ai-app/docs/DECISIONS.md

Decisions taken for Iris to review

Short list of design choices made by the design agent without asking, so they can be judged and reversed later. Detail lives in RUST.md (and IRIS.md for iris API changes); this file is only the summary. Newest first. Items marked DEFERRED are ones the agent chose not to decide alone.

2026-09-07 (platform fonts, not bundled ones)

• Iris's own decision, carried out as directed: removed the 3.6 MB of bundled Noto Sans/Noto Sans Mono TTFs from iris-core and load text from the platform's own font collection instead (fontique's system discovery, already on by default). Matches what the Compose app does -- it takes body text from FontFamily.Default and code text from FontFamily.Monospace, both platform-resolved, and ships no text font of its own. Rejected alternative (the one this pass had left open 2026-09-06): subsetting the bundled Noto Sans to Latin/common punctuation instead of removing it outright, which would have kept identical rendering across devices for a smaller (not zero) size cost; Iris chose to match Compose instead.
• .so -3,748,136 bytes (11,193,608 -> 7,445,472), matching the original 3.6 MB estimate. Fallback still lands on the platform's own tofu for a codepoint no resolved face has (checked with CJK + emoji on desktop) rather than blank space, so the UI_RULES unknown-glyph rule still holds.
• Flagging for Iris, not blocking on it: this fontique version's Android backend never resolves the Monospace generic family at all (confirmed on this checkout's emulator, mono=None in the startup diagnostic) -- a pre-existing ordering bug in fontique's own fonts.xml parsing, not something this change introduced, but this change is what stopped masking it (the bundled mono font used to be registered ahead of the broken platform lookup, so it always won). Effect: code blocks and the tool-card chevron marks fall through to the same face as body text on Android instead of a genuinely monospaced one -- still visible, not blank, just not monospaced. Compose does not share this gap; it resolves FontFamily.Monospace through Android's own Typeface.MONOSPACE, not through fontique. docs/RUST.md's "Platform fonts (2026-09-07)" has the full account and why a fix (an OEM-specific named-family pin, or patching around fontique) is left as a follow-up rather than done in this pass.

2026-09-07 (how a phone log reaches Iris)

• The app sends its own log to ai-server, and Dev Updater shows it as ai-server's runtime log. Iris has no adb/logcat on her phone, and Android forbids one app reading another's logcat, so the app has to carry its own copy and post it somewhere. POST /client-log on ai-server re-emits each line into that server's own tracing output; Dev Updater already runs ai-server as a Managed component, whose stdout its own service script redirects to a file and reports through GET /apps/{key}/components/{name}/logs?kind=runtime, which the phone app's log dialog already offers as a Runtime tab for a server component. So no change to Dev Updater at all -- one route on ai-server, and the client in client-core.

  Rejected: posting to Dev Updater's own server (the first candidate). It would need a new authenticated write route on a TLS surface whose module doc says every route on it "is, or decides, the bytes that get handed to REQUEST_INSTALL_PACKAGES next"; a per-app device-log store; a change to component_logs so an APK component can have a runtime log; a change to the phone app's hasBothKinds = component.kind == "server" gate and to what hasRuntimeLogs means on the wire; and -- the real cost -- a second enrollment for the iris app, since it has no CA or token for Dev Updater and Dev Updater mints tokens per device by QR. Five changes across two repos against one route, for the same line landing in the same viewer.

  Rejected: a share intent from a debug button (a log file in the app's external files dir, shared by hand). It works today and needs no server, but every line costs Iris a manual export and a message, which is the round trip through a person this was meant to remove. It is still the fallback when the tunnel is down, and GrapheneOS's own per-app log export already covers the crash case (that is how the ToolInput.highlighted crash was reported).

• The ring is in client-core, not in the Android crate. A bounded in-memory ring (2000 lines or 256 KiB, whichever bites first) behind a log::Log backend that forwards to whichever logger the platform already installed, so logcat and a desktop terminal see exactly what they saw before. The platform supplies only its own logger and its destination. Copy report appends the ring to what goes on the clipboard, and flushes the uploader first.

