iris: ship an icon font subset, and delete the drawn mark

Iris asked why `mark` existed at all -- "the font should be working if
it's working for compose and nerd fonts are bundled". It was not: the
Compose app draws its icons from its own committed Nerd Fonts subset,
while iris, which bundles no font since 2026-09-07, was setting the
disclosure mark with bare geometric codepoints (U+25B8/25BE/25B4) out of
whatever face the platform resolved -- an empty box on her phone, a dot
on this VM. The 2026-09-07 note that "iris had no equivalent icon font to
keep" is the gap: it had none because it had never had one.

So iris ships the same kind of subset. iris/core/build-icon-font.sh is
the Compose script with its own GLYPHS list, writing a 992-byte
nerd_icons.ttf with three Material Design glyphs from the Mono face;
iris::icon names the codepoints; Family::Icons is how text asks for them.
The variant names an intention rather than a font name -- only TextData
knows what the file registered as, and it resolves it during shaping --
and it is a named family, never a generic one, so nothing falls back into
it for text and an icon cannot fall back out of it onto a system face
that happens to have the codepoint.

every_icon_is_in_the_bundled_font maps each constant through the shipped
font's charmap, which is the guard the script's "the two lists have to
agree" comment asks for. FontDiagnostics gains icon_family, so a build
whose font failed to register says so instead of drawing tofu; the
emulator reports icons=Some("Symbols Nerd Font Mono").

widget/mark.rs is deleted. It drew one correct triangle, but every
further icon would have been another rasteriser, and an icon as text
takes the size, colour and baseline of the line it sits in for free.

Looked at rather than only compiled: closed and open marks in
run-headless.sh phone --phone either side of a tap, and the collapse
bar's up mark under IRIS_TOOLS_EXPANDED=1.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

AGENTS.md

@@ -72,9 +72,10 @@ Module-by-module intent is in `docs/PLAN.md`'s "Backend layout".
     conditions. Read it before touching anything under that branch.
   - `docs/IRIS.md`, `docs/IRIS_TODO.md`, `docs/DECISIONS.md`,
     `docs/LAYOUT.md`, `docs/TEXTURES.md`, `docs/CLIENT_CORE.md` — iris's
-    own public API log, working list, decisions log, layout/render design,
-    and texture-atlas design, and the client-core crate's design,
-    respectively.
+    own build log (**any major addition or design decision, not only
+    public API** -- Iris, 2026-09-08), working list, decisions log,
+    layout/render design, and texture-atlas design, and the client-core
+    crate's design, respectively.
 - `.dev-updater.ron` — what Dev Updater builds here: the server (run as
   `service: Managed(…)`, supervised by Dev Updater's own implementation
   rather than a script kept here) and the APK, in parallel. It points at
@@ -100,6 +101,18 @@ the **Mono** face, where every glyph is one em square, which is what makes
 two icon buttons the same width without either being given one — and why
 `GLYPH_SIZE` is smaller than it looks like it should be.
 
+**The Rust app does the same, from its own subset**:
+`iris/core/build-icon-font.sh` -> `iris/core/assets/fonts/nerd_icons.ttf`,
+with the codepoints named in `iris/core/src/icon.rs` and drawn as text
+with `Family::Icons`. Same rule about the two lists agreeing (there is a
+test, `every_icon_is_in_the_bundled_font`), same Mono face, same Material
+Design family so an icon means the same thing in both apps. Its subset is
+separate rather than shared because subsetting only what one app draws is
+the point. This is the **only** font iris bundles — body and monospace
+text come from the platform (docs/DECISIONS.md, 2026-09-07), and an icon
+is the opposite case: a small closed set of codepoints no system font is
+guaranteed to have.
+
 ## Checking your work
 
 - **Server**: `./run-tests.sh` from the repo root (or `cargo test` from
@@ -297,6 +310,17 @@ Each exists because something was invisible without it.
   into it. The emulator is for JNI, the IME, insets, the surface
   lifecycle and one verification run before a build goes to the phone --
   not for iterating on layout.
+- **The emulator is a GLES rig, deliberately** (Iris, 2026-09-08;
+  docs/DECISIONS.md). Its guest has no hardware Vulkan -- only SwiftShader
+  in software -- while its GLES *is* the host's real GPU through virgl at
+  ES 3.1, so an ordinary build's runtime fallback lands there by itself
+  and nothing should pass `force-gles` to arrange it. The Vulkan path is
+  verified on the desktop build and on Iris's phone. Do not boot the
+  emulator with SwiftShader Vulkan to "test the Vulkan path": that
+  measures a software rasteriser and steers iris away from the one
+  hardware-accelerated backend it has there. Every run says which adapter
+  drew it (`iris renderer:` in logcat, printed by `run-bench.sh`); read
+  that line before reading a number.
 
 ### Driving the UI
 

android-shell/Cargo.lock

@@ -82,7 +82,9 @@ checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
 name = "client-core"
 version = "0.1.0"
 dependencies = [
+ "base64",
  "event-model",
+ "log",
  "pulldown-cmark",
  "serde",
  "serde_json",

app/ui-sandbox.sh

@@ -404,12 +404,27 @@ done
 # Percent-encoded because the app URL-decodes the deep link's query: a
 # token with '+' in it enrols as one with a space, and nothing reports it.
 enc=$(python3 -c 'import sys, urllib.parse; print(urllib.parse.quote(sys.argv[1], safe=""))' "$TOKEN")
+# The CA rides in the link (`wg_app_link::enroll::ca_param`: base64url of
+# the DER, which needs no percent-encoding). The Compose app ignores it and
+# pins the copy its APK was built with; the iris app has no baked copy at
+# all -- it is cross-compiled and could be pointed at any machine -- so
+# without this it enrols and then trusts nothing. Minted here rather than by
+# `--enroll-link` because this token is the sandbox's own, carried across
+# restarts so the emulator stays enrolled (see the top of this file).
+ca=$(python3 - "$CERTS/ca.pem" <<'CA'
+import base64, sys
+pem = open(sys.argv[1]).read()
+body = pem.split("-----BEGIN CERTIFICATE-----")[1].split("-----END CERTIFICATE-----")[0]
+der = base64.b64decode("".join(body.split()))
+print(base64.urlsafe_b64encode(der).decode().rstrip("="))
+CA
+)
 cat <<INFO
 sandbox: server $pid on 127.0.0.1:$PORT, log $LOG
 sandbox: 9 invented Claude Code sessions under $PROJECTS (one of them ${BIG_MB}MB)
 
   enrol the emulator (once; it survives sandbox restarts):
-    adb shell "am start -a android.intent.action.VIEW -d 'aiapp://enroll?host=10.0.2.2&port=$PORT&token=$enc'"
+    adb shell "am start -a android.intent.action.VIEW -d 'aiapp://enroll?host=10.0.2.2&port=$PORT&token=$enc&ca=$ca'"
 
   drive it:
     ./ui-sandbox.sh spawn [title]        an echo session; prints its id

client-core/Cargo.lock

@@ -46,7 +46,9 @@ checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
 name = "client-core"
 version = "0.1.0"
 dependencies = [
+ "base64",
  "event-model",
+ "log",
  "pulldown-cmark",
  "serde",
  "serde_json",

client-core/Cargo.toml

@@ -37,6 +37,15 @@ ureq = { version = "3", features = ["json"] }
 # the same parser at the same version, rather than a hand-written splitter
 # that would drift from it.
 pulldown-cmark = "0.13.4"
+# The enrollment link's `ca` parameter is base64url of the CA's DER
+# (`config::parse_link`). Same version `wg-app-link` already pins for the
+# minting half, so a workspace that has both resolves one copy.
+base64 = "0.23"
+# The logging facade only -- `log_ring` implements a `log::Log` backend and
+# wraps whichever real one the platform installed (`android_logger` on the
+# phone, `env_logger` on the desktop), which is why neither of those is a
+# dependency here. See `log_ring`'s module doc.
+log = { version = "0.4.28", features = ["std"] }
 
 
 [dev-dependencies]

client-core/src/config.rs

@@ -6,33 +6,57 @@
 //! the same text a phone would scan as a QR, with no second format
 //! invented for it (RUST.md's E4).
 //!
-//! What this type deliberately does not decide: where it is persisted, and
-//! under what file permissions. A phone seals its token in the Android
-//! Keystore; a desktop client has its own `$XDG_CONFIG_HOME/<app>/`
-//! directory and its own file-mode conventions (MACHINE.md: owner-only,
-//! never in the repo). Both are caller-specific, so they stay out of this
-//! crate per the code rules' "ask for the least you need" -- see
-//! `iris/desktop-app/src/config.rs` for the desktop instance.
+//! [`EnrollmentStore`] persists one of these as JSON, owner-only, in a
+//! directory the caller names -- `$XDG_CONFIG_HOME/ai-app-desktop` for the
+//! desktop app, the app-private files directory on Android. **Which**
+//! directory is the only part left to the platform: the format, the file
+//! mode and the "nothing saved yet is not an error" answer are the same on
+//! both, and were written twice before this.
+//!
+//! JSON rather than the project's usual RON: `wg-app-link`'s RON house
+//! rules (`format`) are for configs a person hand-edits, and this file
+//! never is one -- only the app itself writes or reads it.
 
+use base64::Engine;
 use serde::{Deserialize, Serialize};
+use std::io;
+use std::path::{Path, PathBuf};
 
 /// One enrolled server: reachable at `https://{host}:{port}`, authenticated
-/// with `token` as a bearer header. Does not carry the pinned CA -- that is
-/// a public certificate rather than a secret, and where to find it differs
-/// by caller (a phone pins the one its APK was built against; a desktop
-/// client is told a path).
+/// with `token` as a bearer header.
+///
+/// `ca_pem` is the trust anchor to pin, when the link carried one (the
+/// `ca` parameter, `wg_app_link::enroll::ca_param`). It is optional
+/// because an app built on the machine its server runs on pins the CA at
+/// build time and needs nothing from the link; one built elsewhere -- the
+/// iris Android client is cross-compiled in a VM and run against the
+/// host's server -- has no other way to get it. A public certificate
+/// rather than a secret, so it costs the link nothing but length.
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct EnrolledServer {
     pub host: String,
     pub port: u16,
     pub token: String,
+    /// `#[serde(default)]` so an enrollment saved before this field
+    /// existed still loads, as the enrolled server it always was.
+    #[serde(default)]
+    pub ca_pem: Option<String>,
 }
 
 impl EnrolledServer {
-    /// Parses `aiapp://enroll?host=H&port=P&token=T` (query order does not
-    /// matter; unrecognised keys are ignored). `token` is percent-decoded,
-    /// since `ui-sandbox.sh` encodes it precisely because a raw token can
-    /// contain `+`, which turns into a space if left to a naive splitter.
+    /// Parses `aiapp://enroll?host=H&port=P&token=T[&ca=B]` (query order
+    /// does not matter; unrecognised keys are ignored). `token` is
+    /// percent-decoded, since `ui-sandbox.sh` encodes it precisely because
+    /// a raw token can contain `+`, which turns into a space if left to a
+    /// naive splitter.
+    ///
+    /// `ca` is base64url of the certificate's DER and is rebuilt into PEM
+    /// here, because that is what every consumer of it wants
+    /// (`UreqTransport::new`, and the file a person points `curl --cacert`
+    /// at). A `ca` that does not decode fails the whole link rather than
+    /// enrolling a server with no trust anchor: the link said which
+    /// certificate to pin, and quietly not pinning it is the one outcome
+    /// nothing downstream could notice.
     pub fn parse_link(link: &str) -> Result<Self, String> {
         let query = link.split_once('?').map(|(_, q)| q).ok_or_else(|| {
             format!(
@@ -44,6 +68,7 @@ impl EnrolledServer {
         let mut host = None;
         let mut port = None;
         let mut token = None;
+        let mut ca = None;
         for pair in query.split('&') {
             let Some((key, value)) = pair.split_once('=') else {
                 continue;
@@ -53,6 +78,7 @@ impl EnrolledServer {
                 "host" => host = Some(value),
                 "port" => port = Some(value),
                 "token" => token = Some(value),
+                "ca" => ca = Some(value),
                 _ => {}
             }
         }
@@ -63,8 +89,14 @@ impl EnrolledServer {
             .parse()
             .map_err(|e| format!("'{link}''s port ('{port_str}') is not a number: {e}"))?;
         let token = token.ok_or_else(|| format!("'{link}' is missing 'token'"))?;
+        let ca_pem = ca.map(|ca| pem_from_link_param(&ca)).transpose()?;
 
-        Ok(Self { host, port, token })
+        Ok(Self {
+            host,
+            port,
+            token,
+            ca_pem,
+        })
     }
 
     /// Where a `client_core::api::UreqTransport` reaches this server.
@@ -73,6 +105,80 @@ impl EnrolledServer {
     }
 }
 
+/// The `ca` parameter (base64url of DER, unpadded) as a PEM certificate.
+fn pem_from_link_param(ca: &str) -> Result<String, String> {
+    let der = base64::engine::general_purpose::URL_SAFE_NO_PAD
+        .decode(ca.as_bytes())
+        .map_err(|e| format!("the link's 'ca' is not base64url ({e})"))?;
+    let body = base64::engine::general_purpose::STANDARD.encode(&der);
+    let mut pem = String::from("-----BEGIN CERTIFICATE-----\n");
+    for line in body.as_bytes().chunks(64) {
+        pem.push_str(std::str::from_utf8(line).expect("base64 is ASCII"));
+        pem.push('\n');
+    }
+    pem.push_str("-----END CERTIFICATE-----\n");
+    Ok(pem)
+}
+
+/// Where one client keeps the enrollment it should not have to be told
+/// about a second time. `dir` is the caller's, because that is the only
+/// part that differs by platform -- see this module's doc.
+pub struct EnrollmentStore {
+    dir: PathBuf,
+}
+
+impl EnrollmentStore {
+    pub fn new(dir: impl Into<PathBuf>) -> Self {
+        Self { dir: dir.into() }
+    }
+
+    pub fn dir(&self) -> &Path {
+        &self.dir
+    }
+
+    fn file(&self) -> PathBuf {
+        self.dir.join("enrollment.json")
+    }
+
+    /// Writes `server` under `dir`, creating it if needed, and sets the
+    /// file owner-only -- it carries a bearer token, the same reason
+    /// `server/`'s own token store is 0600.
+    pub fn save(&self, server: &EnrolledServer) -> io::Result<()> {
+        std::fs::create_dir_all(&self.dir)?;
+        let path = self.file();
+        let json = serde_json::to_vec_pretty(server)
+            .expect("EnrolledServer holds nothing that fails to serialise");
+        std::fs::write(&path, json)?;
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600))?;
+        }
+        Ok(())
+    }
+
+    /// `Ok(None)` when nothing has been enrolled yet, rather than an error
+    /// -- "not enrolled" is an ordinary first-run state, not a failure
+    /// (UI_RULES' "a deliberate choice is not a problem to report" applies
+    /// just as well to a file that simply hasn't been written yet).
+    pub fn load(&self) -> io::Result<Option<EnrolledServer>> {
+        let path = self.file();
+        match std::fs::read(&path) {
+            Ok(bytes) => {
+                let server = serde_json::from_slice(&bytes).map_err(|e| {
+                    io::Error::new(
+                        io::ErrorKind::InvalidData,
+                        format!("{} is not a valid enrollment ({e})", path.display()),
+                    )
+                })?;
+                Ok(Some(server))
+            }
+            Err(e) if e.kind() == io::ErrorKind::NotFound => Ok(None),
+            Err(e) => Err(e),
+        }
+    }
+}
+
 fn percent_decode(s: &str) -> String {
     let bytes = s.as_bytes();
     let mut out = Vec::with_capacity(bytes.len());
@@ -108,6 +214,7 @@ mod tests {
                 host: "127.0.0.1".to_string(),
                 port: 8547,
                 token: "abcDEF123".to_string(),
+                ca_pem: None,
             }
         );
         assert_eq!(server.base_url(), "https://127.0.0.1:8547");
@@ -141,6 +248,115 @@ mod tests {
         );
     }
 
