys
 /// gfxinfo` cannot see a `SurfaceView`'s own GPU-drawn frames at all
@@ -41,6 +113,11 @@ pub struct FrameReport {
     /// "Where iris's frame time goes" CPU/GPU split, added 2026-09-05).
     /// `ring[i] - submit_ring[i]` is that frame's `redraw_to_submit` half.
     submit_ring: Box<[Duration; RING_CAPACITY]>,
+    /// The absolute (0-based, since the last `reset`) frame index each
+    /// `ring`/`submit_ring` slot's sample belongs to -- what `phase_stats`
+    /// slices against `PhaseMark::start_index` to tell which recorded
+    /// frames fall in which phase.
+    index_ring: Box<[u64; RING_CAPACITY]>,
     /// How many of `ring`'s slots hold a real sample -- saturates at
     /// `RING_CAPACITY`, unlike `total_frames` below which keeps counting.
     len: usize,
@@ -50,6 +127,12 @@ pub struct FrameReport {
     /// correct even once the ring itself only holds the most recent frames.
     total_frames: u64,
     janky_frames: u64,
+    /// `mark_phase` calls since the last `reset`, oldest first -- see
+    /// `phase_stats`. Empty on an ordinary run that never calls
+    /// `mark_phase`, so `phase_stats` returns an empty `Vec` and a caller
+    /// prints no "per phase:" section at all, matching RUST.md's "empty/
+    /// absent on an ordinary 'Copy' press, which never marks a phase."
+    phases: Vec<PhaseMark>,
 }
 
 /// One resolved reading. `Display` is the log line both the "Frame report"
@@ -107,10 +190,12 @@ impl FrameReport {
         Self {
             ring: Box::new([Duration::ZERO; RING_CAPACITY]),
             submit_ring: Box::new([Duration::ZERO; RING_CAPACITY]),
+            index_ring: Box::new([0; RING_CAPACITY]),
             len: 0,
             pos: 0,
             total_frames: 0,
             janky_frames: 0,
+            phases: Vec::new(),
         }
     }
 
@@ -131,6 +216,7 @@ impl FrameReport {
     pub fn record_split(&mut self, total: Duration, submit_to_present: Duration) {
         self.ring[self.pos] = total;
         self.submit_ring[self.pos] = submit_to_present;
+        self.index_ring[self.pos] = self.total_frames;
         self.pos = (self.pos + 1) % RING_CAPACITY;
         self.len = (self.len + 1).min(RING_CAPACITY);
         self.total_frames += 1;
@@ -142,12 +228,89 @@ impl FrameReport {
     /// Clears every counter and every sample -- what the "Reset frame
     /// report" control calls, so a report covers only what was scrolled
     /// after the button was pressed (the same reason `FrameStats.kt`'s
-    /// `reset()` exists on the Compose side).
+    /// `reset()` exists on the Compose side). Also clears every phase
+    /// mark, so a fresh run starts with no "per phase:" section until it
+    /// marks one of its own.
     pub fn reset(&mut self) {
         self.len = 0;
         self.pos = 0;
         self.total_frames = 0;
         self.janky_frames = 0;
+        self.phases.clear();
+    }
+
+    /// Marks the start of a named phase at the current moment -- every
+    /// frame recorded from here until the next `mark_phase` (or `reset`)
+    /// belongs to it. RUST.md's "Benchmark v2": a scripted bench run calls
+    /// this once per phase (fling/stream/type/keyboard) so `phase_stats`
+    /// can slice one whole run's frames by what was happening during each.
+    pub fn mark_phase(&mut self, name: &str) {
+        self.phases.push(PhaseMark {
+            name: name.to_string(),
+            start_index: self.total_frames,
+            start_at: Instant::now(),
+        });
+    }
+
+    /// One [`PhaseStats`] per `mark_phase` call since the last `reset`,
+    /// oldest first. `now` closes the last phase's wall-clock span (there
+    /// is no "next phase" instant to use for it); `refresh_hz` is what
+    /// each phase's own `late`/`late_percent` is judged against, read from
+    /// the display rather than assumed -- RUST.md's "Benchmark v2": "late
+    /// count/% against the display's refresh rate."
+    pub fn phase_stats(&self, now: Instant, refresh_hz: f32) -> Vec<PhaseStats> {
+        if self.phases.is_empty() || refresh_hz <= 0.0 {
+            return Vec::new();
+        }
+        let budget = Duration::from_secs_f64(1.0 / refresh_hz as f64);
+        self.phases
+            .iter()
+            .enumerate()
+            .map(|(i, phase)| {
+                let (end_index, end_at) = match self.phases.get(i + 1) {
+                    Some(next) => (next.start_index, next.start_at),
+                    None => (self.total_frames, now),
+                };
+                let frames = end_index.saturating_sub(phase.start_index);
+                let mut samples: Vec<Duration> = (0..self.len)
+                    .filter(|&j| {
+                        let idx = self.index_ring[j];
+                        idx >= phase.start_index && idx < end_index
+                    })
+                    .map(|j| self.ring[j])
+                    .collect();
+                let complete = samples.len() as u64 >= frames;
+                if samples.is_empty() {
+                    return PhaseStats {
+                        name: phase.name.clone(),
+                        frames,
+                        duration: end_at.saturating_duration_since(phase.start_at),
+                        late: 0,
+                        late_percent: 0.0,
+                        p50: Duration::ZERO,
+                        p90: Duration::ZERO,
+                        p99: Duration::ZERO,
+                        worst: Duration::ZERO,
+                        complete,
+                    };
+                }
+                samples.sort_unstable();
+                let pct = |p: usize| samples[(samples.len() * p / 100).min(samples.len() - 1)];
+                let late = samples.iter().filter(|&&d| d > budget).count() as u64;
+                PhaseStats {
+                    name: phase.name.clone(),
+                    frames,
+                    duration: end_at.saturating_duration_since(phase.start_at),
+                    late,
+                    late_percent: 100.0 * late as f64 / samples.len() as f64,
+                    p50: pct(50),
+                    p90: pct(90),
+                    p99: pct(99),
+                    worst: *samples.last().expect("checked not empty above"),
+                    complete,
+                }
+            })
+            .collect()
     }
 