• The destination is baked in at build time, from the build machine's own files (AI_APP_LOG_HOST/_PORT/_TOKEN plus the pinned CA) -- all three or none, never two. The same trust boundary the transcript config and the Compose APK's CA already use: nothing secret is committed, and an APK is good for the server that built it. A build told nothing still keeps its ring and still copies it; the diagnostics pane says which of "not tried yet", "failing -- " and "no server configured" it is, because otherwise all three look like silence.

2026-09-06 (how a tool call looks, P1b)

• A card that never got a result says "no result", in yellow, and it is a state Compose cannot say. A call that finished having printed nothing and a call whose turn was interrupted before anything came back both leave an empty output. Compose draws both as an ordinary finished call, which reads as a fact somebody established. There are five states now, each with a word and a colour: nothing at all for a call that worked, "running" (grey), "your turn" (peach, Compose's own wording and colour), "failed" (red), "no result" (yellow).

• A failed call is drawn as failed, which needed a field on the wire. is_error is on the CLI's tool_result and was being dropped; the server now carries it to the phone. Reversible, but the alternative is a card that says a call succeeded because it cannot tell.

• A group's cards do not each carry their own surface. Compose gives each card a fill and squares the corners where it faces a neighbour, so a run reads as one object broken into parts. iris has no per-corner radius, and -- more to the point -- a group built the way Compose builds it hit a framework layout defect that drew every card's text a card below its own box. So a group is one surface with its cards on it, separated by a small gap, and the 4dp inset Compose holds them off the edge by is gone. Worth revisiting once the layout defect is fixed (docs/IRIS_TODO.md).

• A long tool output is capped at 80 lines or 4 kB with a "Show all N lines". Compose draws the whole thing, and gets away with it because its Text inside a LazyColumn lays out lazily; here the output is one text widget and shaping a hundred kilobytes of it costs what the file editor's 32 kB limit was measured against. If iris's text gets cheaper, this is the number to move.

• A card's command is clipped, not pannable, and its summary line is clipped rather than ellipsised. Both are framework gaps rather than choices (scrollable_on on a non-editable text draws nothing; there is no overflow ellipsis), and both are worse than Compose today. Named here because they are visible.

2026-09-06 (how a markdown block looks, P1a)

• A table is drawn as padded monospace columns, not as a grid. Your call to reverse. Compose draws a real grid: cells on a tint, each column with a 136dp floor, scrolling sideways when there are too many. iris has no grid widget, and building one would be a widget per markdown feature -- which is the thing the block model exists to avoid. In a monospace face a character count is a pixel width, so padding each cell to its column's width is alignment, the widths are still measured from the cells, and a table that is too wide pans sideways through the same mechanism a code fence already uses. The header is bold with a rule under it, and a long cell wraps inside its column (capped at 28 characters, which is what fits three columns across a phone). What it trades: no cell borders, and a table looks like code rather than like a table. If you want the grid, it is a new widget and it is a day's work.
• Three block frames, and only three. A heading, paragraph and list are plain text with spans; a fence and a table are a rounded panel that does not wrap; a quote is a bar with the text padded past it. Everything else markdown says is expressed in span styles, which cost no widgets and no layout nodes. So a new markdown feature is a span, not a widget.
• A list's marker is part of the text, so a wrapped item's second line returns to the left margin. Compose keeps it indented by giving the marker its own column. Doing the same here needs per-line indent in iris's text attributes; it is written down rather than done, because the list items in a real reply are usually one line.
• A link opens on a tap and not on the end of a drag. A press that panned the transcript past a link, or that held long enough to start a selection, does not follow it -- decided by the same gesture machine that decides pan-versus-select, so there is one rule rather than two that can disagree.