+    /// The CA travels as base64url of the DER and comes back out as the
+    /// PEM every consumer of it wants -- the same round trip
+    /// `wg_app_link::enroll::ca_param` mints.
+    #[test]
+    fn a_ca_in_the_link_comes_back_as_pem() {
+        let der = [0x30u8, 0x82, 0x01, 0xfb, 0x3e, 0x7f];
+        let param = base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(der);
+        let server =
+            EnrolledServer::parse_link(&format!("aiapp://enroll?host=h&port=1&token=t&ca={param}"))
+                .unwrap();
+        let pem = server.ca_pem.expect("the link carried a CA");
+        assert!(pem.starts_with("-----BEGIN CERTIFICATE-----\n"), "{pem}");
+        assert!(
+            pem.trim_end().ends_with("-----END CERTIFICATE-----"),
+            "{pem}"
+        );
+        assert_eq!(
+            base64::engine::general_purpose::STANDARD
+                .decode(
+                    pem.lines()
+                        .filter(|l| !l.starts_with("-----"))
+                        .collect::<String>()
+                )
+                .unwrap(),
+            der
+        );
+    }
+
+    /// A link with no `ca` is an ordinary link, not a broken one: an app
+    /// that pins at build time mints and reads exactly these.
+    #[test]
+    fn no_ca_parameter_is_none_not_an_error() {
+        let server = EnrolledServer::parse_link("aiapp://enroll?host=h&port=1&token=t").unwrap();
+        assert_eq!(server.ca_pem, None);
+    }
+
+    /// The half that cannot be noticed later: a `ca` that does not decode
+    /// must fail the link rather than enrolling with nothing pinned.
+    #[test]
+    fn a_ca_that_does_not_decode_fails_the_link() {
+        let err =
+            EnrolledServer::parse_link("aiapp://enroll?host=h&port=1&token=t&ca=not!base64url")
+                .unwrap_err();
+        assert!(err.contains("ca"), "{err}");
+    }
+
+    #[test]
+    fn a_saved_enrollment_reads_back_the_same() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = EnrollmentStore::new(dir.path());
+        let server = EnrolledServer {
+            host: "127.0.0.1".to_string(),
+            port: 8547,
+            token: "tok".to_string(),
+            ca_pem: Some("-----BEGIN CERTIFICATE-----\nQUJD\n-----END CERTIFICATE-----\n".into()),
+        };
+        store.save(&server).unwrap();
+        assert_eq!(store.load().unwrap(), Some(server));
+    }
+
+    #[test]
+    fn nothing_saved_yet_is_none_not_an_error() {
+        let dir = tempfile::tempdir().unwrap();
+        assert_eq!(EnrollmentStore::new(dir.path()).load().unwrap(), None);
+    }
+
+    /// An enrollment written before `ca_pem` existed still loads.
+    #[test]
+    fn an_enrollment_without_a_ca_still_loads() {
+        let dir = tempfile::tempdir().unwrap();
+        let store = EnrollmentStore::new(dir.path());
+        std::fs::create_dir_all(dir.path()).unwrap();
+        std::fs::write(
+            dir.path().join("enrollment.json"),
+            br#"{"host":"h","port":1,"token":"t"}"#,
+        )
+        .unwrap();
+        assert_eq!(store.load().unwrap().unwrap().ca_pem, None);
+    }
+
+    #[test]
+    #[cfg(unix)]
+    fn the_saved_file_is_owner_only() {
+        use std::os::unix::fs::PermissionsExt;
+        let dir = tempfile::tempdir().unwrap();
+        let store = EnrollmentStore::new(dir.path());
+        store
+            .save(&EnrolledServer {
+                host: "h".to_string(),
+                port: 1,
+                token: "t".to_string(),
+                ca_pem: None,
+            })
+            .unwrap();
+        let mode = std::fs::metadata(dir.path().join("enrollment.json"))
+            .unwrap()
+            .permissions()
+            .mode();
+        assert_eq!(mode & 0o777, 0o600);
+    }
+
+    #[test]
+    fn a_corrupt_file_is_named_in_the_error() {
+        let dir = tempfile::tempdir().unwrap();
+        std::fs::write(dir.path().join("enrollment.json"), b"not json").unwrap();
+        let err = EnrollmentStore::new(dir.path()).load().unwrap_err();
+        assert!(err.to_string().contains("enrollment.json"));
+    }
+
     #[test]
     fn a_non_numeric_port_is_named_in_the_error() {
         let err = EnrolledServer::parse_link("aiapp://enroll?host=h&port=x&token=t").unwrap_err();

client-core/src/lib.rs

@@ -8,6 +8,7 @@ pub mod config;
 pub mod durations;
 pub mod event_stream;
 pub mod highlight;
+pub mod log_ring;
 pub mod markdown_blocks;
 pub mod notifications;
 pub mod sse;