     /// `None` if nothing has been recorded since the last reset -- the
@@ -186,6 +349,28 @@ impl FrameReport {
             gpu_wait_p50: median(submit_samples),
         })
     }
+
+    /// `(late count, late percent)` over every sample still in the ring,
+    /// judged against `refresh_hz`'s own frame budget rather than the
+    /// fixed 60Hz `JANK_THRESHOLD` -- RUST.md's "Benchmark v2": "late
+    /// count/% against the display's refresh rate... print 'at N Hz (X ms
+    /// budget)' like Compose does." A separate method from `report()`
+    /// rather than a parameter on it, so `report()`'s own `janky_percent`
+    /// (and the exact-boundary test pinned to `JANK_THRESHOLD`) is
+    /// unaffected for every existing caller that never measured a real
+    /// refresh rate. `(0, 0.0)` with nothing recorded or a non-positive
+    /// `refresh_hz`.
+    pub fn late_at_hz(&self, refresh_hz: f32) -> (u64, f64) {
+        if self.len == 0 || refresh_hz <= 0.0 {
+            return (0, 0.0);
+        }
+        let budget = Duration::from_secs_f64(1.0 / refresh_hz as f64);
+        let late = self.ring[..self.len]
+            .iter()
+            .filter(|&&d| d > budget)
+            .count() as u64;
+        (late, 100.0 * late as f64 / self.len as f64)
+    }
 }
 
 impl Default for FrameReport {
@@ -296,4 +481,68 @@ mod tests {
         // same pattern here.
         assert!(stats.worst <= Duration::from_millis(5));
     }
+
+    #[test]
+    fn no_marks_means_no_phases() {
+        let mut r = FrameReport::new();
+        r.record(Duration::from_millis(5));
+        assert!(r.phase_stats(Instant::now(), 60.0).is_empty());
+    }
+
+    #[test]
+    fn phases_slice_frames_by_when_they_were_marked() {
+        let mut r = FrameReport::new();
+        r.mark_phase("a");
+        for _ in 0..5 {
+            r.record(Duration::from_millis(10)); // 10ms: late at 60Hz (16.7ms budget)... no, 10<16.7, not late
+        }
+        r.mark_phase("b");
+        for _ in 0..3 {
+            r.record(Duration::from_millis(20)); // 20ms: late at 60Hz
+        }
+        let now = Instant::now();
+        let phases = r.phase_stats(now, 60.0);
+        assert_eq!(phases.len(), 2);
+        assert_eq!(phases[0].name, "a");
+        assert_eq!(phases[0].frames, 5);
+        assert_eq!(phases[0].late, 0);
+        assert_eq!(phases[0].worst, Duration::from_millis(10));
+        assert_eq!(phases[1].name, "b");
+        assert_eq!(phases[1].frames, 3);
+        assert_eq!(phases[1].late, 3);
+        assert_eq!(phases[1].late_percent, 100.0);
+        assert_eq!(phases[1].worst, Duration::from_millis(20));
+        assert!(phases[0].complete);
+        assert!(phases[1].complete);
+    }
+
+    #[test]
+    fn the_last_phase_runs_until_now() {
+        let mut r = FrameReport::new();
+        r.mark_phase("only");
+        r.record(Duration::from_millis(1));
+        std::thread::sleep(Duration::from_millis(20));
+        let now = Instant::now();
+        let phases = r.phase_stats(now, 60.0);
+        assert_eq!(phases.len(), 1);
+        assert!(phases[0].duration >= Duration::from_millis(20));
+    }
+
+    #[test]
+    fn reset_clears_phase_marks() {
+        let mut r = FrameReport::new();
+        r.mark_phase("a");
+        r.record(Duration::from_millis(1));
+        r.reset();
+        assert!(r.phase_stats(Instant::now(), 60.0).is_empty());
+    }
+
+    #[test]
+    fn late_at_hz_uses_the_given_refresh_rate_not_the_fixed_60hz_constant() {
+        let mut r = FrameReport::new();
+        // 10ms is under 60Hz's 16.7ms budget but over 120Hz's 8.3ms one.
+        r.record(Duration::from_millis(10));
+        assert_eq!(r.late_at_hz(60.0), (0, 0.0));
+        assert_eq!(r.late_at_hz(120.0), (1, 100.0));
+    }
 }