2026-09-06 (composer scroll and the streaming block model)

• A streamed message becomes a column of per-block widgets. Decided by the design agent; recorded here because it is the shape of every message on screen. A transcript row is one TextEdit today, so a streamed delta re-shapes the entire message through parley on every event -- the stream phase is the one place iris is behind Compose on your phone (p50 18.2ms vs 13.4ms). A row becomes a column of one widget per markdown block (paragraph, heading, fence, list, table) and a delta replaces only the last block, keeping every earlier block's layout. Rejected: splitting parley's layout at block boundaries inside one text widget (couples iris's text widget to markdown structure, and parley has no incremental API), and caching shaped runs per paragraph inside TextEdit (a second cache with its own invalidation beside the glyph cache). Chosen because P1's markdown block model is needed anyway, so the split happens once, in client-core, and iris stays a text renderer. Status: designed, not built -- this pass spent its budget on the composer's three layout defects; docs/RUST.md has the design and the pass conditions.
• The composer's overflowing text now scrolls on a finger, capped at six lines and clipped to the bar. Reverses the "still does not scroll" item below.
• A widget may not report a dp length (see IRIS.md). A rule for widget authors, enforced by a debug_assert!; nothing changes for app code.

2026-09-06 (stale-primitives and touch-scroll pass)

• A vertical drag inside a focused composer now scrolls rather than selects. Android's own EditText does this -- a vertical drag scrolls the field, and only a long press starts a selection -- so the platform decided it. What it costs: you can no longer drag straight down inside the composer to select several lines of what you typed; use a long press and then drag, or drag sideways. Say if that trade is wrong for you.
• Scroll gets a finger pan but no fling. List flings; a scroll area does not, because it has no per-frame tick to animate one and the areas it wraps are at most a screenful (Android does not fling a six-line text box either). Easy to add later if a scroll area ever wraps something long.
• The composer still does not scroll its overflowed text, though the mechanism it needs is now in place. Wrapping the field in .scrollable() was tried and reverted the same day: Scroll measures its content and container against the window, so inside the MaxSize that caps the composer at six lines the two are in different spaces and the field pans itself entirely out of the bar (measured on the emulator with 474 characters in it -- the bar collapsed to its padding). Fixing that means Scroll measuring against its own offered box, which is a change to a widget the transcript and the bench shell both use, so it is its own piece of work rather than a rider on this one.

2026-09-06 (defect pass)

• The keyboard-open diagnostics overlay is gone; the capture only logs now. It was added when on_insets_changed was not firing at all and there was no way to get a report off the phone. It fires reliably since the activity went edge-to-edge -- and what that looks like in use is a full-screen report covering the app every time the keyboard opens, with its own Copy/Close buttons sitting underneath the keyboard, so it cannot be dismissed (reproduced on the emulator this pass: two tap 'CLOSE' runs left it up). An interruption for something nobody asked for, over the app you are trying to type into. The named Diagnostics button still shows the same text on demand, and the new iris surface:/iris insets: log lines carry the lifecycle a logcat pull needs. Reversible: capture_keyboard_diagnostics is still the one place this is decided, and PlatformHandle::show_diagnostics_overlay is still there.

• The bench shell's report pane is sized to its report, not to a share of the window. It held .height(rest(1)) beside the transcript's rest(2), so an empty TextEdit reserved a third of every screen -- which is what Iris's "the app does not start with keyboard spacing correct" screenshot was showing, with the composer two thirds down and black below it. It is .max_height(dp(260)) now and sits above the transcript rather than under the composer, where it was eating the navigation-bar clearance. Cost: a filled report is clipped at 260dp rather than scrolling (a Scroll there drew itself off the top of the screen, since Scroll pins to the end of its content and reports its content's full length to the parent -- worth fixing in Scroll, not worked around here). "Copy report" and logcat still have the whole thing.