client-core/src/log_ring.rs

@@ -0,0 +1,812 @@
+//! The app's own recent log, held in memory so it can be read back
+//! without `logcat`.
+//!
+//! **Why this exists**: Iris tests iris builds on a GrapheneOS phone with
+//! no `adb`, and Android forbids one app reading another's logcat, so
+//! nothing outside the process can recover what it wrote. The only way a
+//! line reaches her is for the app to carry its own copy. This is that
+//! copy: a bounded ring every `log::info!` in the process lands in, on top
+//! of whichever platform logger was already installed (`android_logger`,
+//! `env_logger`) rather than instead of it -- see [`RingLogger`].
+//!
+//! Two consumers, both reading the same ring rather than each keeping
+//! their own: the bench app's `Copy report`/`Diagnostics` (which reads
+//! [`LogRing::tail_text`] and [`LogRing::summary`]) and whatever hands the
+//! log out of the process -- on Android, the `DevLogProvider` Dev Updater
+//! queries, which reads [`LogRing::since`] and [`LogRing::newest_seq`].
+//! That is why reading does not consume: a line already handed over must
+//! still be in the report, and a report taken twice must say the same
+//! thing.
+
+use std::collections::VecDeque;
+use std::sync::{Arc, Mutex, OnceLock};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+/// How many lines a default ring holds, and how many bytes of message.
+///
+/// Both bounds apply -- whichever bites first -- because the two failure
+/// modes are different: a flood of short lines exhausts the count, and one
+/// pathological line (a stack trace, a pretty-printed JSON body) exhausts
+/// the bytes. A ring bounded only by lines can hold megabytes; one bounded
+/// only by bytes can be emptied by a single line.
+pub const DEFAULT_MAX_LINES: usize = 2000;
+pub const DEFAULT_MAX_BYTES: usize = 256 * 1024;
+
+/// How many of the ring's newest lines [`LogRing::tail_text`] includes.
+/// Sized for a phone's share sheet rather than for the ring itself: 150
+/// lines of `HH:MM:SS.mmm LEVEL target: message` is a few KiB, comfortably
+/// short of whatever made pasting the full (up to 2000-line) ring into a
+/// chat's message box laggy on Iris's phone. The full ring is still
+/// reachable through `devlog`'s provider, so this only bounds what a
+/// report inlines.
+pub const COPY_REPORT_TAIL_LINES: usize = 150;
+
+/// One recorded line. `seq` is assigned by the ring and only ever
+/// increases, so a reader that remembers where it got to can ask for what
+/// came after -- and a gap in the sequence is exactly the lines the bound
+/// dropped.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct LogLine {
+    pub seq: u64,
+    /// Milliseconds since the unix epoch, from the app's own clock. The
+    /// app's rather than the receiver's: a line is timestamped when it
+    /// happened, and an upload can be minutes later or never.
+    pub at_ms: u64,
+    pub level: log::Level,
+    pub target: String,
+    pub message: String,
+}
+
+impl LogLine {
+    /// Roughly what the line costs the ring. The two `String`s dominate;
+    /// the fixed fields are counted as a flat overhead so a ring of empty
+    /// messages still has a bound.
+    fn weight(&self) -> usize {
+        self.target.len() + self.message.len() + 32
+    }
+
+    /// `12:34:56.789 INFO  iris::android: the message`, the shape a
+    /// person skims. Time of day only -- the date is in the report's own
+    /// header, and a ring never spans one.
+    pub fn format(&self) -> String {
+        format!(
+            "{} {:<5} {}: {}",
+            clock_time(self.at_ms),
+            self.level,
+            self.target,
+            self.message
+        )
+    }
+}
+
+/// `HH:MM:SS.mmm` in UTC from a unix millisecond count, without a date
+/// library: the only field this needs is the time of day, and dividing out
+/// the day is the whole calculation. Deliberately not local time -- the
+/// phone's offset is not knowable here, and a report that says UTC is
+/// comparable with the server's log, which is what it gets read against.
+fn clock_time(at_ms: u64) -> String {
+    let ms = at_ms % 1000;
+    let secs_of_day = (at_ms / 1000) % 86_400;
+    format!(
+        "{:02}:{:02}:{:02}.{:03}",
+        secs_of_day / 3600,
+        (secs_of_day % 3600) / 60,
+        secs_of_day % 60,
+        ms
+    )
+}
+
+/// Now, in unix milliseconds. Saturating rather than panicking on a clock
+/// before the epoch: a wrong timestamp in a diagnostic is not worth taking
+/// the app down for.
+pub fn now_ms() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_millis() as u64)
+        .unwrap_or(0)
+}
+
+#[derive(Debug)]
+struct Inner {
+    lines: VecDeque<LogLine>,
+    bytes: usize,
+    max_lines: usize,
+    max_bytes: usize,
+    next_seq: u64,
+    /// How many lines the bounds have discarded since the ring was made.
+    /// Reported rather than inferred, so "the log starts here" and "the
+    /// log was cut off here" are distinguishable -- the unknown state the
+    /// UI rules ask for.
+    dropped: u64,
+}
+
+/// A bounded, shareable ring of recent log lines. Cloning shares the ring;
+/// there is one per process and every holder sees the same lines.
+#[derive(Debug, Clone)]
+pub struct LogRing(Arc<Mutex<Inner>>);
+
+impl LogRing {
+    pub fn new(max_lines: usize, max_bytes: usize) -> Self {
+        assert!(
+            max_lines > 0 && max_bytes > 0,
+            "a ring with no room holds nothing"
+        );
+        Self(Arc::new(Mutex::new(Inner {
+            lines: VecDeque::new(),
+            bytes: 0,
+            max_lines,
+            max_bytes,
+            next_seq: 0,
+            dropped: 0,
+        })))
+    }
+
+    /// The bounds this project ships with: [`DEFAULT_MAX_LINES`] and
+    /// [`DEFAULT_MAX_BYTES`].
+    pub fn with_defaults() -> Self {
+        Self::new(DEFAULT_MAX_LINES, DEFAULT_MAX_BYTES)
+    }
+
+    /// A poisoned lock is a bug in a panicking logger, not a reason to
+    /// take the app down a second time -- the ring is a diagnostic, and
+    /// losing it must not be worse than the fault it was recording.
+    fn with<R>(&self, f: impl FnOnce(&mut Inner) -> R) -> R {
+        let mut guard = match self.0.lock() {
+            Ok(guard) => guard,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        f(&mut guard)
+    }
+
+    /// Records a line, evicting the oldest until both bounds hold again.
+    pub fn push(&self, level: log::Level, target: &str, message: String) {
+        self.with(|inner| {
+            let line = LogLine {
+                seq: inner.next_seq,
+                at_ms: now_ms(),
+                level,
+                target: target.to_string(),
+                message,
+            };
+            inner.next_seq += 1;
+            inner.bytes += line.weight();
+            inner.lines.push_back(line);
+            // `!is_empty()` rather than `len() > 1`: one line larger than
+            // the whole byte bound is kept, because dropping it would
+            // leave the ring silently empty while lines were arriving.
+            while inner.lines.len() > inner.max_lines
+                || (inner.bytes > inner.max_bytes && inner.lines.len() > 1)
+            {
+                if let Some(evicted) = inner.lines.pop_front() {
+                    inner.bytes -= evicted.weight();
+                    inner.dropped += 1;
+                }
+            }
+        })
+    }
+
+    /// Every line held, oldest first.
+    pub fn snapshot(&self) -> Vec<LogLine> {
+        self.with(|inner| inner.lines.iter().cloned().collect())
+    }
+
+    /// The lines with a sequence number at or after `seq`, oldest first,
+    /// and the sequence to ask from next time. Does not consume: see this
+    /// module's doc for why.
+    pub fn since(&self, seq: u64) -> (Vec<LogLine>, u64) {
+        self.with(|inner| {
+            let lines: Vec<LogLine> = inner
+                .lines
+                .iter()
+                .filter(|line| line.seq >= seq)
+                .cloned()
+                .collect();
+            let next = lines.last().map(|line| line.seq + 1).unwrap_or(seq);
+            (lines, next)
+        })
+    }
+
+    pub fn len(&self) -> usize {
+        self.with(|inner| inner.lines.len())
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+
+    pub fn dropped(&self) -> u64 {
+        self.with(|inner| inner.dropped)
+    }
+
+    /// The sequence number of the newest line held, or `None` for a ring
+    /// nothing has been written to.
+    ///
+    /// What a reader needs to notice that this process **restarted**: the
+    /// ring is in memory, so a new process starts again at zero, and a
+    /// reader holding a cursor from the previous one would otherwise ask
+    /// for lines after a number nothing will reach for hours and see
+    /// nothing at all -- silently, which is worse than seeing the log
+    /// begin again. Answering `None` rather than 0 for an empty ring is
+    /// the same distinction [`Self::summary`] draws: "nothing has been
+    /// logged" is not a sequence number.
+    pub fn newest_seq(&self) -> Option<u64> {
+        self.with(|inner| inner.lines.back().map(|line| line.seq))
+    }
+
+    /// When the newest line was written, in unix milliseconds, or `None`
+    /// for a ring nothing has been written to.
+    pub fn last_at_ms(&self) -> Option<u64> {
+        self.with(|inner| inner.lines.back().map(|line| line.at_ms))
+    }
+
+    /// Every line held, formatted one per line -- what `Copy report`
+    /// appends.
+    pub fn to_text(&self) -> String {
+        self.snapshot()
+            .iter()
+            .map(LogLine::format)
+            .collect::<Vec<_>>()
+            .join("\n")
+    }
+
+    /// The newest `max_lines` lines, formatted, with a first line naming
+    /// how many older ones were left out of *this* text when the ring held
+    /// more than that -- what `Copy report` appends instead of
+    /// [`Self::to_text`].
+    ///
+    /// Iris's own report: pasting the full ring (over a thousand lines on
+    /// a session that ran with tracing on) into a phone's message box was
+    /// what "causes a lot of lag" meant (docs/IRIS_TODO.md, 2026-09-07
+    /// night) -- nothing is actually lost, since `devlog`'s provider still
+    /// hands Dev Updater's Runtime tab the whole ring; this only caps what
+    /// gets inlined into a share.
+    pub fn tail_text(&self, max_lines: usize) -> String {
+        let lines = self.snapshot();
+        if lines.len() <= max_lines {
+            return lines
+                .iter()
+                .map(LogLine::format)
+                .collect::<Vec<_>>()
+                .join("\n");
+        }
+        let omitted = lines.len() - max_lines;
+        let tail = lines[lines.len() - max_lines..]
+            .iter()
+            .map(LogLine::format)
+            .collect::<Vec<_>>()
+            .join("\n");
+        format!("{omitted} earlier lines omitted; full log in Dev Updater's Runtime tab\n{tail}")
+    }
+
+    /// The newest `max_lines` lines, formatted, or `None` if the ring is
+    /// locked at this instant.
+    ///
+    /// For the one caller that must not block: **the panic hook**. A panic
+    /// raised while this ring's own lock was held -- an allocation failing
+    /// inside [`Self::push`], an assertion in a `log::Log` on the way here
+    /// -- would deadlock the hook against the thread that is panicking,
+    /// and the process would hang instead of aborting, with nothing
+    /// written anywhere. Losing the context lines is the right trade
+    /// against that, and `None` says which happened rather than looking
+    /// like an empty log.
+    pub fn try_tail_text(&self, max_lines: usize) -> Option<String> {
+        let guard = match self.0.try_lock() {
+            Ok(guard) => guard,
+            // A poisoned lock is uncontended, so its contents are still
+            // readable -- the same judgement as `with`.
+            Err(std::sync::TryLockError::Poisoned(poisoned)) => poisoned.into_inner(),
+            Err(std::sync::TryLockError::WouldBlock) => return None,
+        };
+        let lines = &guard.lines;
+        let from = lines.len().saturating_sub(max_lines);
+        Some(
+            lines
+                .iter()
+                .skip(from)
+                .map(LogLine::format)
+                .collect::<Vec<_>>()
+                .join("\n"),
+        )
+    }
+
+    /// One line for a diagnostics pane: how much is held, how much was
+    /// dropped, and when the last line arrived. "no lines yet" is its own
+    /// wording rather than a count of zero with a made-up time, because
+    /// "nothing has been logged" and "logging is not running" would
+    /// otherwise look the same.
+    pub fn summary(&self) -> String {
+        let (len, dropped, last) = self.with(|inner| {
+            (
+                inner.lines.len(),
+                inner.dropped,
+                inner.lines.back().map(|line| line.at_ms),
+            )
+        });
+        match last {
+            None => "app log: no lines yet".to_string(),
+            Some(at) => {
+                let dropped = if dropped > 0 {
+                    format!(", {dropped} dropped")
+                } else {
+                    String::new()
+                };
+                format!(
+                    "app log: {len} lines held{dropped}, last {}",
+                    clock_time(at)
+                )
+            }
+        }
+    }
+}
+
+/// Whether a target belongs to this app's own crates (`iris` or
+/// `client_core`) rather than a dependency's -- `starts_with` guarded by an
+/// exact match or a `::` so an unrelated crate that merely begins with the
+/// same letters (there is no such crate today, but the check should not
+/// rely on that) is never mistaken for one of ours.
+fn is_own_target(target: &str) -> bool {
+    target == "iris"
+        || target.starts_with("iris::")
+        || target == "client_core"
+        || target.starts_with("client_core::")
+}
+
+/// Whether a line at `level` from `target` belongs in the ring, given
+/// whether tracing is on right now.
+///
+/// This is the one filter docs/IRIS_TODO.md's "logs way too big" entry
+/// asked for, applied once here rather than at each `debug!` call site:
+/// Info and above always ring, from anything, because a real warning or
+/// error from a dependency is worth keeping. Debug and Trace ring only
+/// from this app's own targets, and only while tracing is switched on --
+/// otherwise `naga::front`/`wgpu_core`/`jni` log at Debug unconditionally
+/// (the process logger's own level, set once at install and unrelated to
+/// tracing), which is what filled the ring with 1339 lines of it and
+/// dropped 4050 more before this existed. `iris`'s own Debug lines already
+/// self-gate on `iris::diagnostics::trace_enabled` at their call sites
+/// (commit 992c472); this is the backstop for lines this crate does not
+/// control.
+fn ring_accepts(level: log::Level, target: &str, trace_enabled: bool) -> bool {
+    level <= log::Level::Info || (trace_enabled && is_own_target(target))
+}
+
+/// A `log` backend that records into a [`LogRing`] **and** forwards to the
+/// logger the platform already installs, so nothing that reads the
+/// platform's log (`logcat`, a terminal) changes.
+///
+/// The inner logger is passed in rather than chosen here: `client-core`
+/// has no business depending on `android_logger` or `env_logger`, and
+/// which one is right is exactly what differs between the two platforms
+/// (the sharing rule in AGENTS.md).
+pub struct RingLogger {
+    ring: LogRing,
+    inner: Box<dyn log::Log>,
+    /// Whether `iris::input`/`iris::frame`-style tracing is switched on
+    /// right now, consulted by [`ring_accepts`]. A plain fn pointer rather
+    /// than a dependency on `iris::diagnostics::trace_enabled` directly:
+    /// `client-core` sits below `iris` (AGENTS.md's "dependencies flow one
+    /// direction"), so the platform crate that depends on both is the one
+    /// that wires this closure through, the same way it already supplies
+    /// `inner`.
+    trace_enabled: fn() -> bool,
+}
+
+impl RingLogger {
+    pub fn new(ring: LogRing, inner: Box<dyn log::Log>, trace_enabled: fn() -> bool) -> Self {
+        Self {
+            ring,
+            inner,
+            trace_enabled,
+        }
+    }
+}
+
+impl log::Log for RingLogger {
+    /// True for anything `log`'s own max level lets through: the ring
+    /// wants everything the *inner* logger might also want, even where the
+    /// platform logger would filter it out. Which lines the ring itself
+    /// keeps is decided in [`Self::log`] by [`ring_accepts`].
+    fn enabled(&self, _metadata: &log::Metadata) -> bool {
+        true
+    }
+
+    fn log(&self, record: &log::Record) {
+        if ring_accepts(record.level(), record.target(), (self.trace_enabled)()) {
+            self.ring
+                .push(record.level(), record.target(), record.args().to_string());
+        }
+        if self.inner.enabled(record.metadata()) {
+            self.inner.log(record);
+        }
+    }
+
+    fn flush(&self) {
+        self.inner.flush();
+    }
+}
+
+/// Installs a [`RingLogger`] as the process logger and answers the ring it
+/// records into.
+///
+/// Fails only if a logger is already installed, which is a programmer
+/// error (two initialisation paths) rather than a recoverable condition --
+/// the caller is named in the error so it is findable.
+pub fn install(
+    ring: LogRing,
+    inner: Box<dyn log::Log>,
+    max_level: log::LevelFilter,
+    trace_enabled: fn() -> bool,
+) -> Result<(), log::SetLoggerError> {
+    log::set_boxed_logger(Box::new(RingLogger::new(ring, inner, trace_enabled)))?;
+    log::set_max_level(max_level);
+    Ok(())
+}
+
+/// The one ring this process records into.
+///
+/// **A deliberate process-global, where this project's rules otherwise say
+/// pass context explicitly.** What is being modelled is already one: `log`
+/// has exactly one backend per process, set once, and every `log::info!`
+/// anywhere in the binary goes to it. A ring handed around as a parameter
+/// would be a *second* answer to "which lines exist" -- the report would
+/// show one ring while the logger filled another, and which one a caller
+/// got would depend on how far down the call tree it was. The tests above
+/// all use their own [`LogRing`], so nothing here needs this to be
+/// testable.
+static PROCESS_RING: OnceLock<LogRing> = OnceLock::new();
+
+/// The process's ring, created on first use with the default bounds.
+/// Safe to call before [`install_process_logger`] -- it will simply be
+/// empty.
+pub fn process_ring() -> &'static LogRing {
+    PROCESS_RING.get_or_init(LogRing::with_defaults)
+}
+
+/// Installs [`process_ring`] as the recording half of the process logger,
+/// forwarding to `inner` (the platform's own logger, already configured).
+/// The platform half of AGENTS.md's sharing rule is `inner`; everything
+/// else is shared. `trace_enabled` is the platform's own trace toggle
+/// (`iris::diagnostics::trace_enabled` on Android) -- see
+/// [`ring_accepts`] and the field doc on `RingLogger` for why it is
+/// passed in rather than called directly.
+pub fn install_process_logger(
+    inner: Box<dyn log::Log>,
+    max_level: log::LevelFilter,
+    trace_enabled: fn() -> bool,
+) -> Result<(), log::SetLoggerError> {
+    install(process_ring().clone(), inner, max_level, trace_enabled)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use log::Level;
+
+    fn fill(ring: &LogRing, count: usize) {
+        for n in 0..count {
+            ring.push(Level::Info, "test", format!("line {n}"));
+        }
+    }
+
+    #[test]
+    fn lines_come_back_oldest_first() {
+        let ring = LogRing::new(10, 1 << 20);
+        fill(&ring, 3);
+        let text: Vec<String> = ring.snapshot().into_iter().map(|l| l.message).collect();
+        assert_eq!(text, ["line 0", "line 1", "line 2"]);
+    }
+
+    #[test]
+    fn the_line_bound_drops_the_oldest_and_says_how_many() {
+        let ring = LogRing::new(3, 1 << 20);
+        fill(&ring, 5);
+        let text: Vec<String> = ring.snapshot().into_iter().map(|l| l.message).collect();
+        assert_eq!(text, ["line 2", "line 3", "line 4"], "the newest survive");
+        assert_eq!(ring.len(), 3);
+        assert_eq!(ring.dropped(), 2, "and the loss is reported, not silent");
+    }
+
+    #[test]
+    fn the_byte_bound_bites_before_the_line_bound_when_lines_are_large() {
+        // Room for 1000 lines but only a few hundred bytes.
+        let ring = LogRing::new(1000, 300);
+        for n in 0..10 {
+            ring.push(Level::Info, "t", format!("{n}{}", "x".repeat(100)));
+        }
+        assert!(
+            ring.len() < 10,
+            "the byte bound evicted: {} held",
+            ring.len()
+        );
+        assert!(ring.dropped() > 0);
+        assert!(
+            ring.snapshot().last().unwrap().message.starts_with('9'),
+            "and it evicted from the old end"
+        );
+    }
+
+    /// The case the `len() > 1` guard exists for: one line larger than the
+    /// whole bound must still be readable, or a ring that is over budget
+    /// reads as a ring nothing was written to.
+    #[test]
+    fn one_oversized_line_is_kept_rather_than_leaving_the_ring_empty() {
+        let ring = LogRing::new(100, 64);
+        ring.push(Level::Error, "t", "y".repeat(5000));
+        assert_eq!(ring.len(), 1);
+        assert_eq!(ring.dropped(), 0);
+    }
+
+    #[test]
+    fn sequence_numbers_only_increase_and_survive_eviction() {
+        let ring = LogRing::new(2, 1 << 20);
+        fill(&ring, 5);
+        let seqs: Vec<u64> = ring.snapshot().into_iter().map(|l| l.seq).collect();
+        assert_eq!(seqs, [3, 4], "a gap is exactly what was dropped");
+    }
+
+    #[test]
+    fn since_returns_only_what_is_new_and_the_next_cursor() {
+        let ring = LogRing::new(100, 1 << 20);
+        fill(&ring, 3);
+        let (first, cursor) = ring.since(0);
+        assert_eq!(first.len(), 3);
+        assert_eq!(cursor, 3);
+
+        let (none, cursor) = ring.since(cursor);
+        assert!(none.is_empty(), "nothing new yet");
+        assert_eq!(cursor, 3, "and the cursor does not move");
+
+        ring.push(Level::Warn, "test", "later".into());
+        let (more, cursor) = ring.since(cursor);
+        assert_eq!(more.len(), 1);
+        assert_eq!(more[0].message, "later");
+        assert_eq!(cursor, 4);
+    }
+
+    /// The restart signal: a reader that saw sequence 4 and is now told
+    /// the newest is 0 knows the process is not the one it was reading.
+    #[test]
+    fn the_newest_sequence_says_where_the_ring_is_and_nothing_for_an_empty_one() {
+        let ring = LogRing::new(100, 1 << 20);
+        assert_eq!(ring.newest_seq(), None, "an empty ring has no newest line");
+        fill(&ring, 5);
+        assert_eq!(ring.newest_seq(), Some(4));
+
+        let restarted = LogRing::new(100, 1 << 20);
+        fill(&restarted, 1);
+        assert_eq!(
+            restarted.newest_seq(),
+            Some(0),
+            "a fresh ring starts again, which is exactly what a reader has to notice"
+        );
+    }
+
+    #[test]
+    fn reading_does_not_consume() {
+        let ring = LogRing::new(100, 1 << 20);
+        fill(&ring, 2);
+        let (sent, _) = ring.since(0);
+        assert_eq!(sent.len(), 2);
+        assert_eq!(ring.len(), 2, "the report still has them after an upload");
+        assert_eq!(ring.to_text().lines().count(), 2);
+    }
+
+    #[test]
+    fn tail_text_is_the_whole_ring_untouched_when_under_the_cap() {
+        let ring = LogRing::new(100, 1 << 20);
+        fill(&ring, 5);
+        assert_eq!(ring.tail_text(150), ring.to_text());
+    }
+
+    #[test]
+    fn tail_text_trims_to_the_newest_lines_and_says_how_many_were_left_out() {
+        let ring = LogRing::new(1000, 1 << 20);
+        fill(&ring, 200);
+        let tail = ring.tail_text(150);
+        let mut lines = tail.lines();
+        assert_eq!(
+            lines.next().unwrap(),
+            "50 earlier lines omitted; full log in Dev Updater's Runtime tab"
+        );
+        let rest: Vec<&str> = lines.collect();
+        assert_eq!(rest.len(), 150, "exactly the cap, after the header line");
+        assert!(
+            rest[0].ends_with("line 50"),
+            "the oldest line kept is the 50th, not line 0: {}",
+            rest[0]
+        );
+        assert!(rest.last().unwrap().ends_with("line 199"));
+    }
+
+    #[test]
+    fn try_tail_text_gives_the_newest_lines_with_no_header() {
+        let ring = LogRing::new(1000, 1 << 20);
+        fill(&ring, 200);
+        let tail = ring.try_tail_text(80).expect("nothing holds the lock");
+        let lines: Vec<&str> = tail.lines().collect();
+        assert_eq!(lines.len(), 80, "the cap, and no header: this is a file");
+        assert!(lines[0].ends_with("line 120"), "{}", lines[0]);
+        assert!(lines[79].ends_with("line 199"), "{}", lines[79]);
+    }
+
+    /// The whole point of the `try_`: the panic hook calls this from a
+    /// thread that may already hold the ring's lock, and a blocking read
+    /// there would hang the process instead of aborting it.
+    #[test]
+    fn try_tail_text_answers_none_rather_than_blocking_on_a_held_lock() {
+        let ring = LogRing::new(10, 1 << 20);
+        fill(&ring, 3);
+        let held = ring.0.lock().expect("fresh ring");
+        assert_eq!(ring.try_tail_text(80), None);
+        drop(held);
+        assert!(ring.try_tail_text(80).is_some());
+    }
+
+    #[test]
+    fn an_empty_ring_says_so_rather_than_reporting_a_time() {
+        let ring = LogRing::with_defaults();
+        assert_eq!(ring.summary(), "app log: no lines yet");
+        assert_eq!(ring.last_at_ms(), None);
+        assert!(ring.is_empty());
+    }
+
+    #[test]
+    fn the_summary_names_dropped_lines_only_when_there_are_some() {
+        let ring = LogRing::new(2, 1 << 20);
+        fill(&ring, 2);
+        assert!(!ring.summary().contains("dropped"), "{}", ring.summary());
+        fill(&ring, 2);
+        assert!(ring.summary().contains("2 dropped"), "{}", ring.summary());
+    }
+
+    #[test]
+    fn a_line_formats_as_time_level_target_message() {
+        let line = LogLine {
+            seq: 0,
+            // 1970-01-01T12:34:56.789Z, so the arithmetic is checkable by
+            // hand rather than against another clock.
+            at_ms: (12 * 3600 + 34 * 60 + 56) * 1000 + 789,
+            level: Level::Info,
+            target: "iris::android".into(),
+            message: "surface created".into(),
+        }
+        .format();
+        assert_eq!(line, "12:34:56.789 INFO  iris::android: surface created");
+    }
+
+    /// The forwarding half: a line reaches the ring *and* the logger the
+    /// platform already had, and one the inner logger filters out is still
+    /// in the ring.
+    #[test]
+    fn the_ring_logger_forwards_to_the_inner_logger() {
+        use log::Log;
+        struct Collect(Arc<Mutex<Vec<String>>>, log::Level);
+        impl Log for Collect {
+            fn enabled(&self, metadata: &log::Metadata) -> bool {
+                metadata.level() <= self.1
+            }
+            fn log(&self, record: &log::Record) {
+                self.0.lock().unwrap().push(record.args().to_string());
+            }
+            fn flush(&self) {}
+        }
+
+        let seen = Arc::new(Mutex::new(Vec::new()));
+        let ring = LogRing::with_defaults();
+        // Own target, tracing on: this is the case where the ring and the
+        // inner logger disagree, which is the thing under test -- a
+        // foreign target is covered separately below.
+        let logger = RingLogger::new(
+            ring.clone(),
+            Box::new(Collect(seen.clone(), Level::Info)),
+            || true,
+        );
+        logger.log(
+            &log::Record::builder()
+                .args(format_args!("kept"))
+                .level(Level::Info)
+                .target("iris::test")
+                .build(),
+        );
+        logger.log(
+            &log::Record::builder()
+                .args(format_args!("filtered"))
+                .level(Level::Debug)
+                .target("iris::test")
+                .build(),
+        );
+
+        assert_eq!(
+            *seen.lock().unwrap(),
+            ["kept"],
+            "the inner logger's own filter still applies"
+        );
+        let held: Vec<String> = ring.snapshot().into_iter().map(|l| l.message).collect();
+        assert_eq!(
+            held,
+            ["kept", "filtered"],
+            "own-target debug still rings while tracing is on"
+        );
+    }
+
+    /// The bug this filter fixes: `naga`/`wgpu_core`/`jni` log at Debug
+    /// unconditionally, and used to flood the ring even though nothing in
+    /// this app asked for their Debug output. A foreign target's Debug
+    /// line must not ring even while tracing is on -- tracing controls
+    /// this app's own diagnostics, not a dependency's chatter.
+    #[test]
+    fn a_foreign_targets_debug_line_never_rings_even_while_tracing_is_on() {
+        use log::Log;
+        struct Discard;
+        impl Log for Discard {
+            fn enabled(&self, _: &log::Metadata) -> bool {
+                true
+            }
+            fn log(&self, _: &log::Record) {}
+            fn flush(&self) {}
+        }
+
+        let ring = LogRing::with_defaults();
+        let logger = RingLogger::new(ring.clone(), Box::new(Discard), || true);
+        logger.log(
+            &log::Record::builder()
+                .args(format_args!("naga debug spam"))
+                .level(Level::Debug)
+                .target("naga::front")
+                .build(),
+        );
+        logger.log(
+            &log::Record::builder()
+                .args(format_args!("naga warning"))
+                .level(Level::Warn)
+                .target("wgpu_core::device")
+                .build(),
+        );
+
+        let held: Vec<String> = ring.snapshot().into_iter().map(|l| l.message).collect();
+        assert_eq!(
+            held,
+            ["naga warning"],
+            "Info-and-above always rings; foreign Debug never does"
+        );
+    }
+
+    #[test]
+    fn ring_accepts_is_own_target_debug_only_while_tracing() {
+        assert!(
+            ring_accepts(Level::Info, "wgpu_core::device", false),
+            "Info+ from anything, tracing off"
+        );
+        assert!(
+            ring_accepts(Level::Warn, "jni", true),
+            "Info+ from anything, tracing on"
+        );
+        assert!(
+            !ring_accepts(Level::Debug, "jni", true),
+            "foreign Debug, tracing on: still excluded"
+        );
+        assert!(
+            !ring_accepts(Level::Debug, "iris::sense", false),
+            "own Debug, tracing off: excluded"
+        );
+        assert!(
+            ring_accepts(Level::Debug, "iris::sense", true),
+            "own Debug, tracing on: included"
+        );
+        assert!(
+            ring_accepts(Level::Trace, "client_core::api", true),
+            "own Trace, tracing on: included"
+        );
+    }
+
+    #[test]
+    fn is_own_target_matches_the_crate_or_its_modules_only() {
+        assert!(is_own_target("iris"));
+        assert!(is_own_target("iris::sense"));
+        assert!(is_own_target("client_core"));
+        assert!(is_own_target("client_core::log_ring"));
+        assert!(!is_own_target("iris_something_else"));
+        assert!(!is_own_target("naga::front"));
+        assert!(!is_own_target("jni"));
+    }
+}

docs/DECISIONS.md

@@ -5,6 +5,214 @@ 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.
 