iris/src/android/attr.rs

@@ -12,6 +12,10 @@ impl<T: HasAndroidUiState> FocusHost for T {
         self.android_state_mut().focus = id;
     }
 
+    fn is_focused(&self, id: WeakWidget<TextEdit>) -> bool {
+        self.android_state().focus == Some(id)
+    }
+
     fn focus_gained(&mut self, region: Option<PixelRegion>) {
         // Showing the keyboard is a JNI call (`InputMethodManager.showSoftInput`),
         // and this runs deep inside the platform-agnostic sensor dispatch

iris/src/android/ime.rs

@@ -49,6 +49,52 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
     fn focus(&self) -> Option<WeakWidget<TextEdit>> {
         self.state.android_state().focus
     }
+
+    /// Tell Gboard where the caret/selection and the composing region
+    /// actually are, via `InputMethodManager.updateSelection` -- every one
+    /// of android-view's own demo's `set_composing_text_internal`/`render`
+    /// calls this, and this bridge never did, which is what left Gboard's
+    /// own model of the field diverging from `TextEdit`'s real one after
+    /// the very first edit (RUST.md's P0 box, "doesn't enter it until I
+    /// hit space, and also doesn't move cursor forward" -- Gboard holds
+    /// its composing keystrokes back until it believes the app has caught
+    /// up, and without this call it never does). Called from
+    /// [`IrisViewPeer::after_input`], the one tail every touch/key/IME
+    /// callback already runs through, rather than duplicated at each of
+    /// this file's mutating methods.
+    ///
+    /// `candidates_start`/`candidates_end` report the composing region;
+    /// `-1, -1` when nothing is composing, matching `EditorInfo`'s own
+    /// convention. `compose_len` is tracked in `char`s (this module's doc
+    /// comment), so this reports it as that many UTF-16 units back from the
+    /// caret -- exact for the common BMP case, the same approximation
+    /// `set_composing_text` already makes.
+    pub(super) fn update_ime_selection(&mut self, ctx: &mut CallbackCtx) {
+        let Some(focus) = self.focus() else { return };
+        let text = &self.rsc[focus];
+        let Some(sel) = text.selection_range() else {
+            return;
+        };
+        let content = text.text();
+        let sel_start = byte_to_utf16(content, sel.start) as i32;
+        let sel_end = byte_to_utf16(content, sel.end) as i32;
+        let compose_len = self.state.android_state().compose_len;
+        let (comp_start, comp_end) = if compose_len > 0 {
+            let caret = byte_to_utf16(content, text.caret().unwrap_or(sel.end)) as i32;
+            (caret - compose_len as i32, caret)
+        } else {
+            (-1, -1)
+        };
+        let imm = ctx.view.input_method_manager(&mut ctx.env);
+        imm.update_selection(
+            &mut ctx.env,
+            &ctx.view,
+            sel_start,
+            sel_end,
+            comp_start,
+            comp_end,
+        );
+    }
 }
 
 impl<State: AndroidAppState> InputConnection for IrisViewPeer<State> {

iris/src/android/view.rs

@@ -325,6 +325,13 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
             show_soft_input(&mut ctx.env, &ctx.view);
         }
 