2026-09-05

• iris no longer asks every device for compute-shader limits it never uses. adapter.request_device (both iris/src/android/render.rs and iris/src/default/render.rs) used Limits::default() plus an override for max_buffer_size, and Limits::default() unconditionally requests desktop-tier compute limits (max_compute_workgroups_per_dimension: 65535, per wgpu_types) even though nothing in iris/iris-core creates a ComputePipeline or writes a @compute shader stage — confirmed by grepping the whole tree, not assumed. That crashed request_device outright on the Android emulator's software GL path (EMU_GPU=software, --features force-gles): SwiftShader's GL reports itself as OpenGL ES 3.0, which has no compute shaders at all, so the adapter's real limit is 0 against the unconditional request for 65535 — RUST.md's "Software mode ... crashes for a third, different reason," 2026-09-05, earlier today. The same would happen on any real GLES-3.0-only Android device, not just the emulator. Fixed by a new iris_core::device_limits() (iris/core/src/render/mod.rs), shared by both platform backends so the two requests cannot drift, that zeros the six max_compute_* fields explicitly rather than switching to a downlevel Limits preset — Limits::downlevel_webgl2_defaults() was considered and rejected: it also zeros max_storage_buffers_per_shader_stage, and shader.wgsl's vertex stage reads four var<storage> buffers (rects, glyphs, masks, move_offsets), so that preset would trade the compute crash for a bind-group-layout one on the same downlevel hardware this is meant to support. No capability check or fallback path was needed since nothing is being disabled — the request is simply narrowed to what the pipeline actually uses. rigs/gpu-probe's own mirrored limits (it is deliberately its own crate, not a workspace member, so it cannot call device_limits() directly) were updated to match, and confirm IRIS DEVICE: ok against this VM's own Vulkan and GL adapters. Not verified this pass: the specific SwiftShader-ES-3.0 crash this fixes, on-device — the EMU_GPU=software cold boot this needs would have force-restarted this checkout's emulator while another session was actively running its own app on it (com.example.aiapp had window focus at the time), so it was left for a pass when the emulator is free rather than disrupting that session. Everything reachable without the emulator is clean: cargo fmt/clippy --workspace --all-targets/test --workspace, cargo ndk build/clippy for iris-android-app with force-gles, and gpu-probe against this VM's own Vulkan and GL(ES 3.2, which still has compute and so would not have reproduced the crash even before this fix — not a substitute for the real ES-3.0 test).

• P0's Compose half is built and smoke-tested on the emulator — the bench build type, the shared app/bench-fixture/ transcript, and an in-process fake backend (BenchFixture.kt/BenchNetwork.kt) that answers TranscriptSource/EventStream from an in-memory event log instead of a real server, so the fold and paging under test are the real ones. Full account, the smoke run's report, and what is deliberately left (the iris half, the real on-phone runs) are in RUST.md's P0 box. Not a decision to review so much as the gate itself now being runnable — flagged here because it is the first half of something Iris explicitly asked to see before P1.

• P0's iris half is also built and smoke-tested on the emulator, 2026-09-05. A new bench Cargo feature on iris-android-app, on top of transcript-screen: the same checked-in fixture (include_str!, no asset pipeline needed), the same 24-swipe scroll loop animated through List::scroll and the same 400-event/20s streaming phase through fold_event, "Run benchmark"/"Copy report" as named accessible controls, and the same three added report fields (process CPU time, peak RSS, battery current) via direct JNI calls (bench_jni.rs::PlatformHandle) since android_view has no BatteryManager/ClipboardManager wrapper of its own. One small public API addition to get there: AndroidAppState::platform_ready (IRIS.md), a default-no-op lifecycle hook handing an implementor a JavaVM + GlobalRef it can call Java through from any thread. Packaged with a new release build type on iris-android-app's own Gradle project (there was previously only debug), signed with the same key app/build-apk.sh generates. Smoke run and the full report are in RUST.md's P0 box; not attempted this pass: the real on-phone runs and Iris's pass/fail call, which is the actual gate.