+## 2026-09-08 (iris ships an icon font, and the drawn mark is deleted)
+
+- **Directed by Iris.** Her question on seeing `widget::mark`: "why does
+  mark exist? The font should be working if it's working for compose and
+  nerd fonts are bundled." It was not: the Compose app draws its icons
+  from **its own committed Nerd Fonts subset**, while iris was setting
+  the disclosure mark with bare Unicode geometric codepoints
+  (U+25B8/25BE/25B4) out of whatever face the platform resolved -- an
+  empty box on her phone, a dot on this VM. The 2026-09-07 entry below,
+  which said "iris had no equivalent icon font to keep", is what left
+  that gap: iris had no icon font because it had never had one, not
+  because it needed none.
+- **So iris now bundles the same kind of subset**:
+  `iris/core/build-icon-font.sh` writes
+  `iris/core/assets/fonts/nerd_icons.ttf` (992 bytes, three Material
+  Design glyphs today), `iris::icon` names the codepoints, and
+  `Family::Icons` draws them. This does **not** reopen the platform-fonts
+  decision: body and monospace text still come from the platform, and an
+  icon is the opposite case -- a small, closed, known set of codepoints,
+  which is exactly the division the Compose app already makes.
+- **`iris::widget::mark` is deleted** (added earlier the same day). It
+  drew a correct triangle, but only a triangle, and every further icon
+  would have been another bespoke rasteriser. An icon as text also takes
+  the size, colour and baseline of the line it sits in for free.
+
+## 2026-09-08 (the emulator is a GLES machine, and Vulkan is verified elsewhere)
+
+- **Directed by Iris, carried out here**: "make sure the setup uses GL for
+  the android emulator and remove any vulkan requirements. That'll be
+  tested through both the desktop version as well as my phone." So the
+  emulator is settled as a GLES rig and nothing chases hardware Vulkan in
+  it any more; the Vulkan path is covered by the desktop build and by her
+  phone.
+- **Nothing had to be forced to make that true.** Measured in the guest
+  the same day: the emulator has no hardware Vulkan at all (its only
+  Vulkan is SwiftShader, in software) and its GLES is the host's real RX
+  7900 XT through virgl at ES 3.1. iris's existing runtime fallback --
+  `Backends::PRIMARY`, no adapter, rebuild on `Backends::GL` -- already
+  lands there, verified end to end with an ordinary (no `force-gles`)
+  debug APK.
+- **The emulator and the phone therefore run the same binary**, differing
+  only in what that binary finds. That is deliberate and worth not
+  undoing: a build flag that changed the backend would mean the thing
+  measured on the emulator is not the thing shipped. `force-gles` stays,
+  but only for pinning the backend on a machine that *does* have Vulkan
+  (the desktop), and never for a phone build.
+- **Every run now says which adapter drew it.** The Android renderer logs
+  the full adapter line at startup the way the desktop already did -- only
+  the backend enum was logged before, which cannot separate `Gl` on the
+  host's GPU from `Gl` on SwiftShader, or a phone's real Vulkan from a
+  software one. `run-bench.sh` prints that line before any number.
+- **No Vulkan requirement was found in iris to remove.** `device_limits()`
+  asks for nothing beyond wgpu's defaults (and zeroes the compute fields),
+  neither backend requires a feature, and both probe rather than
+  `.expect()` an adapter. What was removed was the *documentation* telling
+  people to boot the emulator with SwiftShader Vulkan.
+
+## 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.
+- **Gap found, then closed same day**: this fontique version's Android
+  backend never resolved the `Monospace` generic family at all (confirmed
+  on this checkout's emulator, `mono=None` in the startup diagnostic) --
+  two pre-existing bugs in fontique's own `fonts.xml` parsing stacked (an
+  ordering bug, and a `<family name="monospace">` declaration whose
+  `<font>` children the backend's parser never reads), 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). Checked `linebender/parley`'s `main` branch
+  on GitHub: neither bug is fixed there, so there was no newer release to
+  bump to. Fixed instead in `iris-core` itself
+  (`TextData::patch_android_monospace`, Android-only): reads
+  `/system/etc/fonts.xml`'s own `"monospace"` declaration for the font
+  filename it names, then registers whichever of fontique's actually-
+  scanned families owns that file as the `Monospace` generic -- the same
+  authority Compose's `Typeface.MONOSPACE` resolves through, without
+  pinning an OEM-specific family name. Verified on this checkout's
+  emulator: `mono=Some("Droid Sans Mono")`, and a screenshot showing the
+  bench-fixture's code block and tool-card values in a visibly monospaced
+  face beside sans body text; the desktop `fontconfig` backend is
+  unaffected (still resolves monospace correctly, confirmed unchanged).
+  docs/RUST.md's "Platform fonts (2026-09-07)" has the full account.
+
+## 2026-09-07 (a phone log reaches Iris through Dev Updater's own tab)
+
+**Supersedes the "how a phone log reaches Iris" entry below, same day.**
+Iris's call once the route was working: put it in Dev Updater properly
+rather than smuggling the lines through `ai-server`'s log.
+
+- **The app exposes its own log on the device, and Dev Updater reads it
+  there.** A `ContentProvider` at `<applicationId>.devlog`, one table of
+  lines queried with `?since=<seq>` so a poll is incremental, plus a
+  `status` row (`held`, `dropped`, `newest_seq`). Dev Updater's phone app
+  polls it while the component's **Runtime** tab is open and forwards what
+  is new to its own build machine, into that APK component's runtime log
+  -- so the same tab renders both kinds and the history outlives the
+  phone. No tunnel, no token, no second enrolment: the two apps are on the
+  same phone.
+
+  **It is a contract, not a feature for iris.** Written down in
+  dev-updater's `README.md` ("An app's own log"), so any app that server
+  delivers gets the tab by implementing it; the Compose app in `app/` can
+  do the same later. That is the reason it beat the route below on its
+  second look -- the earlier one only ever worked for the one project that
+  had a server, and put a phone's lines under a *different component* than
+  the one they came from.
+
+- **Read access is `protectionLevel="normal"`, and that is a real trade.**
+  `signature` is what this wants and is not available: Dev Updater and the
+  apps it delivers are built on one machine but signed with different
+  locally generated keys, so a signature permission would be held by
+  nothing at all. What `normal` costs is that any app on that phone which
+  requests `dev.updater.permission.READ_DEVLOG` by name can read another
+  app's dev log. Accepted because these are development builds on a
+  development phone and the alternative was no log; stated in the manifest
+  beside the declaration and in dev-updater's README so it is not
+  rediscovered as a surprise.
+
+- **The provider polls rather than notifying.** `notifyChange` was not
+  implemented: the ring is filled by a `log::Log` backend on whatever
+  thread logged, and giving that a route to a `ContentProvider` means
+  plumbing a callback through `client-core` for every platform. Dev
+  Updater's contract therefore says it polls (about a second, only while
+  the tab is open), which is what keeps implementing the contract cheap --
+  a provider that does notify loses nothing.
+
+- **What was deleted, so there is one mechanism**: `client-core`'s
+  `log_upload` module, `POST /client-log` on `ai-server`, the
+  `AI_APP_LOG_HOST`/`_PORT`/`_TOKEN` baking in `iris/android-app/build.rs`
+  (which left that file with nothing to do, so it is gone too), and the
+  uploader fields on both Android clients. Kept: the ring, `RingLogger`,
+  `install_process_logger`, and the Diagnostics line counting what is
+  held. The upload-status line there is now **"devlog provider:
+  content://<authority>"** -- named from what the provider registered
+  rather than composed from the package here, so a screenshot of that pane
+  is evidence the contract is live and says which package's log it is.
+
+## 2026-09-07 (how a phone log reaches Iris) -- superseded, see above
+
+- **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,
+  and what the entry above went on to build -- the estimate below was
+  right about the work and wrong about it being too much).
+  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) --
+  *gone; the provider above replaced it.* What is worth keeping from it is
+  the reason it went: an APK good only for the server that built it cannot
+  be built in this VM for Iris's phone, which is the case that mattered.
+  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 -- <why>" 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
@@ -442,3 +650,55 @@ marked **DEFERRED** are ones the agent chose not to decide alone.
   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.
+
+## 2026-09-07: the enrolment link carries the CA, so an APK need not be built where its server runs
+
+**Problem.** Every phone build pinned the CA of the machine that compiled
+it -- the Compose app from `GeneratePinnedCert`, the iris app from
+`build.rs` reading `$XDG_CONFIG_HOME/ai-app/certs/ca.pem`. That is fine
+while the two are the same machine and impossible when they are not, which
+is exactly the iris client's situation: cross-compiled in this VM,
+delivered to a phone, run against `ai-server` on the host. Baking the
+host/port/token as well made it worse -- a token in a built artifact.
+
+**Decided: the CA rides in the enrolment link**, as `&ca=<base64url of the
+DER>` (`wg_app_link::enroll::ca_param`), optional and per mint. The app
+that opens the link pins what the link said, and an APK built anywhere
+works against whatever server it is pointed at.
+
+Two alternatives were worked out and rejected.
+
+- **A CA *fingerprint* in the link, pinned at the TLS handshake.** The
+  smallest link (43 more characters) and the strongest shape, but `ureq`
+  3.4 exposes no hook for a custom `rustls` `ServerCertVerifier`: its
+  `TlsConfig` builds the `ClientConfig` itself, so this needs a hand-written
+  `Connector` on the `unversioned` API and `rustls` as a direct dependency
+  of `client-core`. A lot of machinery in the one crate that must stay
+  light.
+- **A fingerprint in the link plus an unauthenticated `GET /ca.pem`.**
+  Small code, but it needs a first connection with verification disabled,
+  and it breaks a documented, tested posture -- `auth.rs`'s "gates every
+  route with zero unauthenticated endpoints", which is a load-bearing
+  decision rather than an implementation detail. Not something to change
+  silently for this.
+
+**What it costs**, measured rather than guessed: on this project's P-256
+CA the link goes from 89 bytes to 652, and `print_enrollment`'s terminal
+QR from 45x23 to 93x47 characters. That is why the parameter is the
+minter's choice per call: `ai-server` passes it (its iris client needs it),
+`dev-updater` passes `None` (its app is built on the machine it talks to,
+and its QR stays scannable in an 80-column terminal). The URI printed under
+the QR is the fallback either way, and is the path Dev Updater's Enroll
+button already uses -- it opens the link with `ACTION_VIEW`, so Android
+offers whichever apps registered the scheme, which needed no change here.
+
+The CA is a public certificate, so putting it in the QR leaks nothing the
+token did not already: photographing the terminal still costs exactly the
+token, which is rotatable.
+
+**The log upload's destination is moot**, so it is not wired to this. On
+the same day Iris decided Dev Updater will read an APK's runtime log from
+an on-device ContentProvider instead, which removes `log_upload`,
+`POST /client-log` and the `AI_APP_LOG_*` baking altogether -- so the
+enrolment landed without touching any of them, for that change to delete
+whole.

docs/IRIS.md

@@ -1,13 +1,249 @@
-# iris: notable public API changes
+# iris: the log of how it is being built
 
-For Iris to read on her own time. Each entry is a change to iris's public
-surface that a widget author or app author would notice: a trait method
-added, removed or re-shaped; a type that callers construct differently; a
-capability that moved. Small and trivial changes do not go here.
+For Iris to read on her own time. An entry is anything **major**: a new
+capability or widget, a design decision and what it was chosen over, a
+mechanism that changed shape, a defect whose root cause says something
+about the framework -- and the public-surface changes a widget or app
+author would notice, which is all this file used to hold (widened on
+Iris's instruction, 2026-09-08: "any major additions or design things
+should be added there, not just public API stuff"). Small and trivial
+things still stay out.
 
 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.
 
+## 2026-09-08: a gesture can be cancelled, and the pointer belongs to the input handler
+
+Two changes to how a drag ends, from defects on Iris's phone (a code
+fence panned sideways made the transcript jump on the next tap, and made
+the fence itself snap back).
+
+**`CursorSense::Cancel`, and `GestureOutcome::Cancelled`.** Taking
+pointer capture cuts every other widget off from the press completely --
+no `PressEnd`, no `Drop` -- so anything else tracking that press was left
+with a gesture open at an origin belonging to a finger long gone, and the
+next touch anywhere was measured from it. A widget that loses a capture
+race is now told, exactly once. It is a separate sense from `Drop`
+deliberately: `Drop` means "your gesture finished" and callers act on it
+(a fling, a tap, a link followed), which is precisely wrong here.
+
+**`CursorSense::drag_senses()`** is what a widget driving a `DragGesture`
+registers -- the frames plus `unclick`, `Drop` and `Cancel`. Both ways a
+gesture can end, stated once rather than remembered per call site;
+forgetting `Drop` is what left a `Scroll` panning from a stale position.
+
+**The pointer's state left `UiRenderState`.** `capture_pointer`,
+`release_pointer` and `captured_pointer` are gone from it. Capture and
+the pressed set are `PointerInput` -- the cursor senses' `Event::Global`,
+a new associated type for state an event owns that belongs to no single
+widget -- held by the event manager that runs the dispatch and reached
+by `&mut`, with no lock anywhere. A handler asks through
+`ctx.data.pointer` (`PointerRequests`: `capture(id)`, `release()`,
+`holder()`).
+
+    // before -- interior mutability on whatever structure was reachable
+    ctx.data.render.capture_pointer(id);
+    // after
+    ctx.data.pointer.capture(id);
+
+`DragGesture::handle` and `Scroll::drag` take `&PointerRequests` where
+they took `&UiRenderState`. `task_on` also lost a `Data: Send` bound it
+never needed -- the future it spawns never sees the event's data, and
+that bound was the whole reason the pointer state had been behind a
+`Mutex`.
+
+## 2026-09-08: `mark(dir, dp, colour)` -- a drawn triangle, and a scroll area's opening edge
+
+**`iris::widget::mark`** draws a filled, antialiased triangle pointing
+along a `Dir`, at a size in dp. It replaces the disclosure codepoints
+U+25B8/25BE/25B4, which were a bet that the platform's fonts have them --
+once iris stopped bundling its own faces, Iris's phone drew an empty box.
+It rasterises one oversampled bitmap into the ordinary texture path and
+scales it into the box asked for, so no new primitive was needed and it
+is correct at any density.
+
+**`scrollable_on` now opens at the beginning of its content, and
+`scrollable_to_end(axis)` is the other one** -- pinned to the end and
+staying there while the content grows, which is what a composer wants and
+what everything did before. A code fence was opening at the end of its
+longest line, in the middle of a word. `Scroll::new` takes the edge as a
+third argument rather than deciding for its caller.
+
+The design point behind that bug is worth more than the bug: `Scroll`
+held its content's length as an `f32` that was `0.0` both for "there is
+nothing here" and for "I have not drawn yet". Those lead somewhere
+different, and the code could not ask which it had -- so the first
+frame's clamp computed a scroll range of zero, read `amt == len` as
+"sitting at the end", and pinned itself there. It is an `Option` now,
+and the clamp declines to answer a question it cannot yet answer. Any
+measurement iris caches from a previous frame has this shape (LAYOUT.md
+section 4's one-frame lag is the general case), so the rule is: give the
+unmeasured state its own value, not a plausible number.
+
+## 2026-09-08: masks have a shape -- `.masked_by(shape)`, and clipping applies to touch
+
+A mask no longer carries a rectangle. It carries **the slot of a
+primitive already drawn**, and the fragment stage evaluates that
+primitive's own coverage at each masked pixel and multiplies it into the
+alpha -- the same rounded-rect SDF the primitive itself is drawn with.
+Nothing about the shape is copied, so a rounded container's corner and
+the corner its content is cut to cannot fall out of step, and nested
+masks multiply rather than intersect: a pixel inside two feathered
+corners is dimmed by both.
+
+    // before -- the mask clipped to the padded box, the rounding was
+    // only painted behind it, and the two knew nothing of each other
+    field.scrollable_on(Axis::X)
+        .masked()
+        .pad(dp(FRAME_PAD_DP))
+        .background(rect(fill).radius(dp(FRAME_RADIUS_DP)))
+
+    // after -- one rect, drawn and clipped to
+    field.scrollable_on(Axis::X)
+        .pad(dp(FRAME_PAD_DP))
+        .masked_by(rect(fill).radius(dp(FRAME_RADIUS_DP)))
+
+`.masked()` is unchanged for callers and still clips to the widget's own
+box; under it, it now writes an undrawn rect primitive and points the
+mask at that, so square-cornered clipping is the same mechanism rather
+than a special case. `.masked_by(shape)` draws `shape` behind the
+content, in its own layer, and clips to the first primitive it drew.
+There is no radius or shape argument anywhere -- that is the point.
+
+**A press now has to be inside the shape, not just the box.** A corner
+the container rounded away is not there to be tapped, which needed the
+coverage function on the CPU as well as in the shader;
+`iris/tests/mask_sdf.rs` runs the shader's own text against the Rust one
+over a grid of points so the two cannot drift apart.
+
+One limit worth knowing before reaching for it: **a mask's shape must be
+a rect**, asserted by name. Clipping to a glyph or an image would need,
+respectively, a CPU-side alpha plane for the hit test and a bind-group
+switch the fragment stage cannot make. The shader has the branch where
+either would go.
+
+## 2026-09-07: iris runs on a GLES-only Android device, and reports the renderer it cannot build
+
+`AndroidRenderer::new` asked wgpu for `Backends::PRIMARY`, which does not
+include `GL`. A device that offers a Vulkan driver with no adapter behind
+it -- this checkout's emulator -- therefore had no adapter at all, and the
+`.expect` on that turned into a crash loop with nothing on screen. It now
+probes for a `PRIMARY` adapter first and falls back to `Backends::GL` when
+there is none, so **Vulkan still wins wherever it has an adapter** and
+nothing changes on a phone.
+
+The probe deliberately runs on an instance that never touches the window:
+an Android window can be connected to one graphics API only, so an
+instance carrying both backends lets Vulkan claim the window and leaves
+the GLES surface unusable. That is why this is a second instance rather
+than one wider `Backends` value.
+
+The other half a caller sees: `AndroidRenderer::new` already returned
+`Result<Self, String>`, and now **every** way it can fail goes through
+that -- no surface, no adapter, no device, as well as the bind-group
+validation failure it was originally written for. `surface_changed` puts
+that string on screen and in the log ring instead of aborting.
+
+## 2026-09-07: `VelocityTracker` takes positions, not deltas
+
+A flick released at the wrong speed because the tracker averaged. It now
+does what Compose's touch scrolling does, and that changes what a caller
+feeds it.
+
+    // before -- one frame's motion
+    tracker.add_sample(dy, now);
+    // after -- where the finger was
+    tracker.add_position(pos.axis(axis), now);
+
+`VelocityTracker::velocity` is a port of Compose's `VelocityTracker1D`
+with `Strategy.Lsq2`: a degree-2 least-squares fit through the last 20
+positions, differentiated at the newest sample, with Compose's 100ms
+horizon, 40ms stopped-gap and three-sample minimum. Positions rather than
+deltas because a fit needs points on a curve -- Compose itself throws on
+differential data for this strategy.
+
+Three consequences a caller sees. **A gesture with fewer than three
+samples answers `0.0`**, where the average answered a number from two;
+that is Compose's answer too, and on the phone a 120Hz flick delivers
+four or five. **A finger that rests for more than 40ms before lifting
+answers `0.0`** rather than flinging at the speed it arrived with.
+**`add_position` must be called in time order** -- the same debug assert
+as before, now load-bearing for the fit's x-axis.
+
+Also new: `VelocityTracker::samples_display` (the held samples as
+`t_ms:position`, printed by `DragGesture` at debug level so a flick
+reported from a phone can be replayed), `DragArbiter::axis`, and
+`sense::MAX_FLING_VELOCITY_DP_S` (8000, `ViewConfiguration`'s own).
+`List::fling` now applies that maximum against its own density and
+ignores anything at or under 1px/s, which is Compose's pair of thresholds
+exactly -- there is deliberately no 50dp/s minimum, because Compose's
+scrolling never consults the one in `ViewConfiguration`.
+
+## 2026-09-07: `client-core` carries the app's own log
+
+Not iris itself but the crate beside it, and it is a new public surface an
+app author will use: `client_core::log_ring`. Because Iris's phone has no
+`logcat`, an app now keeps a bounded copy of its own log and hands it to
+Dev Updater on the device.
+
+Before, an app installed a platform logger and that was the end of it:
+
+    android_logger::init_once(config);           // Android
+    // nothing at all on the desktop
+
+After, the platform's logger becomes the *inner* logger of a ring that
+records everything alongside it -- `logcat` and a terminal see exactly
+what they saw before:
+
+    client_core::log_ring::install_process_logger(
+        Box::new(android_logger::AndroidLogger::new(config)),
+        LevelFilter::Debug,
+    )?;
+    let ring = client_core::log_ring::process_ring();   // 2000 lines / 256 KiB
+    ring.to_text();                                     // for a report
+    ring.summary();                                     // "1801 lines held, 12 dropped, last 20:09:24"
+
+    // and, for whatever hands the log out of the process:
+    let (lines, next) = ring.since(cursor);   // inclusive of `cursor`
+    ring.newest_seq();                        // None for a ring nothing was written to
+
+`process_ring` is a deliberate process-global, unusually for this project:
+`log` already has exactly one backend per process, and a ring passed around
+as a parameter would be a second answer to "which lines exist".
+
+**Amended later the same day.** `client_core::log_upload` and
+`ai-server`'s `POST /client-log` are **gone** -- an app no longer sends
+its log anywhere. It exposes it on the device instead, and Dev Updater
+reads it there: on Android that is a `ContentProvider` at
+`<applicationId>.devlog`, which is Dev Updater's own contract (its
+`README.md`, "An app's own log") rather than anything iris-specific.
+`LogRing::newest_seq()` is the one addition that went with it: a reader
+holding a cursor uses it to notice the process **restarted**, since the
+ring is in memory and a new process starts again at sequence zero.
+The reasoning and the rejected alternatives are in docs/DECISIONS.md,
+2026-09-07.
+
+## 2026-09-07: `TextData` no longer bundles a font
+
+Iris's call: "remove the font for now; just match what compose does."
+`TextData::default()` used to embed six Noto Sans/Noto Sans Mono `.ttf`s
+(3.6 MB, `include_bytes!`) and register them ahead of the platform's own
+fonts in the `SansSerif`/`Monospace` fallback lists. That registration is
+gone; `TextData::default()`'s signature is unchanged, but what it produces
+now depends entirely on `fontique`'s platform discovery (already on by
+default, previously shadowed) -- Roboto/Roboto Flex on Android, whatever
+the desktop's fontconfig resolves on Linux. No caller-visible type or
+method changed, but every consumer of `iris-core` text now renders with
+whatever the host platform's fonts are, not a fixed bundled face -- worth
+knowing if you were relying on pixel-identical text across devices.
+`.so` shrank by 3.75 MB. One real gap surfaced by the switch: this
+fontique version's Android backend never resolves the `Monospace`
+generic family (a fontique ordering bug, not new in this change), so
+`Family::Monospace` text falls through to the same face as
+`SansSerif` on Android rather than a true monospaced one -- still
+visible, not blank, just not monospaced. docs/RUST.md's "Platform fonts
+(2026-09-07)" has the full account.
+
 ## 2026-09-07: a headless harness, replayed touch, and physical-pixel desktop layout
 