+        // RUST.md's P0 box, "doesn't enter it until I hit space, and also
+        // doesn't move cursor forward": Gboard needs `updateSelection`
+        // after every edit to keep its own model of the field in sync, or
+        // it holds keystrokes back rather than trusting a screen it
+        // believes is stale. See `update_ime_selection`'s own doc.
+        self.update_ime_selection(ctx);
+
         let ui_state = self.state.android_state_mut();
         ui_state.cursor.end_frame();
         if self.render.needs_redraw(&ui_state.root, self.rsc.widgets()) {
@@ -650,6 +657,22 @@ impl<State: AndroidAppState> ViewPeer for IrisViewPeer<State> {
         let content_scale = self.state.android_state().content_scale;
         match AndroidRenderer::new(window, width as u32, height as u32, content_scale) {
             Ok(renderer) => {
+                // A genuinely new renderer means a genuinely new GPU device
+                // and a fresh, empty glyph atlas -- the CPU-side glyph
+                // cache (`TextData::atlas`) and the texture bookkeeping it
+                // is built on (`UiData::textures`) both outlive `renderer`
+                // itself (they live on `self.rsc`, not on `AndroidRenderer`),
+                // so without this they would keep pointing at the *old*
+                // device's now-gone textures -- the app-switch counterpart
+                // to the keyboard-resize glyph wipe this same function's
+                // `already_live` branch above already fixed by reusing the
+                // renderer instead of rebuilding it. One mechanism either
+                // way: this call only runs on the branch that actually
+                // builds a new renderer, exactly where invalidation is
+                // needed, never on the reuse branch, where it would throw
+                // away perfectly valid GPU state for nothing.
+                self.rsc.ui.text.atlas.clear();
+                self.rsc.ui.textures.reset();
                 self.state.android_state_mut().renderer = Some(renderer);
                 self.render(ctx);
             }

iris/src/attr.rs

@@ -22,6 +22,13 @@ pub trait FocusHost {
     /// it was hit in (`None` when the widget could not be located, which
     /// happens for one it was just deselected from).
     fn focus_gained(&mut self, region: Option<PixelRegion>);
+    /// Whether `id` is the current focus target -- what [`select`] uses to
+    /// tell a fresh press (which must wait to see whether it becomes a tap
+    /// or a drag before focusing/showing the IME, Iris 2026-09-06: "if I
+    /// swipe over the input bar it brings up the keyboard") from a drag
+    /// continuing inside a field that was already focused (an ordinary
+    /// drag-to-select, unaffected).
+    fn is_focused(&self, id: WeakWidget<TextEdit>) -> bool;
 }
 
 /// Helper shared by every `FocusHost` impl, so the double-click window is
@@ -33,6 +40,17 @@ pub fn recent_click(last_click: &mut Instant) -> bool {
     recent
 }
 
+/// `PressStart`/`Pressing`/`PressEnd`, all for the left button -- what
+/// [`Selector`]/[`Selectable`] register instead of [`CursorSense::
+/// click_or_drag`], so their shared handler (`on_press`, below) sees every
+/// frame of a gesture and can tell a completed tap from a drag itself,
+/// rather than reacting to `PressStart` alone the way `click_or_drag`'s
+/// consumer used to (Iris, 2026-09-06: "if I swipe over the input bar it
+/// brings up the keyboard").
+fn press_track() -> CursorSenses {
+    CursorSense::click() | CursorSense::Pressing(CursorButton::Left) | CursorSense::unclick()
+}
+
 pub struct Selector;
 
 impl<Rsc: HasEvents, W: Widget + 'static> WidgetAttr<Rsc, W> for Selector
@@ -42,7 +60,7 @@ where
     type Input = WeakWidget<TextEdit>;
 
     fn run(rsc: &mut Rsc, container: WeakWidget<W>, id: Self::Input) {
-        rsc.register_event(container, CursorSense::click_or_drag(), move |ctx, rsc| {
+        rsc.register_event(container, press_track(), move |ctx, rsc| {
             let region = ctx.data.render.window_region(&id, &*rsc).unwrap();
             let id_pos = region.top_left;
             let container_pos = ctx
@@ -53,14 +71,14 @@ where
                 .top_left;
             let pos = ctx.data.pos + container_pos - id_pos;
             let size = region.size();
-            select(
+            on_press(
                 rsc,
                 ctx.data.render,
                 ctx.state,
                 id,
                 pos,
                 size,
-                ctx.data.sense.is_dragging(),
+                ctx.data.sense,
             );
         });
     }
@@ -75,31 +93,71 @@ where
     type Input = ();
 
     fn run(rsc: &mut Rsc, id: WeakWidget<TextEdit>, _: Self::Input) {
-        rsc.register_event(id, CursorSense::click_or_drag(), move |ctx, rsc| {
-            select(
+        rsc.register_event(id, press_track(), move |ctx, rsc| {
+            on_press(
                 rsc,
                 ctx.data.render,
                 ctx.state,
                 id,
                 ctx.data.pos,
                 ctx.data.size,
-                ctx.data.sense.is_dragging(),
+                ctx.data.sense,
             );
         });
     }
 }
 