• The intermittent touch-scroll dropout is root-caused and fixed: a missed ACTION_DOWN hit-test, not the previously-suspected coalesced first ACTION_MOVE. Diagnosed by temporary logcat tracing of every touch event, DragArbiter state transition and Selection::drag dispatch (removed once confirmed), reproduced on this checkout's own emulator against a real sandbox session. The trace showed the actual mechanism: a gesture's ACTION_DOWN lands wherever the finger actually is, which is not guaranteed to fall inside the same row-local sensor region a later ACTION_MOVE in the same gesture lands in (a row's own padding/gap, or its non-selectable sender-name header, is pointer-transparent to iris::sense::CursorSense). When that happens, the widget that ends up handling the gesture never saw PressStart, so DragArbiter sits in Idle — which answers every subsequent frame with Undecided and has no way to tell "no press is happening" from "a press is happening but I missed its start," so it never recovers on its own for the rest of that gesture. One real trace showed exactly this: touch Down/Move/Up all delivered correctly, but zero PressStart reaching the arbiter, state=Idle unchanged from first frame to last. Fixed at the call site that has the context to recover (iris::transcript_ui::selection::Selection::drag, iris/transcript-ui/src/selection.rs): a new DragArbiter::is_idle() (iris/src/sense.rs) lets it notice a Pressing frame arriving with the arbiter still Idle — which can only mean a missed PressStart, since a Pressing sense requires the button to genuinely be down — and start the press there instead of where it was missed. Three new unit tests in sense.rs's drag_arbiter_tests and one in transcript-ui's selection::tests (the latter fails on the code before this fix). Commit follows. Not the same failure the earlier pass's DECISIONS.md DEFERRED item speculated about (a coalesced first ACTION_MOVE skipping slop detection) — that hypothesis is now ruled out; the arbiter's own slop/long-press logic was never wrong. RUST.md's I5 box, "Touch-scroll dropout root-caused, 2026-09-05" has the full trace.

• P0, a phone benchmark gate before any porting, asked for by Iris 2026-09-05: "before P1 I'd like to see benchmarks & also maybe stress test on my own phone ... If it doesn't match compose reasonably well then I don't think I'd wanna continue." Design (RUST.md's P0 box has the detail): the same embedded synthetic fixture in both apps with no server needed; the same scripted scroll loop then a streaming phase, run programmatically since the phone has no usable system tracing and no agent can drive it; the same report from both (frames, janky %, p50/p90/ p99, process CPU time, peak RSS, battery current where readable) with a copy button; the iris app under its own id and the Compose one as a new bench build type with an id suffix, so neither replaces her production install; two arm64 APKs plus instructions delivered under ~/host/bench/. The gate is hers: iris within a reasonable margin of Compose release on p50, p99 and CPU time, no crashes, no visible stutter. If it fails, the port stops.

• The rest of the port is one UI crate, iris/app-ui, grown out of iris/transcript-ui rather than started beside it. It holds a Screen enum plus a back stack — the Rust equivalent of AppRoot.kt's when — and iris/desktop-app/iris/android-app become thin entry points over it. Chosen over a fresh crate because transcript-ui already has the right generic shape (Rsc: HasEvents + Rsc::State: FocusHost) and the client-core/event-model path dependencies every later screen needs, so growing it in place is the smaller diff. Platform-only code (notification service, share target, QR scanner, Keystore token, deep-link enrolment) stays in the E3/E5 Java shell (android-shell/ + app/shellApp) rather than moving into this crate, since none of it is a screen. The Android APK is built by cargo xtask apk (E5), merging the app-ui cdylib into the E3 shell so there is one app rather than a demo shell plus a service shell. app/androidApp (the Compose app) stays untouched and is the baseline every step is measured against, until parity is reached (P7 decides the switch, and is itself a load-bearing decision left to Iris). Order is by risk to the daily-use path: session screen first (P1, where every hard behaviour already lives), then the shell merge and a real phone install (P2), then root tabs (P3), the explorer (P4), settings/enrolment (P5), desktop parity (P6), and the cutover itself (P7). Full plan: RUST.md's "The port, in order (decided 2026-09-05)".