 Layer 1 and 2 of docs/RUST.md's "Three test layers".
@@ -1067,3 +1303,166 @@ and per-block-row work (RUST.md's "Verification pass over Tasks A and B").
   measured. With the counter it is a test: one delta into a 100-paragraph
   reply shapes exactly **1** text layout, the same as into a
   one-paragraph one.
+
+## 2026-09-07: `iris::diagnostics` -- a trace toggle for input/frame lines, gating four existing per-frame `debug!` calls
+
+One new public module and one behaviour change to four existing log
+lines, from Iris's "add another button to copy input event info ...
+instrument a lot of the code with timings" request (RUST.md's own
+section has the full account).
+
+- **`iris::diagnostics::set_trace(bool)`/`trace_enabled() -> bool`**, a
+  process-global switch, off by default. It gates two new diagnostics
+  (`sense::log_input_event`, one line per platform pointer sample under
+  target `iris::input`; `diagnostics::log_frame`, one line per frame
+  under `iris::frame`, with the frame number, the frame clock, time
+  since the last input, layout/draw durations, `RedrawKind`, primitives
+  on screen, and whether something is animating) and, as of a same-day
+  review finding (D1), four *older* `debug!` lines that were previously
+  unconditional: `android::view`'s two `render():` lines, `widget::
+  list`'s `iris fling tick:`, `widget::text`'s `iris text render:`, and
+  `sense`'s `iris drag release samples:`. Not `log::log_enabled!`/
+  `log::set_max_level`, because the app installs its logger at
+  `LevelFilter::Debug` already and the ring records everything that
+  level lets through regardless of target — the gate has to live on
+  this side. **Not wired to a control**: the Diagnostics pane is in
+  `bench_client.rs`, off-limits while another agent had it open; this
+  is the whole surface a button needs.
+- **`UiRenderState` gained `RedrawKind`, `frame_number()`, `epoch()`,
+  `last_layout_duration()`, `last_redraw_kind()`,
+  `active_primitive_count()`, `note_input(Instant)` and
+  `time_since_input(Instant) -> Option<Duration>`** (`iris-core`). All
+  read back by `log_frame`; `note_input` is called once from
+  `SensorUi::run_sensors`, which both backends and the harness already
+  share, so a frame's `since_input` is comparable across all three
+  without either platform doing its own bookkeeping.
+- **`iris::harness::TouchAction` gained `word() -> &'static str`**, the
+  inverse of its own `parse` -- what a caller (here, `Harness::touch`)
+  hands the input logger so a `.touch` file and an `iris::input` line
+  agree on one spelling of each action.
+- **`iris_core::Axis` gained `Debug`** — a one-line derive, needed to log
+  which axis a drag committed to.
+- **`iris/benches/report_to_touch.py`** (new): turns a report's
+  `iris::input` lines back into a `.touch` file, expanding inline
+  historical samples into their own lines first. Round-tripped against
+  the harness in `iris/transcript-fixture/tests/input_log_roundtrip.rs`.
+
+## 2026-09-07: the phone app is told which server to talk to, and pins from the link
+
+Not an iris API change -- a client-facing one, in the crates around it,
+worth knowing because it changes what a build of the Android app *is*.
+
+- **An iris APK is no longer tied to the machine that compiled it.** It
+  used to have the server's host, port, token and CA compiled in, which
+  made a build good for exactly one emulator/server pair and put a token
+  in the artifact. Now it registers `aiapp://enroll` like the Compose app:
+  open the link (Dev Updater's Enroll button already offers it, and the
+  phone asks which app should take it) and the app stores where to go and
+  what to trust.
+- **The CA rides in the link** as `&ca=<base64url DER>`, which is what
+  makes the above possible at all -- a pinned certificate cannot be baked
+  into an APK cross-compiled somewhere else. Optional, so the projects
+  that do build on their own machine keep the short link and the small QR.
+  docs/DECISIONS.md, 2026-09-07, has why not a fingerprint.
+- **`client_core::config` now holds the storage as well as the parsing**:
+  `EnrolledServer` gained an optional `ca_pem`, and `EnrollmentStore` (the
+  0600 JSON file, moved out of `desktop-app`) is one implementation for
+  both the desktop and the phone -- only the directory differs.
+  `desktop-app --ca` is now the override for a link that carried no CA
+  rather than a required flag.
+
+## 2026-09-08: a new GPU device re-uploads its textures instead of forgetting them, and a mark is one texture per shape
+
+Two defects with one cause: **`widget::mark` built a texture per widget**,
+so a transcript screen had one 48x48 standalone image, one bind group and
+one draw call *per folded card* rather than one per picture -- and the
+Android surface-rebuild path assumed no long-lived widget held a texture
+handle at all.
+
+- **`Textures::reset` is gone; `Textures::reupload` replaces it.** A new
+  GPU device holds none of the old one's textures, but this side still
+  holds their pixels, so the answer is to queue every slot for upload
+  again in slot order (empty slots included, as `PushFree`, so the
+  indices after a hole still land where they were) rather than to throw
+  the slot numbering away. Resetting left every live `TextureHandle`
+  naming a slot nothing recognised: the first frame after the emulator's
+  Vulkan-to-GLES fallback panicked with *"texture slot 89 is not a live
+  standalone image: None"*, before anything had been touched.
+- **The glyph atlas is no longer cleared on that path either**, which
+  falls out of the same change: its pages are slots here and their pixels
+  are on this side, so re-uploading restores exactly the atlas that was
+  there. An app switch no longer re-rasterises every glyph on screen.
+- **`Textures::shared(key, make)`** (new): the one texture for a
+  description, built on the first ask and handed out again after, keyed
+  by a `SharedTextureKey { owner, id }` the caller packs *exactly* rather
+  than hashes. The map holds its own reference, so a shared slot is never
+  freed and never recycled under a widget still drawing it. `mark()` is
+  its first caller: three marks now exist for the whole transcript screen
+  (open, closed, collapse) instead of one per card, and the rasterising is
+  paid once.
+
+## 2026-09-08: an app's own log survives the process that wrote it
+
+`devlog`'s provider could only ever show the run that was still up. After
+a crash, Dev Updater's query starts the app process **for the provider
+alone** -- no activity runs, so `MainActivity.nativeSetFilesDir` never
+fired and the panic hook's file was never replayed. The Runtime tab
+therefore showed one line, `iris devlog: serving this app's log at ...`,
+which is exactly the run nobody needs.
+
+- **`DevLogProvider.nativeReady` now takes the files directory too**, and
+  `app_log::set_crash_dir` is called from whichever of the provider and
+  the activity runs first (it deletes the file, so the second says
+  nothing).
+- **The panic hook saves context, not just the panic**: the dying run's
+  last 80 log lines go into the file with it, and are replayed into the
+  new run's ring ahead of the panic line, so the Runtime tab reads
+  chronologically -- what the app was doing, then what killed it, then
+  this run. They are read with a new non-blocking
+  `LogRing::try_tail_text`, because a panic raised while the ring's own
+  lock was held would otherwise deadlock the hook and hang the process
+  instead of aborting it.
+
+## 2026-09-08: iris ships an icon font, and `widget::mark` is gone
+
+Iris's question -- "why does mark exist? The font should be working if
+it's working for compose and nerd fonts are bundled" -- and its answer:
+the Compose app draws icons from its own committed Nerd Fonts subset,
+while iris was setting the disclosure mark with bare Unicode geometric
+codepoints out of whatever face the platform resolved. So iris now does
+what Compose does.
+
+- **`iris::icon`** (new module): the codepoints iris draws, one constant
+  each -- `OPEN`, `CLOSED`, `COLLAPSE` today. Every one has to have a
+  matching entry in `iris/core/build-icon-font.sh`'s `GLYPHS`, which is
+  what builds the shipped `iris/core/assets/fonts/nerd_icons.ttf` (992
+  bytes, Material Design, Mono face). `every_icon_is_in_the_bundled_font`
+  fails the build if the two lists drift.
+- **`Family::Icons`** (new variant): how any text asks for that family.
+  Before/after:
+
+      // was
+      mark(if open { Dir::DOWN } else { Dir::RIGHT }, 9.0, MUTED)
+      // now
+      text(if open { icon::OPEN } else { icon::CLOSED }, 9.0, MUTED)
+          .family(Family::Icons)
+
+  It names an intention, not a font name: only `TextData` knows what the
+  bundled file registered as, and it resolves the variant during shaping
+  (`TextData::resolve_family`, also public). A *named* family rather than
+  a generic one, so nothing falls back into it for ordinary text and an
+  icon cannot fall back out of it onto a system face that happens to have
+  the codepoint.
+- **`iris::widget::mark` is removed** -- added earlier the same day and
+  superseded within it. It drew one correct triangle; every further icon
+  would have been another rasteriser, and an icon as text takes the size,
+  colour and baseline of the line it sits in for free.
+- **`FontDiagnostics::icon_family`** (new field), in the startup log line
+  and the Diagnostics pane: which family the icons resolved to, so a
+  build whose bundled font failed to register says so instead of drawing
+  tofu.
+
+This does not reopen the 2026-09-07 platform-fonts decision. Body and
+monospace text still come from the platform's own collection; an icon is
+the opposite case, a small closed set of codepoints no system font is
+guaranteed to have, and it is the same division the Compose app makes.

docs/IRIS_TODO.md

@@ -701,12 +701,15 @@ Iris's report, verbatim, with a screenshot. Phone: Mali-G715 (Vulkan),
     has no rules for stays plain rather than being coloured by the
     nearest one's.
 
-- [ ] **Masks defined relative to each other.** Wanted: mask A multiplies
-  by something *and also* applies mask B — a mask can reference a parent
-  mask, the way the move chain references a parent offset. Today masks
-  are independent regions. Design it beside the move chain (same shape:
-  a parent index and a bounded walk in the shader); do it when a real
-  widget needs it, not before.
+- [x] **Masks defined relative to each other. (Done: chaining
+  2026-09-07 in d507ae4, the multiply 2026-09-08.)** Built exactly
+  beside the move chain, as this asked: `Mask::parent` is a slot index
+  and the fragment stage walks it under the same bound the move chain
+  uses. Each step multiplies the referenced primitive's coverage into
+  the pixel's alpha, so a pixel inside two feathered corners is dimmed
+  by both — the "multiplies by something *and also* applies mask B" half.
+  The real widget that needed it was the transcript's code fence inside
+  the list. See docs/LAYOUT.md's "Masks with a shape".
 - [ ] **Positions as a single float per scroll.** Iris raised, and half
   rejected, letting a scroll update one float rather than positions:
   input handling cares about most elements in a list, so absolute
@@ -798,23 +801,43 @@ has the fuller account.
   `reposition` fault f5b8893 fixed -- it survives that commit. The
   workaround costs the group the 4dp inset its Compose counterpart holds
   its cards off the edge by, so this is worth fixing.
-- [ ] **`scrollable_on(Axis::X)` on a non-editable `Text` draws nothing.**
+- [x] **`scrollable_on(Axis::X)` on a non-editable `Text` draws nothing.**
   The panel is drawn and the text inside it is not. A markdown fence does
   the same to a `TextEdit` and is fine, so it is the widget kind rather
   than the chain. `tool.rs`'s `raw_block` is `masked()` only until this is
   fixed, which means a long command is clipped rather than pannable.
+  **Not reproducible on 2026-09-08**: `raw_block` was changed to
+  `.scrollable_on(Axis::X).pad(..).masked_by(..)` and the command draws
+  normally (`IRIS_TOOLS_EXPANDED=1 iris/run-headless.sh transcript --shot`,
+  the `rm -rf target` card). Something between 09-06 and 09-08 fixed it --
+  the shaped-mask work (`.masked_by`, 38bf630) is the likeliest, since the
+  old chain was `.masked()` *inside* the padding. Left ticked with the
+  original symptom recorded rather than deleted, in case it comes back.
 - [ ] **No overflow ellipsis.** `TextAttrs` can wrap or not wrap; there is
   no "one line, ellipsised" the way `maxLines = 1` + `TextOverflow.
   Ellipsis` gives Compose. A tool card's summary is clipped instead, so
   nothing on screen says it was cut. Whichever end is cut has to be a
   choice when this lands: a path is identified by its tail, a command by
   its head.
-- [ ] **A drawn chevron.** `Chevron.kt` draws its own strokes precisely
-  because a chevron from a font is a glyph a system font may not have --
-  and the bundled `NotoSans-Regular.ttf` indeed has no U+25B8/25BE/25B4,
-  while `NotoSansMono-Regular.ttf` does. `tool.rs` sets the mark in the
-  monospace face as a result. A real fix needs a line/path primitive;
-  iris has rects, text and textures only.
+- [x] **A chevron the platform cannot fail to have.** **Done
+  2026-09-08**, twice. First as `iris::widget::mark(dir, dp, colour)`,
+  which rasterised an antialiased triangle into the ordinary texture path
+  -- correct, but one bespoke shape, and it built a texture *per widget*,
+  which is what crashed the bench (RUST.md's 2026-09-08 evening entry).
+  Then, on Iris's question -- "why does mark exist? The font should be
+  working if it's working for compose and nerd fonts are bundled" -- as
+  what the Compose app has always done: **iris ships its own Nerd Fonts
+  subset** (`iris/core/build-icon-font.sh` -> `iris/core/assets/fonts/
+  nerd_icons.ttf`, 992 bytes, three Material Design glyphs), named in
+  `iris::icon` and drawn with `Family::Icons`. `mark` is deleted. That
+  serves every future icon rather than one triangle, and an icon is text,
+  so it takes the size, colour and baseline of the line it sits in for
+  free. The original entry, for the record: *the bundled fonts were
+  removed on 2026-09-07 in favour of the platform collection, so the mark
+  is a codepoint the phone's own faces may not have -- Iris's 2026-09-08
+  screenshot shows an empty box where it should be, and the desktop render
+  draws it as a small dot. UI_RULES: "don't rely on characters the
+  platform might not have."*
 - [ ] **A tool card's text is not selectable.** `Selection` is keyed
   `(RowKey, block index)` and a card has no markdown blocks, so nothing in
   a card registers. Compose's `SelectionContainer` covers tool output,
@@ -934,7 +957,7 @@ do not duplicate it there.
 