-fn select(
+/// One press-track frame (`PressStart`, `Pressing` or `PressEnd`) over a
+/// selectable field. A field that is *already* focused behaves exactly as
+/// `click_or_drag` always did -- every frame updates the selection, which
+/// is what lets a finger already inside a focused field drag out a
+/// selection. A field that is **not** focused withholds `select`'s
+/// focus-granting side effects (and so the platform-specific `focus_gained`
+/// that shows the keyboard) until the press resolves as a tap: `PressEnd`
+/// with no frame in between having moved past [`DRAG_SLOP`] from where the
+/// press began. A drag recognised before release simply cancels the
+/// pending tap and does nothing further here -- it is not consumed, so
+/// whatever is behind the field (a list to pan) still sees every frame of
+/// it, the same as a drag that never touched a selectable field at all.
+fn on_press(
     rsc: &mut impl UiRsc,
     render: &UiRenderState,
     state: &mut impl FocusHost,
     id: WeakWidget<TextEdit>,
     pos: Vec2,
     size: Vec2,
-    dragging: bool,
+    sense: CursorSense,
 ) {
-    let recent = state.recent_click();
-    id.edit(rsc).select(pos, size, dragging, recent);
-    state.set_focus(Some(id));
-    state.focus_gained(render.window_region(&id, &*rsc));
+    if state.is_focused(id) {
+        let recent = matches!(sense, CursorSense::PressStart(_)) && state.recent_click();
+        id.edit(rsc).select(pos, size, sense.is_dragging(), recent);
+        return;
+    }
+
+    match sense {
+        CursorSense::PressStart(_) => {
+            id.edit(rsc).text.press_origin = Some(pos);
+        }
+        CursorSense::Pressing(_) => {
+            let ctx = id.edit(rsc);
+            if let Some(origin) = ctx.text.press_origin
+                && ((pos.x - origin.x).abs() > DRAG_SLOP || (pos.y - origin.y).abs() > DRAG_SLOP)
+            {
+                // Past the slop before release: this is a drag, not a tap
+                // -- give up the pending focus rather than granting it once
+                // the finger lifts wherever it happens to be by then.
+                ctx.text.press_origin = None;
+            }
+        }
+        CursorSense::PressEnd(_) => {
+            let was_tap = id.edit(rsc).text.press_origin.take().is_some();
+            if was_tap {
+                let recent = state.recent_click();
+                id.edit(rsc).select(pos, size, false, recent);
+                state.set_focus(Some(id));
+                state.focus_gained(render.window_region(&id, &*rsc));
+            }
+        }
+        _ => {}
+    }
 }

iris/src/default/attr.rs

@@ -10,6 +10,10 @@ impl<T: HasDefaultUiState> FocusHost for T {
         self.default_state_mut().focus = id;
     }
 
+    fn is_focused(&self, id: WeakWidget<TextEdit>) -> bool {
+        self.default_state().focus == Some(id)
+    }
+
     fn focus_gained(&mut self, region: Option<PixelRegion>) {
         let state = self.default_state_mut();
         let Some(region) = region else { return };