• iris gets its own measured frame report, rather than waiting on a dumpsys/gfxinfo answer that cannot see a SurfaceView's GPU-drawn frames. iris_core::FrameReport (iris/core/src/render/frame_report.rs) times each frame's wall clock from the same point render()'s redraw starts to just after queue.submit + present() — the span Compose's own render report and gfxinfo both count — into a fixed 4096-entry ring (no allocation per frame; report() is the only place that allocates, and only on a button tap). The report gives total frames, janky % over the same 16.7ms budget gfxinfo uses, P50/P90/P99 and the worst, plus a reset. Exposed the way the Compose app's copy-button report already is: two named controls ("Frame report", "Reset frame report") on the transcript screen, tappable by accessibility name via ui-trace, logging under this crate's fixed android_logger tag (iris-android-app) so a script can grep "iris frame report" the way transcript-bench.sh greps "ai-app render report". The report's own Display line says plainly that it measures up to the present() call returning, not GPU/compositor completion — wgpu's present() is not fenced against either, so presenting that span as "time to reach the screen" would be a measured-looking number that is actually inferred, which the standing UI rule forbids.

• ui-trace gains a hold-then-drag gesture, additive, in emulator-tools. Neither of its two existing actions can produce "hold stationary for LONG_PRESS, then move without lifting" — tap has no hold and swipe X1 Y1 X2 Y2 MS interpolates motion across its whole duration from t=0. A new action presses, waits, then moves to a second point and releases as one continuous touch (raw sendevent/MotionEvent injection, extending whatever mechanism the existing swipe already uses), so DragArbiter's pan-vs-select rule (iris/src/sense.rs, already covered by 8 unit tests against a synthetic clock) can finally be driven on a real device instead of only in a test harness.

• Touch drag on a transcript row follows Android's own rule: a vertical drag pans the list immediately; a stationary press held 500 ms starts a text selection which further dragging extends; a horizontal drag while something is already selected extends that selection without the wait. One DragArbiter per list decides it (iris/src/sense.rs). Chosen over a "text layer always wins" or "list always wins" rule because either loses one of the two gestures a reader expects.

• E4's desktop shape is a new iris/desktop-app crate: a winit window holding transcript-ui's screen beside a session list, talking to a real ai-server through client-core. It enrols by pasting the same aiapp://enroll?… link a phone scans (client-core::config::EnrolledServer) and keeps it owner-only under $XDG_CONFIG_HOME/ai-app-desktop/. The pinned CA is a path given on the command line, not baked in. Chosen so the phone and desktop share one enrolment format and no second one is invented.

• I5's Android integration extends iris-android-app (I2's shell) behind a Cargo feature (transcript-screen), rather than a third shell crate. That project already has the Gradle module, the IrisView/MainActivity Java, and the JNI registration; the only thing a second screen needs on top is a different AndroidAppState, the same axis tabs_ui::build/transcript_ui::build already vary along on the winit side. tabs-screen/transcript-screen are mutually exclusive and each pulls in only its own deps, so the plain tabs build (I2/I4) is untouched.

• Order of remaining work, updated 2026-09-05: the two in-flight pieces and I5's Android integration are all done; next is giving iris its own frame-timing report so item 3 below can be decided by a number.

• DECIDED by Iris, 2026-09-05: iris is the app's framework; Masonry was the calibration. Her words: "I think iris definitely makes more sense based on the limitations we've found." The limitations: Masonry has no touch scroll on Android (E2), no per-span rich text and no cross-row selection on the pinned commit (E2), and its keyboard bridge is a TODO (E1); iris carries the same screen under the Compose baseline on the host GPU (p50 15.0 ms against Compose's 20.0 ms, RUST.md's I5 box). What follows: the E-steps are closed as calibration, and the port proceeds on iris — screens, the shell (E3/E5), and client-core underneath. The item below is kept as the record of what she decided from.