 ## From the phone, 2026-09-07 (build from ed04d4c)
 
-- [ ] **"Some transcript blocks will be hidden until I uncover enough of
+- [x] **"Some transcript blocks will be hidden until I uncover enough of
   them."** Two screenshots of the bench app's transcript at the top
   edge, both wrong in opposite directions: in one, rows scrolled above
   the viewport are still drawn and bleed *through* the header bar
@@ -955,3 +978,175 @@ do not duplicate it there.
   the header's bottom must be masked. Fix both with one rule: a row is
   drawn if any part of it intersects the viewport, and the viewport is
   the list's own region.
+
+  **Done, e922b73 + d507ae4.** Three causes, and the rule above is what
+  they are all fixed with (`List::intersects_viewport`).
+  `iris/transcript-fixture/tests/top_edge.rs` is the layer-1
+  reproduction -- the real screen under a bench-app-shaped header --
+  and each test was confirmed to fail on its own subject and no other.
+
+  1. *Drawn over the header*: **nothing was clipping the list at all**,
+     and a row straddling an edge is drawn in full, so the part above
+     the list was on screen. It could not be `.masked()` before, either:
+     `Painter::set_mask` aborted when an ancestor already had a mask,
+     and the list's own rows use `.masked()` (a code fence, a tool
+     card's title). So masks nest now -- `Mask::parent`, walked in the
+     fragment stage, chained rather than intersected on the CPU because
+     each mask moves with its own widget. `the_list_is_clipped_to_its_
+     own_box`.
+  2. *Rows already scrolled past still drawn*: the layout walk runs from
+     the anchor, `scroll` moves the anchor's offset and nothing else, so
+     panning leaves the anchor's row further and further outside the
+     viewport and **every row between it and the viewport was drawn,
+     every frame** -- measured at 64 rows for a 2012px viewport after 8
+     scrolls of 3000px. `place` skips a row whose known box does not
+     overlap, and `rehome_anchor` puts the anchor back on a visible row
+     each frame without moving anything drawn.
+     `rows_that_have_left_the_viewport_are_not_drawn`.
+  3. *The blank band*: not a culling rule at all -- the list could rest
+     **past its own first row** (`fling_toward_the_start_stops_at_the_
+     first_row` was leaving it 1398px below a 600px viewport, a blank
+     screen, and that test's own assertion could not see it).
+     `clamp_to_content` gives the gap back. Both ends:
+     `scrolling_past_the_first_row_settles_on_it`,
+     `scrolling_past_the_last_row_settles_on_it`. This is also the first
+     item of the later report below.
+
+  What was suspected and is *not* what happened: the visible-range test
+  never compared a row's top against the viewport's top (there was no
+  culling test at all), and `03c6be8`'s header duplicate is untouched by
+  any of this -- it stays open. A row straddling the top edge is drawn
+  both before and after; the test that would catch that mistake
+  (`the_row_across_the_top_edge_is_drawn`) is in place, and fails if the
+  rule is written against the row's top instead of its bottom.
+
+## From the phone, 2026-09-07, later (build from 4274b8b, ai-app-bench b47eb73)
+
+- [x] **"You shouldn't be able to scroll below the bottom (or above
+  top)."** Done in e922b73, as `List::clamp_to_content` rather than as a
+  clamp inside the scroll setter: nothing at the moment of a `scroll`
+  call knows where the content ends (that is what walking the rows finds
+  out), so the correction is measured from the ends the layout walk
+  already placed and written to the anchor. In the app that lands in the
+  same frame -- a scrolled list is dirty, and `redraw_updates` drains
+  the mark the correction sets before the frame is submitted -- so
+  nothing displaced is displayed; only a full-tree redraw (a resize)
+  could show one frame of it. A fling that reaches an end already ends
+  there (`tick_fling`'s `hit_bound`), and now stops *on* the end rather
+  than wherever the spline's last step had put it. Layer-1 tests at both
+  ends, listed in the item above. The list's offset is not clamped to its content range while
+  dragging and/or flinging. Compose's `LazyColumn` never moves content
+  past its ends -- the overscroll *effect* on Android 12+ is a stretch
+  drawn over clamped content, not a displacement. Clamp the offset in
+  one place (`List`'s scroll setter, so drag, fling, page-in and
+  programmatic scroll all go through it) and end a fling that hits the
+  clamp. Test at layer 1: a drag past either end leaves the offset at
+  the end; a fling into the end stops there.
+- [x] **"Flinging now actually works but is slower than Compose's
+  immediately after releasing the flick (the slow down seems
+  correct)."** Done; RUST.md's "The fling started too slow" has the
+  derivation and the table. On `flick-120hz.touch` the release velocity
+  goes from **12250px/s to 15250px/s**, and on an accelerating flick --
+  the shape a real finger makes, and what the recording is too short to
+  show -- from 1080 to 2445px/s. The curve was right; `VelocityTracker`
+  was averaging total motion over the sample span, which cannot tell an
+  accelerating flick from a steady drag.
+  **Two things the plan for this item had wrong, both found by reading
+  the sources rather than remembering them.** Compose's touch path is
+  **not** `Strategy.Impulse`: `scrollable`/`draggable` release through
+  the 2D `VelocityTracker`, which on Android is two
+  `VelocityTracker1D(strategy = Lsq2)` over absolute *positions* -- a
+  degree-2 least-squares fit, differentiated at the newest sample.
+  Impulse is reached only by `DifferentialVelocityTracker`, for mouse
+  wheel and trackpad. And there is **no minimum** fling velocity on that
+  path: `ViewConfiguration.minimumFlingVelocity`'s 50dp/s is used only by
+  `NestedScrollInteropConnection`, while `DefaultFlingBehavior` skips
+  `abs(v) <= 1f` to dodge a NaN from the spline. So iris ports Lsq2, caps
+  at 8000dp/s, and floors at 1px/s -- no 50dp/s threshold Compose does
+  not have. `iris/benches/velocity_reference.py` is the independent
+  transcription the checked-in numbers come from; the negative control
+  (reverting to the average) fails exactly the seven tests about the
+  estimator and none of the rest. The release log gains a debug
+  `iris drag release samples:` line so a flick reported from the phone can
+  be replayed at layer 1.
+- [~] **Input-event and timing report from the phone.** Iris: "add
+  another button to copy input event info so that I can do some stuff
+  manually and then send the event log to you ... instrument a lot of
+  the code with timings so I can give you time reports through the
+  same button." **Built on the log ring, 2026-09-07** (docs/RUST.md's
+  own section): `iris::sense::log_input_event` (one line per platform
+  pointer sample -- Android's `MotionEvent`, historical samples inline;
+  winit's `WindowEvent`; the harness's `TouchScript` line) and
+  `iris::diagnostics::log_frame` (one line per frame: frame number,
+  the frame clock, time since the last input, layout/draw durations,
+  `redraw_all`/`redraw_updates`/neither, primitives on screen,
+  whether something is animating), both under
+  `iris::diagnostics::trace_enabled()`, off by default because the ring
+  is only 2000 lines / 256 KiB and both targets at 120Hz fill that in
+  seconds. `iris/benches/report_to_touch.py` turns a report's
+  `iris::input` lines back into a `.touch` file for layer 1/2 replay --
+  round-tripped in `iris/transcript-fixture/tests/
+  input_log_roundtrip.rs`. **Not wired to a button**: the Diagnostics
+  pane is `iris/android-app/src/bench_client.rs`, open under another
+  agent at the time this landed; `set_trace(bool)` is the whole surface
+  a control needs. `docs/REVIEW-2026-09-07.md`'s D1 (the ring already
+  drowned in per-frame `debug!` lines that predated this pass) is fixed
+  in the same change -- see RUST.md's section for which four call
+  sites.
+
+## From the phone, 2026-09-07, night (build 92985ba, ai-app-bench bf2088b)
+
+Iris pasted a full Copy report (Mali-G715 Vulkan, 2.55, 120Hz). What it
+showed, beyond her words:
+
+- [x] **"Sometimes when I try to catch it while it's still moving
+  (particularly if I drag) then it fails to stop & snap to where finger
+  is." (done 2026-09-07, b87f5a5.)** Built as described below.
+  `DragArbiter::press_start` takes a `PressState` -- what the target
+  looked like at the moment the press landed -- rather than asking the
+  list later, because by then the fling has already been cancelled and
+  the answer is no. The defect layer 1 found doing it: one touch-down
+  reaches every sensor under the finger, so a block and the tool row
+  containing it deliver the same `PressStart` twice, and re-reading the
+  state on the second delivery turned every catch back into an ordinary
+  slop-waiting press. Tests in
+  `iris/transcript-fixture/tests/catch_a_fling.rs`, with
+  `the_same_small_drag_on_a_settled_list_moves_nothing` as the half the
+  change had no reason to touch. **Not yet confirmed from the phone.**
+  The original reading follows. The report's release lines show catches ending as
+  `v=-41`/`v=-274` pans, so the gesture *does* reach `Panning`, but the
+  content under the finger does not follow it while the fling is still
+  running and the slop has not been crossed. Compose: a down while
+  `isScrollInProgress` stops the fling *at the down* and starts the
+  drag immediately with no touch slop (`scrollable`'s
+  `startDragImmediately = isScrollInProgress`); the content is pinned to
+  the finger from the first sample. Port that: `PressStart` on a list
+  with a live fling ends the fling on that sample and enters `Panning`
+  without waiting for `DRAG_SLOP`; a release with no movement is then a
+  `Released(None)`, not a tap (Compose does not deliver a click either).
+  Layer-1 test on a flick followed by a down + small drag 150 ms later:
+  offset tracks the finger sample-for-sample from the down.
+- [x] **"The copy report button seemed impossible to hit until I hit the
+  diagnostics one." (done 2026-09-07, b8ea723).** Not hit-testing: the
+  button logged `iris bench report: nothing to copy -- run the benchmark
+  first` six times and did nothing on screen. A control that silently
+  declines is the UI_RULES failure "a failure is reported where it
+  happened": `copy_report` now always copies something -- the
+  diagnostics pane's own text (with a first line saying no benchmark has
+  run) when nothing has run yet, or the last report otherwise -- and
+  never depends on another button having been pressed first.
+- [x] **"The logs seem way too big to send in this message box, causes a
+  lot of lag." (done 2026-09-07, 7485d78 + b8ea723).** Two causes. (1)
+  The ring was 1339 lines of `naga::front` / `wgpu_core` / `jni` DEBUG
+  output with 4050 dropped: the ring logger accepted every crate at
+  Debug, and the trace gate (992c472) only covered iris's own lines.
+  `client_core::log_ring::ring_accepts` is the one filter now, applied at
+  the ring rather than per callsite: Debug/Trace only from `iris`/
+  `client_core` targets when tracing is on, Info and above from
+  everything else. (2) Copy report appended the whole ring; it now
+  appends `LogRing::tail_text(COPY_REPORT_TAIL_LINES)` (150, named at the
+  constant) with a first line saying how many older lines were left out
+  -- the full ring is still what the devlog provider hands Dev Updater.
+- [x] Keyboard: the report shows `ime_bottom=891 ime_visible=true` then
+  back to 0 on the phone, so the insets now arrive with a height; the
+  push-up was not reported broken this time.

docs/LAYOUT.md

@@ -948,7 +948,7 @@ When this lands, copy this entry into `IRIS.md` (newest first):
 > design, the move-offset mechanism this shipped alongside, and the file
 > list.
 
-## Masks with a shape (decided 2026-09-07, not yet built)
+## Masks with a shape (decided 2026-09-07, built 2026-09-08)
 
 Iris, on the code block's scrolling: "the code block scrolling currently
 masks in an inner rectangle. Ideally masks should have a shape
@@ -1039,3 +1039,38 @@ and that the CPU SDF and the shader agree at a grid of points; a
 `run-headless.sh --phone` screenshot of a scrolled code block shows
 rounded corners with no square pixels poking out at the top and bottom
 of the scrolled content. Record the commands in RUST.md when it lands.
+
+### What was built (2026-09-08), and where it differs
+
+The commands and the screenshot are in docs/RUST.md's queue entry. Four
+places the code is narrower than the design above, each deliberate:
+
+- **No `kind` and no `flags` on `Mask`.** It is `{ primitive, parent }`.
+  The referenced instance already carries its own `binding`, so a copy
+  of it in the mask is a second thing to keep in step; *alpha only* is
+  the only mode there is, so there is nothing to select. Both are a
+  field away if a second mode appears.
+- **A mask's shape must be a rect.** `Painter::set_mask_to` asserts it,
+  by name, rather than leaving the shader to read a `rects` entry that
+  is not there. A glyph would need a CPU-side alpha plane before the
+  hit test could agree with the shader, and a standalone image needs a
+  bind-group switch the fragment stage cannot make (`masks_layout`'s own
+  comment on why an image's bind group must not name the masks buffer).
+  So **the texture-mask pass condition is not met and no texture mask
+  exists** — the point of the reference design is that adding one is a
+  binding check and a sampled alpha, with no new shader path, and the
+  shader's `mask_coverage` already has the branch where it would go.
+- **The shape is a primitive of its own, not always a drawn one.** A
+  plain `.masked()` writes an undrawn `RectPrimitive` at its region
+  (`Drawn::No`, `NOT_DRAWN`) and points the mask at that, so "clip to my
+  box" and "clip to that widget's rounded background" are one mechanism
+  and square-cornered clipping did not become a special case.
+  `.masked_by(shape)` draws `shape` behind the content — in its own
+  layer, the way `Stack` puts a background under its content — and
+  clips to the first primitive it drew.
+- **The CPU/shader agreement is a GPU test**, `iris/tests/mask_sdf.rs`,
+  the only test in the workspace that needs an adapter. It lifts
+  `distance_from_rect` and `rounded_rect_coverage` out of
+  `iris_core::SHAPE_SHADER` *by name* and runs them in a compute pass,
+  so the thing under test is the shader itself rather than a copy of it
+  that would be edited alongside.