iris/src/layout_tests.rs

@@ -183,3 +183,80 @@ fn a_mask_stays_put_while_its_scrolled_content_moves() {
     assert_eq!(mask_delta_before, [0.0, 0.0]);
     assert_eq!(mask_delta_after, [0.0, 0.0]);
 }
+
+/// Reproduces `transcript_ui::composer::build_composer`'s exact tree shape
+/// (a `Rect` background stacked behind a `Span::RIGHT`-wrapped, padded,
+/// `rest`-width `TextEdit`, itself the second child of an outer
+/// `Span::DOWN` beside a `rest(1)`-height sibling) without the event/
+/// resource plumbing `composer.rs`'s builders need, to isolate whether the
+/// bug Iris reported on 2026-09-06 ("text seems to not appear in box")
+/// is this crate's layout engine or something specific to the real
+/// composer/screen. `TextEditable::edit` only needs `UiRsc`, so a plain
+/// insert exercises the exact redraw path a keystroke does.
+fn composer_like_tree(rsc: &mut TestRsc) -> (WeakWidget<TextEdit>, StrongWidget) {
+    let field = wtext("")
+        .editable(EditMode::MultiLine)
+        .text_align(Align::LEFT)
+        .wrap(true)
+        .size(18)
+        .color(UiColor::WHITE)
+        .add(rsc);
+    let bar = (field.pad(dp(12)).width(rest(1)),)
+        .span(Dir::RIGHT)
+        .background(rect(UiColor::new(40, 40, 46, 255)))
+        .add(rsc);
+    let list_stand_in = rect(UiColor::BLACK).height(rest(1)).add(rsc);
+    let tree = (list_stand_in, bar).span(Dir::DOWN).add_strong(rsc).any();
+    (field, tree)
+}
+
+/// The reproduction itself. A window this tall stands in for the keyboard
+/// closed; the second, shorter `resize` stands in for `adjustResize`
+/// shrinking the surface when the IME opens -- exactly the sequence
+/// `IrisViewPeer::surface_changed` drives on a real keyboard open. Typing
+/// happens both before and after, since Iris's report was specifically
+/// that text typed *after* the keyboard was already up did not appear.
+#[test]
+fn composing_text_after_a_keyboard_resize_lands_in_the_bars_own_region() {
+    let mut rsc = TestRsc {
+        ui: UiData::default(),
+    };
+    let (field, root) = composer_like_tree(&mut rsc);
+    let mut render = UiRenderState::new();
+
+    render.resize((1080.0, 2298.0));
+    render.update(&root, &mut rsc);
+    field.edit(&mut rsc).insert("a");
+    render.update(&root, &mut rsc);
+
+    let before_px = render.window_region(&field, &rsc).unwrap();
+    // The field is one line plus 12dp of padding on a 2298-tall window --
+    // nowhere near the whole window's height, and anchored at the bottom.
+    assert!(
+        before_px.bot_right.y - before_px.top_left.y < 200.0,
+        "before a resize: {before_px:?}"
+    );
+    assert!(
+        before_px.top_left.y > 1800.0,
+        "expected the bar near the bottom before a resize: {before_px:?}"
+    );
+
+    // The keyboard opens: a real `surface_changed`/`resize` to a shorter
+    // window, then a further keystroke -- the redraw that must land in the
+    // bar's new (also short) region, not whatever region a provisional
+    // measurement pass used along the way.
+    render.resize((1080.0, 1478.0));
+    render.update(&root, &mut rsc);
+    field.edit(&mut rsc).insert("b");
+    render.update(&root, &mut rsc);
+
+    let after_px = render.window_region(&field, &rsc).unwrap();
+    assert!(
+        after_px.bot_right.y - after_px.top_left.y < 200.0,
+        "after a resize + keystroke: {after_px:?}"
+    );
+    assert!(
+        after_px.top_left.y > 1200.0,
+        "expected the bar near the bottom of the shorter window: {after_px:?}"
+    );
+}

iris/src/widget/list.rs

@@ -489,6 +489,24 @@ impl List {
         true
     }
 
+    /// The anchor's own row index and pixel offset, formatted the same
+    /// shape Compose's `firstVisibleItemIndex`/`firstVisibleItemScrollOffset`
+    /// report (`idx=N/off=Mpx`) -- what RUST.md's "Benchmark v2" fling
+    /// phase reads before/after/between its fling runs so the two apps'
+    /// travel can be compared directly. `more_before`/`more_after`
+    /// sentinels print as `idx=more-before`/`idx=more-after` rather than
+    /// leaking their internal `isize` representation; `idx=none` if the
+    /// list has never drawn (no anchor yet -- e.g. right after
+    /// `jump_to_end` and before the next frame runs `repair_anchor`).
+    pub fn anchor_position_display(&self) -> String {
+        match self.anchor {
+            None => "idx=none".to_string(),
+            Some(a) if a.slot == BEFORE_SLOT => "idx=more-before".to_string(),
+            Some(a) if a.slot == AFTER_SLOT => "idx=more-after".to_string(),
+            Some(a) => format!("idx={}/off={}px", a.slot, a.offset.round() as i64),
+        }
+    }
+
     /// Snap to the newest content (last item, or the `more_after`
     /// sentinel if set), bottom-aligned to the viewport. O(1).
     pub fn jump_to_end(&mut self) {
@@ -1456,4 +1474,22 @@ mod tests {
             first.top
         );
     }
+
+    #[test]
+    fn anchor_position_display_before_any_draw_is_none() {
+        let list = List::new(Axis::Y);
+        assert_eq!(list.anchor_position_display(), "idx=none");
+    }
+
+    #[test]
+    fn anchor_position_display_reports_slot_and_offset() {
+        let mut rsc = TestRsc {
+            ui: UiData::default(),
+        };
+        let (list_weak, root, mut render) = build_flingable_list(&mut rsc);
+        let _ = (&root, &mut render);
+        let list_ref = rsc.ui.widgets.get(&list_weak).unwrap();
+        assert!(list_ref.anchor_position_display().starts_with("idx="));
+        assert!(!list_ref.anchor_position_display().contains("none"));
+    }
 }