• Was DEFERRED — whether to commit to iris over Masonry for ai-app. Updated 2026-09-05 with the clean comparison the recommendation wanted: same sandbox session content, same emulator, EMU_GPU=software, one session. Headline numbers (RUST.md's I5 box, "Clean scroll comparison, 2026-09-05," has the full table and every caveat):

  app	build	frames	janky %	p50	p90	p99	worst
  Compose (in-app report)	debug	1102	99.0% late	33.8ms	50.6ms	79.5ms	--
  Compose (dumpsys gfxinfo)	debug	1499	21.15% (95.66% legacy)	32ms	48ms	150ms (p99)	--
  iris (FrameReport)	release	299	94.65%	79.1ms	98.6ms	117.8ms	212.6ms
  iris (FrameReport, repeat)	release	233	94.42%	109.3ms	130.8ms	147.1ms	150.5ms

  Not a clean apples-to-apples reading, stated plainly rather than smoothed over: iris had to be built release (debug SIGSEGVs on this emulator's Vulkan loader, I4's finding) against Compose's mandated debug build, so this asymmetry likely understates iris's gap rather than the reverse; the three frame-time sources measure different things (Compose's own phase accounting vs. Android's HWUI deadline-miss definition vs. iris's redraw-start-to-present window, the last of which dumpsys gfxinfo cannot see at all for iris's SurfaceView); and both figures are emulator numbers under software rasterisation, which Compose's own in-app report shows already costs 20-34ms/frame in swap+gpu alone under this GPU mode, so a same-mode iris number well above 16.7ms was expected going in for either app. A second pair under -gpu host was not taken this pass. The earlier session's suspected intermittent touch-delivery dropout was not reproduced this pass — the zero-frame results this time traced to this pass's own script bug (a cd that changed which emulator ui-trace targeted), not the emulator; a CPU-load rise during the gesture was observed by a sampler running throughout, but did not correlate with any failure, so the original candidate is neither confirmed nor ruled out. The choice in front of Iris, updated: decide now on the structural-plus-functional case already made (iris works end-to-end where Masonry's scroll gesture doesn't exist at all on Android) plus this table — reading the two build profiles and three jank definitions with the caveats above rather than as a single number — or ask for a same-profile, same-GPU-mode rerun first. RUST.md's I5 box has the full account.

  Updated 2026-09-05, the -gpu host pair taken. Real GPU rendering (force-gles -- the default Vulkan backend has no adapter at all under plain host-GPU boot, confirmed by the exact wgpu error) reverses the software-mode shape:

  app	build	GPU mode	frames	janky %	p50	p90	p99	worst	cpu p50	gpu-wait p50
  Compose (in-app report)	debug	host (virgl)	1268	96.4% late	20.0ms	28.4ms	37.7ms	--	--	--
  iris (FrameReport), best of three, 2026-09-05	release, force-gles	host (virgl)	439	46.24%	15.7ms	23.3ms	31.2ms	57.4ms	1.2ms	13.2ms

  Under real GPU rendering iris's median frame is faster than Compose's, not the 2-3x-slower shape the software-mode table shows. A new split inside FrameReport (redraw-to-submit vs. submit-to-present, commit e2a1fad) says why: iris's own CPU work per frame is a median ~1ms -- almost the entire frame is time spent handing the frame to the driver, not in iris's layout/text/primitive code. This is consistent with the earlier software-mode gap being mostly SwiftShader's CPU rasterisation cost rather than an iris-specific slowness. Still not proof, and now closed as unanswerable rather than merely untaken: a same-mode software force-gles run to isolate the backend was retried 2026-09-05 after fixing the compute-limit crash the first attempt hit, and hit a second, structural wall instead — SwiftShader's ES 3.0 GL path has no storage-buffer capacity at all, and shader.wgsl reads var<storage> buffers unconditionally, so reaching that path needs a shader rewrite, not a limits fix (RUST.md's I5 box, "The three remaining I5 verifications, closed 2026-09-05," item 2). The intermittent touch-scroll dropout this pass also reproduced is root-caused and fixed as of the same date (a missed ACTION_DOWN on a row's padding/header left DragArbiter stuck in Idle); three clean iris-scroll.sh runs post-fix each scrolled all 24/24 swipes, replacing the single-attempt 62-frame reading this table used to carry. RUST.md's I5 box, "Where iris's frame time goes, 2026-09-05, the -gpu host pass," and "The three remaining I5 verifications, closed 2026-09-05," have the full account. The iris-vs-Masonry choice itself is still Iris's to make.