docs/REVIEW-2026-09-07.md

@@ -0,0 +1,459 @@
+# Review, 2026-09-07 — `ba2afba..origin/rustify`
+
+Read-only review of the day's 24 commits: the glyph-atlas fix, the fling
+spline and Lsq2 velocity estimator, keyboard/IME insets and `targetSdk`,
+historical touch samples and the input clock, list culling / clamp /
+anchor re-homing, nested masks and `draw_again`, the headless harness +
+`transcript-fixture` + `rig-input`, desktop density, the release profile,
+platform fonts + the Android monospace patch, and the client-core log ring
+with `POST /client-log`.
+
+**Verified while reviewing** (working tree, which also carries three other
+agents' uncommitted edits — `iris/src/sense.rs`, `iris/core/src/ui/render_state.rs`,
+`iris/src/lib.rs`, `iris/core/src/orientation/axis.rs`, and an untracked
+`iris/src/diagnostics.rs`): `cargo fmt --check` clean in `iris/`,
+`client-core/` and `server/`; `cargo clippy --all-targets` clean in `iris/`
+and `client-core/`; `cargo test --lib -p iris` 101 passed, `cargo test -p
+transcript-fixture` 10 passed. The `iris` doctest target fails to link
+(`extern location for iris_core does not exist`) — a stale build artefact,
+not a code fault, but worth knowing before trusting `cargo test -p iris`
+as a whole.
+
+The work is unusually well documented and the two "a test that compared
+the code with itself" findings the authors made themselves are real and
+were fixed correctly. What follows is what is left.
+
+Counts: **5 defects, 7 risks, 3 tests that cannot fail in the bug's
+direction, 7 rule findings, 2 nits.**
+
+## Fix pass, 2026-09-07 evening
+
+Every finding below carries a **Status** line. In summary: **13 fixed**
+(D1, D4, D5, R1, R5, R7, T1, T2, T3 and four of the rule findings and both
+nits), **6 moot or deferred** (D2, D3, R3, R4 and two rule findings, all
+of them in the phone-logging route that `06b8a1f` deleted or in files the
+devlog agent held open), and **2 not done on purpose** (R2, which waits on
+docs/LAYOUT.md's mask redesign, and R6, which needs Iris's own phone).
+
+The commits are `2ec0fee` (D4), `7e79ec1` (D5), `551c013` (R1), `e10582a`
+(T1-T3), `ff1d6ea` (R5, R7) and `a6a100e` (the rename and the nits). Each
+fix that the rig can express carries a test, and each of those was
+confirmed by breaking its subject on purpose -- the break is recorded
+beside the assertion, so the next reader does not have to re-derive it.
+
+---
+
+## Defects
+
+### D1 — the app's own log ring is drowned by the same day's per-frame `debug!` lines, so the route built to get Iris's logs to her carries almost none of them
+
+`iris/android-app/src/lib.rs:132` installs the ring at `LevelFilter::Debug`,
+and `client-core/src/log_ring.rs:279` (`RingLogger::enabled`) returns
+`true` unconditionally by design, so **every `log::debug!` in the process
+lands in a 2000-line / 256 KiB ring**. In the same commit range that ring
+became the only way a line reaches Iris, three ungated per-frame `debug!`
+callsites are live:
+
+- `iris/src/android/view.rs:446` and `:509` — two lines *per rendered frame*.
+- `iris/src/widget/list.rs:576` — `iris fling tick:`, one line per fling tick.
+- `iris/src/widget/text/mod.rs:81` — one per text shape (many per frame while rows compose).
+
+**Failure scenario.** Iris flicks the transcript on a 120 Hz phone. That is
+~240–360 debug lines a second; the ring's 2000-line bound is exhausted in
+**under ten seconds**, so by the time she presses `Copy report` every
+`log::info!` about what she was actually investigating has been evicted.
+The uploader makes it worse: it sends at most the ring per 10 s wake
+(2000 lines ≈ 200 lines/s) against ~350 lines/s produced, so it also runs
+permanently behind and pushes tens of KB/s of frame spam over the tunnel.
+
+Note that another agent has already built the right mechanism — the
+untracked `iris/src/diagnostics.rs` has `set_trace`/`trace_enabled`, a
+default-off gate, and its module doc states this exact problem in as many
+words. It gates `iris::input`/`iris::frame`; it does **not** gate the four
+callsites above.
+
+*Fix*: put `List::tick_fling`'s line and `view.rs`'s two `render():` lines
+behind `iris::diagnostics::trace_enabled()` (the mechanism that already
+exists for exactly this), and/or record into the ring at `Info` while
+leaving `android_logger` at `Debug`.
+
+**Status:** fixed in `992c472` (verified 2026-09-07: all four callsites, plus `sense.rs`'s drag-release samples line, now sit behind `iris::diagnostics::trace_enabled`, and `input_log_roundtrip` proves both directions).
+
+### D2 — `POST /client-log` can make `ai-server` write an unbounded runtime log at an authenticated client's request
+
+`server/src/routes.rs:1473` bounds the **line count** (500) and nothing
+else. The route sits inside the router that applies
+`DefaultBodyLimit::max(32 * 1024 * 1024)` at `server/src/routes.rs:179`
+(raised for phone photos), so one request may carry 500 lines of ~64 KiB
+each, and each is re-emitted verbatim into `tracing`. There is no
+per-message cap on the server, no rate limit, and the runtime log
+`ai-server` writes is the file Dev Updater tails and never rotates.
+`MAX_MESSAGE_BYTES` (4096) exists only in the *client*
+(`client-core/src/log_upload.rs:33`), i.e. the server trusts a value the
+attacker controls.
+
+**Failure scenario.** A buggy client (a `log::debug!` in a loop is enough —
+see D1) or one holding a leaked bearer token posts 32 MiB every 10 s; the
+host's disk fills and every other component's log goes with it.
+
+*Fix*: give the route its own `DefaultBodyLimit` (the attachments route at
+`:175` is the precedent for a per-route limit) and truncate each `message`
+server-side to the same 4096 bytes rather than assuming the client did.
+
+**Status:** moot -- `POST /client-log` was deleted with the whole upload route (`06b8a1f`), the app hands its log to Dev Updater through an on-device ContentProvider instead. Nothing to bound.
+
+### D3 — lines the ring drops before the uploader sends them vanish with nothing saying so
+
+`LogRing::since` (`client-core/src/log_ring.rs:169`) filters `seq >= cursor`
+and silently returns fewer lines when eviction has passed the cursor;
+`LogUploader::flush_once` (`:94`) then advances to whatever came back.
+`dropped` is counted (`log_ring.rs:109`) and shown in the *local*
+diagnostics pane, but it is never put in the upload body, and
+`ClientLogBody` has no field for it.
+
+**Failure scenario.** The tunnel is down for two minutes; the ring wraps.
+When it comes back, the server log jumps from `#812` to `#5106` with no
+line saying anything was lost. This is precisely the "unknown state
+sharing a value with the empty state" UI_RULES asks to design first, and
+the module doc for `dropped` claims it is "reported rather than inferred"
+— it is, but only on the half of the path nobody is reading.
+
+*Fix*: carry `dropped` (or `firstSeq`) in the batch and have `client_log`
+emit one `warn!` when the sequence is not contiguous with the last batch
+from that `source`.
+
+**Status:** moot -- `client-core/src/log_upload.rs` was deleted with the route (`06b8a1f`). Whatever the ContentProvider does about eviction is that design's question, not this one's.
+
+### D4 — the input clock anchors on the first event's *own* time, so that event's historical samples are dated before the anchor: the ordering assert fires, and release silently collapses them onto one instant
+
+`iris/src/android/view.rs:628` takes the anchor as
+`(Instant::now(), event.event_time_nanos())` from the first `MotionEvent`
+the view ever sees, and `at()` computes
+`anchor_at + (sample_time - anchor_nanos).max(0)`. Historical samples of
+that same event are by definition **earlier** than its own `event_time`.
+
+**Failure scenario.** The first event this view receives is an
+`ACTION_MOVE` (the `DOWN` was delivered to another view, or the view was
+attached mid-gesture). Its historical samples are, say, 12 ms before
+`anchor_nanos`; `at()` clamps all of them to `anchor_at`, so the tracker
+receives three samples with identical timestamps, the Lsq2 fit is
+degenerate, and the flick reads 0 px/s. In a debug build the
+`debug_assert!(ht >= previous)` at `:653` fires first — but `previous`
+starts at `anchor_nanos` (`:651`), which is a value from a *different*
+event, so that assert is also the wrong comparison for the first sample of
+every later event.
+
+*Fix*: anchor on the earliest sample of the first event
+(`historical_event_time_nanos(0)` when `history_size() > 0`, else
+`event_time`), and seed `previous` from the previous event's last sample
+rather than from the anchor.
+
+**Status:** fixed in `2ec0fee`. The arithmetic moved into `sense::PointerClock`, which anchors at `now - (event_time - oldest_sample)` and carries the last sample seen *across* events, so the ordering assert compares against the previous event's last sample rather than the anchor. It lives in `sense` because `iris::android` is `cfg`'d out everywhere but the device: `sense_tests.rs`'s `the_first_events_batched_samples_are_dated_apart` reports `[0ns, 0ns, 0ns]` against the old anchoring.
+
+### D5 — the "before" velocity quoted in four places is not what the reference script prints
+
+`iris/benches/velocity_reference.py`, run today, prints **12250 px/s** for
+`flick-120hz.touch`'s average and **12500 px/s** for "press and one move
+frame". Four places say 11750 for both:
+
+- `docs/RUST.md:900` (`flick-120hz.touch | 11750 px/s`)
+- `docs/RUST.md:905` (`press + one move frame | 11750 px/s`)
+- `docs/IRIS_TODO.md:1026`
+- `iris/transcript-fixture/tests/phone_screen.rs:55`
+
+`iris/src/sense.rs:1406` has the correct 12250, so the two halves of the
+same change disagree. The file that carries the wrong number is the one
+that says "every number below is printed by `velocity_reference.py` … do
+not 'fix' one by running the Rust and copying what it said". One of the
+two rows also being 11750 for a completely different sample set is the
+tell.
+
+*Fix*: replace 11750 with the script's own 12250 / 12500 in those four
+places, or say which run produced 11750.
+
+**Status:** fixed in `7e79ec1`. All four places now say 12250 / 12500, the 1.30x ratio becomes 1.24x, and RUST.md records where 11750 half came from (196 px over a 16.68 ms **60 Hz** frame rather than the recording's own 16 ms -- which explains the flick row and not the other one, so that one was copied).
+
+---
+
+## Risks
+
+### R1 — every new invariant guard is a `debug_assert!`, and the phone runs release
+
+The five guards added today —
+`iris/src/widget/list.rs:1156` (a `List` must be inside a `.masked()`),
+`:1218` (`extents` holds only on-screen rows),
+`iris/src/android/view.rs:653` (historical sample ordering),
+`iris/src/sense.rs:1076` (`poly_fit_least_squares` sample count), and
+`iris/core/src/ui/painter.rs`'s doubled-`set_mask` check — are all
+`debug_assert!`. `docs/RUST.md` records that the bench APK **must** be
+installed as `release` on the emulator (the debug `libmain.so` is 325 MB
+and will not install) and Iris's phone gets release too. So none of these
+can fire on any build anybody actually runs; in release a `List` drawn
+without a mask silently paints over its surroundings again — the exact
+fault e922b73 was written to fix.
+
+*Fix*: for the two that are cheap and once-per-draw (`is_masked`, the
+extents check), consider a plain `assert!` or a one-shot `log::error!`, so
+the guard survives into the build the defect was found in.
+
+**Status:** fixed in `551c013`. `is_masked`, the `extents` check, `set_mask`'s doubled-call check, `Painter::glyphs`'s atlas generation and `List::fling`'s finiteness are `assert!`/`assert_eq!` now; `List::place`'s slot precondition, `poly_fit_least_squares`'s two, and `PointerClock::sample`'s ordering stay `debug_assert!` and say in a comment why. The layer-1 suites pass in `--release` as well as debug, which is what says the promoted ones do not fire on a real replayed flick.
+
+### R2 — a straddling row is now invisible above the list and still tappable through the header
+
+Masks are applied in the fragment shader
+(`iris/core/src/render/shader.wgsl:203`); the CPU hit path
+(`UiRenderState::resolved_region`, `iris/core/src/ui/render_state.rs:709`)
+does not consult `masks` at all. Before today the top of a straddling row
+was drawn over the header *and* hit-testable there; now it is clipped away
+but still hit-testable, which is worse — a tap on "Run benchmark" can land
+on an invisible link in the row behind it. `docs/LAYOUT.md:1012` ("Hit-
+testing applies the shape") is design, not code.
+
+*Fix*: until LAYOUT.md's mask redesign lands, intersect a widget's hit
+region with its mask chain in `resolved_region`; the chain walk already
+exists on the GPU side.
+
+**Status:** not done, deliberately -- docs/LAYOUT.md's mask redesign ("masks reference a drawn primitive instead of copying a shape", `1121d7c`) is where hit-testing gets the shape, and intersecting a chain in `resolved_region` now would be a second mechanism to unpick. Pointer left here rather than a fix.
+
+### R3 — three copies of one wire contract, none of them linked
+
+`client-core/src/log_upload.rs:28` (`MAX_LINES_PER_BATCH = 500`) and
+`server/src/routes.rs:1418` (`CLIENT_LOG_MAX_LINES = 500`) must agree, in
+different crates, with only a comment saying so; the body itself is built
+by hand with `serde_json::json!` on one side and parsed by a
+`#[serde(deny_unknown_fields)]` struct on the other. This project already
+has the mechanism for exactly this — `event-model`, a crate both `server`
+and `client-core` depend on precisely so "the app hand-mirroring it" stops
+happening (`server/Cargo.toml:16` says so).
+
+**Failure scenario.** Somebody raises the client's batch to 1000. Every
+upload now returns 400, the uploader retries the *same* batch from the same
+cursor forever, and the only sign is one line in a diagnostics pane on a
+phone.
+
+*Fix*: move `ClientLogLine`/`ClientLogBody` and the batch constant into a
+shared crate.
+
+**Status:** moot -- both copies went with the route (`06b8a1f`). If a client/server contract comes back, `event-model` is still the answer.
+
+### R4 — `build.rs` bakes in a CA it never asks Cargo to watch, and the bench build now has no rebuild trigger at all
+
+`emit_log_config` (`iris/android-app/build.rs:92`) calls `read_pinned_ca()`
+but emits only `rerun-if-env-changed` for `AI_APP_LOG_HOST/_PORT/_TOKEN` —
+no `rerun-if-changed` for the CA *file*, and (because the bench build
+returns at `:65`, before the transcript path's declarations) no
+`rerun-if-env-changed=AI_APP_CA`/`XDG_CONFIG_HOME` either. Emitting any
+`rerun-if-*` directive turns off Cargo's default "rerun when anything in
+the package changes" heuristic, so the bench build lost the only trigger it
+had.
+
+**Failure scenario.** `~/.config/ai-app` is wiped (AGENTS.md calls this the
+one-way door), `ai-server` mints a new CA, the APK is rebuilt — and
+`build.rs` does not re-run, so the APK still pins the dead CA and every
+upload fails with a TLS error nobody can attribute.
+
+*Fix*: `println!("cargo:rerun-if-changed={}", ca_path.display())` inside
+`read_pinned_ca`, and move the `AI_APP_CA`/`XDG_CONFIG_HOME` declarations
+above the bench early-return.
+
+**Status:** moot -- `iris/android-app/build.rs` was deleted (`06b8a1f`/`d8562d9`): the destination comes from the enrolment link now, so nothing is baked in at build time and there is nothing for Cargo to watch.
+
+### R5 — desktop density is read once and never updated
+
+`iris/src/default/mod.rs:254` reads `content_scale(window)` at startup and
+sets it on both `rsc.ui.text.density` and `render`. `WindowEvent::
+ScaleFactorChanged` is not handled, and `UiRenderer::resize` deliberately
+no longer consults `scale_factor`. Dragging the window to a monitor with a
+different scale leaves every `dp(...)` and every rasterised glyph at the
+old density — the same class of disagreement the commit removed elsewhere.
+It is invisible here (every display on this machine is 1.0), which is why
+it needs writing down.
+
+**Status:** fixed in `ff1d6ea`. `WindowEvent::ScaleFactorChanged` re-reads `content_scale` -- through that function, so `IRIS_SCALE` still pins `--phone`'s density instead of following the monitor -- and `UiRenderState::set_density` marks the tree for a full redraw when the value actually changes, since `Text::shape` keys its cache on `(attrs, width, density)`.
+
+### R6 — removing the bundled fonts removed the guard for a fault that was found on the phone, and the check was run on the desktop
+
+`iris/core/src/primitive/text.rs`'s `register_bundled_fonts` existed
+because "bold spans on a real phone rendered as blank gaps of the correct
+advance width" — the deleted doc says so. Its removal is Iris's own call
+and is recorded properly in `docs/DECISIONS.md`, but the verification
+recorded there is "checked with CJK + emoji **on desktop**", which is the
+half that cannot fail: the fault was Android's font enumeration resolving
+a weight/style. `iris/transcript-ui/src/tool.rs:110`'s comment is honest
+that `CLOSED_MARK`/`OPEN_MARK`/`UP_MARK` (U+25B8/BE/B4) are now "a bet"
+that the platform monospace face has them — which is UI_RULES' "don't rely
+on characters the platform might not have", stated and then accepted.
+
+*Fix*: before the next phone build, look at a bold run and the three
+chevrons on Iris's device specifically; the emulator's font set is not
+evidence for hers.
+
+**Status:** not done here -- it is a *look at it on Iris's phone* item, and no build in this VM is evidence about her device's font set. Carried forward as the review said: before the next phone build, look at a bold run and at `CLOSED_MARK`/`OPEN_MARK`/`UP_MARK` (U+25B8/BE/B4) on her device specifically.
+
+### R7 — the least-squares fit clamps a degenerate norm instead of detecting it
+
+`iris/src/sense.rs:1105`: `1.0 / dot(...).sqrt().max(1e-6)`. Compose's
+`polyFitLeastSquares` treats `norm < 1e-6` as "vectors are linearly
+dependent, no solution" and bails; clamping instead produces a `q` row of
+zeros, a zero on `r`'s diagonal, and a `0/0` that the `is_finite` check at
+`:1059` happens to catch. It works, but it works by accident and the escape
+is not the one the source it is transcribed from takes.
+
+**Status:** fixed in `ff1d6ea`. `poly_fit_least_squares` returns `Option` and bails at `DEGENERATE_NORM` (Compose's `0.000001f`) instead of clamping; `velocity()` answers 0 on `None`. `a_fit_through_linearly_dependent_points_has_no_solution` reports `Some([NaN, NaN, NaN])` with the clamp back in place.
+
+---
+
+## Tests that cannot fail in the direction the bug would go
+
+### T1 — `iris/transcript-fixture/tests/phone_screen.rs:64` computes the expected fling duration with the calculator under test, and asserts it one-sidedly
+
+`let expected = FlingCalculator::new(PHONE_SCALE).duration(velocity);` then
+`assert!(ran_for <= expected + 2 frames)`. This is the same
+"calculator compared with itself" shape the fling-spline commit
+(73f956f) identified and fixed elsewhere, and the direction it can fail in
+is "the fling ran too long" — never "the fling stopped dead", which is
+literally Iris's reported symptom. The companion
+`assert_ne!(before, after)` passes on one pixel of travel. A fling that
+settles on the first tick passes this test.
+
+*Fix*: add a lower bound from `velocity_reference.py`'s number (a fling at
+-15250 px/s at density 2.55 must run ≥ ~1.4 s and travel ≥ ~6000 px), not
+from `FlingCalculator`.
+
+**Status:** fixed in `e10582a`. Both bounds come from `fling_spline_reference.py`, which gained this case's own line (`density=2.55 v=15250.0: distance=11057.424px duration=2.0716s`), and travel is measured in pixels from a row's own on-screen extent (10527px measured). Scaling `tick_fling`'s elapsed by 1000 reports "stopped after 8ms"; scaling its delta by 0.01 reports "travelled 111px".
+
+### T2 — `top_edge.rs:150` checks a row *count* on the leg where the culling bug appeared, and the box only on the other leg
+
+`rows_that_have_left_the_viewport_are_not_drawn` asserts `rows.len() <= 24`
+on the outbound leg and the per-row `inside the box` predicate only on the
+return leg. The doc explains why (an unmeasured row must be drawn to be
+measured), which is correct — but it means the test's name is only true of
+half of it, and a regression that draws 20 rows in the wrong *place* on the
+outbound leg passes.
+
+**Status:** fixed in `e10582a`. The first leg still cannot assert the box (an unmeasured row has to be drawn to be measured), so there is a third leg -- back again, every height known. Widening `intersects_viewport` downwards passes all 40 forward steps and fails at "back 6".
+
+### T3 — `top_edge.rs:116` checks that a mask exists and where it is, not that it reaches anything
+
+`the_list_is_clipped_to_its_own_box` asserts `active.mask != MaskIdx::NONE`
+and that the mask's region lies within the list's box. It never checks the
+row primitives actually reference that mask, so a broken `Mask::parent`
+chain — the thing d507ae4 introduced — would leave this green while a code
+fence inside a row drew unclipped again.
+
+*Fix*: assert that a row primitive's mask chain contains the list's mask
+slot.
+
+**Status:** fixed in `e10582a`. It walks every row primitive's mask chain and requires the list's own slot on it, and rejects a chain that loops. Forcing `Painter::set_mask`'s `parent` to `NONE` fails it with "clips to [Id(1)], a chain that never reaches the list's own mask Id(0)".
+
+---
+
+## Rules
+
+- **`iris/src/widget/list.rs:576` is a second mechanism for per-frame
+  instrumentation.** `iris::diagnostics::trace_enabled` exists for exactly
+  "a default-off `debug!` in a hot path" and this line does not use it.
+  (Cause of D1; the gate is in the untracked `diagnostics.rs`, so at the
+  reviewed commit the line is simply ungated.)
+- **`server/src/routes.rs:1518` (`client_log_time`) duplicates
+  `client-core/src/log_ring.rs:76` (`clock_time`)** — the same arithmetic
+  written twice in two crates, with a comment noting they must agree. Same
+  shared-crate answer as R3.
+- **`client-core/src/log_ring.rs:301`'s doc claims more than the code
+  delivers**: "the caller is named in the error so it is findable" —
+  `log::SetLoggerError` names nobody. `iris/android-app/src/app_log.rs:44`
+  repeats the claim.
+- **Stale comment: `iris/src/android/view.rs:624`** cites
+  `VelocityTracker::add_sample`'s debug assert; the method was renamed to
+  `add_position` in the same commit range.
+- **`MOVE_CHAIN_LIMIT` now bounds two different chains** (move offsets and
+  masks) under a name that says one, in both
+  `iris/core/src/ui/render_state.rs:63` and `shader.wgsl:97`. The shader's
+  comment already calls it "the bound on the parent walk"; the constant
+  should say that too, or masks should get their own.
+- **`iris/src/sense.rs:1434`'s stated negative control is not reproducible
+  as written.** "Reverting `velocity` to `total / span` fails exactly this
+  one, the flick recording, and `phone_screen.rs`" — but `samples` now
+  holds *positions*, so `total / span` over them gives 2750 for the steady
+  drag too, and the commit message for the same change says "exactly seven
+  tests". Two numbers for one experiment.
+- **`iris/android-app/src/bench_client.rs:393`'s `ime_visible` is right and
+  its sibling one line up is not.** `set_bottom_inset(rsc,
+  insets.bottom.max(insets.ime_bottom))` still infers "make room" from a
+  `max`, so during the slide-in the composer is padded by the system-bar
+  inset while `ime_visible` already says the keyboard is up. Harmless
+  today; it is the same conflation the comment beside it warns about.
+
+**Status of the rule findings, 2026-09-07 evening.**
+
+- `list.rs:576`'s ungated per-frame line -- **fixed in `992c472`** with
+  the rest of D1.
+- `routes.rs:1518`'s `client_log_time` duplicating `log_ring.rs`'s
+  `clock_time` -- **moot**: the route was deleted (`06b8a1f`).
+- `log_ring.rs:301`'s "the caller is named in the error" -- **deferred to
+  the devlog agent**; `client-core/src/log_ring.rs` is its file this pass,
+  and `app_log.rs` no longer repeats the claim.
+- `view.rs:624`'s stale `VelocityTracker::add_sample` -- **fixed in
+  `2ec0fee`**; the paragraph was rewritten for the anchoring change and
+  now names `PointerClock` rather than a method that no longer exists.
+- `MOVE_CHAIN_LIMIT` naming two chains -- **fixed in `a6a100e`**: renamed
+  to `PARENT_CHAIN_LIMIT` in `render_state.rs` and `shader.wgsl` at once
+  (it had no other users), with the doc naming both chains it governs.
+- `sense.rs:1434`'s unreproducible negative control -- **fixed in
+  `7e79ec1`**. Rerun with `velocity` reverted to `(newest - oldest) /
+  span`: seven fail in `-p iris` (the flick recording, the accelerating
+  flick, the horizon, the stopped finger, the minimum sample count, both
+  `drag_gesture` flick tests) plus `phone_screen.rs`'s flick. RUST.md's
+  "exactly seven" was right; the doc comment's "exactly this one, the
+  flick recording, and `phone_screen.rs`" was not, and now says the same
+  thing RUST.md does.
+- `bench_client.rs:393`'s `set_bottom_inset(.., max(..))` -- **deferred to
+  the devlog agent**; `iris/android-app/**` was open under it this pass.
+
+## Nits
+
+- `iris/src/sense.rs:798` computes `self.velocity.velocity()` twice on a
+  release when `info` logging is on (once for the outcome, once for the
+  log line) — a full Lsq2 fit each.
+- `iris/transcript-ui/src/selection.rs:303` calls `ui.ui_mut().animate(id)`
+  even when `fling()` bailed (`|v| <= 1.0`, or no anchor). Harmless — the
+  first `tick` unregisters — but it registers an animation that is known
+  not to exist.
+
+---
+
+**Status of the nits, both fixed in `a6a100e`.** `DragGesture`'s release
+computes `velocity()` once into a local both the outcome and the
+`iris drag release:` line read. `selection.rs`'s `animate(id)` is behind
+`is_scrolling()`, which is the same answer `List::fling` itself reached --
+and `phone_screen.rs`'s recorded flick still flings, which is the half
+that says the guard did not turn a working release off.
+
+## Commits reviewed
+
+```
+7e4e26a iris: resolve fontique's Android monospace generic family ourselves
+84a13e8 iris: a fling starts at Compose's velocity, which is a curve fit and not an average
+452c442 docs/RUST.md: queue -- logging landed; iris app enrolment ...
+238057a docs: the phone-logging decision, how to use it, and two build-apk traps
+896c93a iris: drop bundled Noto Sans, match Compose's platform-font fonts
+690161e docs: the transcript's edges were three faults, and what the rig found
+e922b73 iris: a transcript row is drawn if it overlaps the viewport, and clipped to it
+d507ae4 iris-core: masks nest instead of aborting, and a widget can ask to be drawn again
+9ed01e2 docs: phone report 2026-09-07 later -- overscroll, low initial fling velocity ...
+5be9f1b iris-android-app: keep the app's own log, put it in Copy report, upload it
+977bdb9 client-core: the app's own log ring, and POST /client-log to get it off a phone
+9cd1263 docs/RUST.md: queue -- APK size done, the embedded-fonts question left for Iris
+42af780 iris android-app: strip+LTO+cgu1+opt-level=s halve libmain.so, no feature trim needed
+4274b8b Merge remote-tracking branch 'origin/rustify' into worktree-agent-ace98b0bdaf33ffff
+73f956f iris: the fling curve was the identity function, and the keyboard was a targetSdk
+038f6a3 docs: the test rig's layers 1 and 2, with their commands and their limits
+1121d7c docs/LAYOUT.md: masks reference a drawn primitive instead of copying a shape ...
+232de0e iris: a phone-shaped desktop window, driven by the same touch recordings
+e430880 docs: phone report 2026-09-07, rows at the transcript's top edge culled early ...
+a999bd1 docs: masks with a shape (LAYOUT.md, decided 2026-09-07) and the orchestrator queue
+6840edf iris-android-app: the bench's fixture half comes from transcript-fixture
+3332201 iris: a headless in-process harness, and the bench fixture as a shared crate
+7f4ea7e docs/TODO.md: Compose app crash from Iris's phone log export, reversed AnnotatedString range
+591128e AGENTS.md: the phone app and the planned desktop app share widgets and styling
+```