iris/src/widget/text/edit.rs

@@ -32,6 +32,15 @@ pub struct TextEdit {
     #[cfg_attr(target_os = "android", allow(dead_code))]
     history: Vec<(String, Option<Selection>)>,
     double_hit: Option<usize>,
+    /// Where an in-flight press over this field began, while it is still
+    /// undecided whether the gesture is a tap (focus/show the IME) or a
+    /// drag (attr.rs's `Selector`/`Selectable`, Iris 2026-09-06: a swipe
+    /// over the composer must not summon the keyboard). `None` both before
+    /// any press and once the gesture has been decided either way --
+    /// `attr.rs` is the only reader/writer, kept `pub(crate)` rather than
+    /// behind an accessor since it is pure bookkeeping with no invariant
+    /// beyond "some press is undecided," same shape as `double_hit` above.
+    pub(crate) press_origin: Option<Vec2>,
     pub mode: EditMode,
 }
 
@@ -48,6 +57,7 @@ impl TextEdit {
             selection: None,
             history: Default::default(),
             double_hit: None,
+            press_origin: None,
             mode,
         }
     }
@@ -698,6 +708,67 @@ mod tests {
         assert_eq!(content(&t), "に");
     }
 
+    /// `android/ime.rs`'s `set_composing_text` calls `replace` and expects
+    /// the caret to land right after the inserted text, growing with it on
+    /// every re-send -- the buffer-level half of RUST.md's P0 box ("doesn't
+    /// enter it until I hit space, and also doesn't move cursor forward").
+    #[test]
+    fn composing_advances_the_caret_with_the_growing_text() {
+        let (mut t, mut d) = edit("", EditMode::SingleLine);
+        ctx(&mut t, &mut d).set_caret(0);
+        ctx(&mut t, &mut d).replace(0, "h");
+        assert_eq!(t.caret(), Some(1));
+        ctx(&mut t, &mut d).replace(1, "hi");
+        assert_eq!(content(&t), "hi");
+        assert_eq!(t.caret(), Some(2));
+        ctx(&mut t, &mut d).replace(2, "hit");
+        assert_eq!(content(&t), "hit");
+        assert_eq!(t.caret(), Some(3));
+    }
+
+    /// The IME's `commitText` (`android_view::InputConnection::commit_text`'s
+    /// default body): finish a composition in place, same as a real word
+    /// boundary (a space) landing after Gboard's composing span.
+    #[test]
+    fn committing_composed_text_leaves_it_in_place_with_the_caret_after_it() {
+        let (mut t, mut d) = edit("say ", EditMode::SingleLine);
+        ctx(&mut t, &mut d).set_caret(4);
+        ctx(&mut t, &mut d).replace(0, "hi");
+        assert_eq!(content(&t), "say hi");
+        // `finish_composing_text`/`commit_text` do not themselves touch the
+        // buffer -- only the IME's own `compose_len` bookkeeping resets, in
+        // `android/ime.rs`. Confirms the buffer already holds committed
+        // text as plain, uncomposed content: a further `replace(0, " ")`
+        // (the space that ends the word) appends rather than overwriting.
+        ctx(&mut t, &mut d).replace(0, " ");
+        assert_eq!(content(&t), "say hi ");
+        assert_eq!(t.caret(), Some(7));
+    }
+
+    /// `TextEditCtx::delete_byte_range` is `deleteSurroundingText`'s entry
+    /// point once `android/ime.rs` has converted UTF-16 code units to
+    /// bytes -- exercised directly here in bytes, since the UTF-16 math
+    /// itself is `android/ime.rs`'s own `byte_to_utf16`/`utf16_to_byte`,
+    /// outside this widget-only test module.
+    #[test]
+    fn delete_byte_range_removes_exactly_that_range() {
+        let (mut t, mut d) = edit("hello world", EditMode::SingleLine);
+        ctx(&mut t, &mut d).delete_byte_range(5, 11);
+        assert_eq!(content(&t), "hello");
+        assert_eq!(t.caret(), Some(5));
+    }
+
+    /// `set_cursor_byte` is `setSelection`'s entry point -- collapses to a
+    /// caret at the given byte offset regardless of any span that was there.
+    #[test]
+    fn set_cursor_byte_collapses_to_a_caret_there() {
+        let (mut t, mut d) = edit("hello world", EditMode::SingleLine);
+        ctx(&mut t, &mut d).select_all();
+        ctx(&mut t, &mut d).set_cursor_byte(5);
+        assert_eq!(t.selected_text(), None);
+        assert_eq!(t.caret(), Some(5));
+    }
+
     #[test]
     fn motion_moves_the_caret_and_shift_extends_a_span() {
         let (mut t, mut d) = edit("abc", EditMode::SingleLine);

iris/transcript-ui/src/composer.rs

@@ -8,14 +8,58 @@
 //! measures whatever vertical space is left each frame -- nothing here
 //! computes a height by hand, and growing this field is exactly the
 //! O(1)-move-chain case LAYOUT.md and I3's benchmark already measured.
+//!
+//! **Rebuilt 2026-09-06** (Iris's phone report on the dc01f88 build: the
+//! grey bar drawn as a short, fixed strip with the typed text ~150px below
+//! it on black, and empty black between the bar and the keyboard). One
+//! widget now, top to bottom: an opaque background sized to its content
+//! (`.background`, the same `Stack` idiom the header row's `HEADER_SURFACE`
+//! already uses), the field inside `dp` padding and capped at
+//! [`MAX_LINES`] before it scrolls instead of growing forever, and an
+//! outer [`Pad`] whose `bottom` [`TranscriptScreen::set_bottom_inset`]
+//! rewrites in place whenever the keyboard opens/closes -- never rebuilt,
+//! since `field` is strongly owned inside this tree and this crate's
+//! widgets cannot be re-parented once added (this module's own comment
+//! below on why `build_composer` hands back a **weak** id).
 
 use iris::prelude::*;
 
+/// Caps the field's growth at roughly six lines of its own 18px text
+/// before it scrolls instead of consuming the whole screen -- an
+/// approximation (line-height and padding folded into one round `dp`
+/// number) rather than a value derived from the font's real metrics,
+/// which nothing in this crate exposes to a caller today.
+const MAX_LINES: f32 = 6.0;
+const APPROX_LINE_HEIGHT_DP: f32 = 24.0;
+const FIELD_PAD_DP: f32 = 12.0;
+
 /// `field` is exposed so the caller can read its content on submit
 /// (`field.edit(rsc).text()`) and clear it afterward
 /// (`field.edit(rsc).set("")`).
 pub struct Composer {
     pub field: WeakWidget<TextEdit>,
+    /// The bar's own outer padding -- only `bottom` is ever changed, by
+    /// [`Self::set_bottom_inset`]. A `Pad` around the whole bar rather than
+    /// a rebuilt tree, because `field` lives inside it and cannot be
+    /// re-added to a new wrapper once it is strongly owned here.
+    outer_pad: WeakWidget<Pad>,
+}
+
+impl Composer {
+    /// Called by the platform shell (Android's `on_insets_changed`, e.g.)
+    /// whenever the space below the bar changes: the IME's own inset while
+    /// it is open, the navigation-bar inset otherwise. Takes a plain
+    /// `f32` in the caller's own physical-pixel units rather than an
+    /// Android-specific insets type, so this crate stays usable from the
+    /// winit backend too, which has no navigation bar to report.
+    /// Rewrites the existing `Pad` in place (marking it dirty through the
+    /// ordinary `Widgets::get_mut` path) instead of swapping in a new one,
+    /// so the field's focus, selection and in-progress text are untouched.
+    pub fn set_bottom_inset(&self, rsc: &mut impl UiRsc, inset: f32) {
+        if let Some(pad) = rsc.ui_mut().widgets.get_mut(&self.outer_pad) {
+            pad.padding.bottom = Len::abs(inset);
+        }
+    }
 }
 
 /// Returns the composer plus its own bar as a **weak** id -- the caller
@@ -39,10 +83,20 @@ where
         .label("Message")
         .add(rsc);
 
-    let bar: WeakWidget = (field.pad(dp(12)).width(rest(1)),)
-        .span(Dir::RIGHT)
+    // One widget: an opaque bar sized to its own content (`.background`'s
+    // `Stack{child: 1}`, the header row's own idiom) wrapping the padded,
+    // height-capped field -- not a background rect and a field drawn as
+    // two independent siblings, which is what let the two disagree on
+    // where the bar actually was.
+    let content = field
+        .pad(dp(FIELD_PAD_DP))
+        .max_height(dp(APPROX_LINE_HEIGHT_DP * MAX_LINES + FIELD_PAD_DP * 2.0))
+        .scrollable()
+        .width(rest(1))
         .background(rect(UiColor::new(40, 40, 46, 255)))
         .add(rsc);
 
-    (Composer { field }, bar)
+    let outer_pad: WeakWidget<Pad> = content.pad(Padding::ZERO).add(rsc);
+
+    (Composer { field, outer_pad }, outer_pad)
 }