iris/Cargo.lock

@@ -720,7 +720,9 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 name = "client-core"
 version = "0.1.0"
 dependencies = [
+ "base64",
  "event-model",
+ "log",
  "pulldown-cmark",
  "serde",
  "serde_json",
@@ -1735,6 +1737,7 @@ dependencies = [
  "accesskit_winit",
  "android-view",
  "arboard",
+ "bytemuck",
  "image",
  "iris-core",
  "iris-macro",
@@ -3662,6 +3665,7 @@ dependencies = [
  "client-core",
  "event-model",
  "iris",
+ "log",
  "serde_json",
  "transcript-ui",
  "winit",

iris/Cargo.toml

@@ -60,15 +60,20 @@ accesskit_android = "0.8.0"
 send_wrapper = "0.6.0"
 
 [features]
-# RUST.md's I5 "Where iris's frame time goes" diagnosis: forces the Android
-# `wgpu::Instance` to `Backends::GL` instead of `Backends::PRIMARY`, so the
-# same build can be measured against SwiftShader's software Vulkan ICD (the
-# default) or virgl's GLES path, without a second env-var plumbing path that
-# nothing on this machine can hand to an already-launched Android process
-# (there is no `am start` environment and no system-property reader here to
-# add one). Read by `android/render.rs` and, so the GLES path can be
-# reproduced on a machine with a real GPU rather than only in the emulator,
-# by `default/render.rs`:
+# RUST.md's I5 "Where iris's frame time goes" diagnosis: pins the
+# `wgpu::Instance` to `Backends::GL` instead of `Backends::PRIMARY`, so one
+# build can be measured on either backend. A compile-time feature rather
+# than an env var because nothing on this machine can hand an env var to an
+# already-launched Android process (there is no `am start` environment and
+# no system-property reader here to add one).
+#
+# **Not needed to get GLES in the emulator**, whatever the history here
+# says: the emulator's guest has no hardware Vulkan at all, so an ordinary
+# build's runtime fallback lands on GLES by itself (docs/RUST.md, "What the
+# emulator gives a GPU app"). Keeping the emulator on the same binary the
+# phone runs is the point. What this feature is still for is forcing GLES
+# on a machine that *does* have Vulkan -- the desktop -- which is why
+# `default/render.rs` reads it too:
 #     ./run-headless.sh transcript --shot /tmp/x.png -- -p transcript-ui \
 #         --features iris/force-gles
 force-gles = []
@@ -79,6 +84,9 @@ tokio = { workspace = true, features = ["sync", "rt", "rt-multi-thread", "time"]
 # package is fine -- cargo excludes dev-dependencies from the graph used
 # to build the library itself, so this only matters for `--examples`.
 tabs-ui = { path = "tabs-ui" }
+# `tests/mask_sdf.rs` only: the grid it hands the GPU and the coverages it
+# reads back. wgpu and pollster are ordinary dependencies already.
+bytemuck = { workspace = true }
 
 # Plain Instant-timed binaries, not criterion -- see benches/message_list.rs's
 # header for why. `harness = false` opts out of the unstable `#[bench]`
@@ -110,6 +118,35 @@ exclude = ["android-app"]
 version = "0.1.0"
 edition = "2024"
 
+# Debug info is the reason a `cargo test --workspace` here was taking half
+# an hour, and it is worth the paragraph. Measured 2026-09-08: with rustc's
+# default `debug = true`, linking this workspace's test binaries wrote
+# **~54 GB** (one single test binary's linker wrote 16.9 GB) and left an
+# **88 GB** `target/`. Eight test binaries each statically link the whole
+# wgpu + naga + winit + parley graph, and at the default every one of them
+# gets a full copy of that graph's DWARF written into it. On a btrfs at 83%
+# full the linkers then sat in `handle_reserve_ticket` -- uninterruptible,
+# waiting on space reservation -- at about 20 MB/s between them, which is
+# what "the tests are slow" actually was. Not CPU: the machine was 87% idle
+# throughout.
+#
+# `line-tables-only` keeps what is actually read from a backtrace -- the
+# file and line of every frame, which is what a panicking test prints and
+# what gdb needs to name the frames of a segfault. What it gives up is
+# inspecting variables in a debugger; when that is wanted, ask for it on
+# the command line for that one run rather than paying for it on every
+# build:
+#
+#     RUSTFLAGS="-C debuginfo=2" cargo test -p iris --test whatever
+[profile.dev]
+debug = "line-tables-only"
+
+# The tests are what this is really for; `cargo test` uses `dev` for
+# dependencies and `test` for the test targets themselves, so setting only
+# `dev` leaves the eight big binaries at the default.
+[profile.test]
+debug = "line-tables-only"
+
 [workspace.dependencies]
 pollster = "0.4.0"
 winit = "0.30.12"

iris/android-app/Cargo.lock

@@ -744,7 +744,9 @@ checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"
 name = "client-core"
 version = "0.1.0"
 dependencies = [
+ "base64",
  "event-model",
+ "log",
  "pulldown-cmark",
  "serde",
  "serde_json",

iris/android-app/Cargo.toml

@@ -73,6 +73,18 @@ bench = ["transcript-screen", "dep:transcript-fixture", "dep:libc", "dep:tokio"]
 
 [profile.release]
 panic = "abort"
+# Measured 2026-09-07 (docs/RUST.md's "APK size" subsection): together these
+# take libmain.so from 18,546,488 to 11,193,608 bytes (-39.7%) and the APK
+# from 20,678,956 to 13,326,076 bytes (-35.5%), arm64-v8a release. `strip`
+# also works around AGP's own stripReleaseDebugSymbols failing silently on
+# this .so ("packaging them as they are"). `opt-level = "s"` over `"z"`:
+# `z` measured another ~800 KB smaller but was not checked against iris's
+# own frame-time bench, so it is not worth the unmeasured risk -- see the
+# doc for the number and the follow-up this leaves.
+strip = true
+lto = "fat"
+codegen-units = 1
+opt-level = "s"
 
 [profile.dev]
 panic = "abort"

iris/android-app/app/src/main/AndroidManifest.xml

@@ -24,8 +24,45 @@
                 <category android:name="android.intent.category.LAUNCHER" />
             </intent-filter>
 
+            <!-- The enrollment link Dev Updater's Enroll button opens
+                 (what `ai-server` mints), the same one the Compose app
+                 in `app/` registers: which app answers it is the phone
+                 owner's choice at the moment of the tap, and both being
+                 offered is the intended behaviour rather than a clash.
+                 BROWSABLE so a link tapped in another app reaches here,
+                 and `android:host` so this app is not offered for every
+                 aiapp:// URI a future route invents. -->
+            <intent-filter>
+                <action android:name="android.intent.action.VIEW" />
+                <category android:name="android.intent.category.DEFAULT" />
+                <category android:name="android.intent.category.BROWSABLE" />
+                <data android:scheme="aiapp" android:host="enroll" />
+            </intent-filter>
+
             <meta-data android:name="android.app.lib_name" android:value="main" />
         </activity>
+
+        <!-- This app's own recent log, for Dev Updater to read on the
+             phone. Iris runs these builds with no adb, and Android
+             forbids one app reading another's logcat, so this is the
+             only way a log::info! here reaches her. The shape is Dev
+             Updater's contract (its README.md, "An app's own log"), not
+             something invented for this app.
+
+             The authority carries ${applicationId}, so the bench package
+             and the ordinary one each get their own and neither can read
+             the other's log. Exported, because the whole point is
+             another app reading it, and guarded by a permission Dev
+             Updater declares at protectionLevel="normal" (a signature
+             permission is not available: the two apps are signed with
+             different locally generated keys). Read-only: insert,
+             update and delete throw. -->
+        <provider
+            android:name=".DevLogProvider"
+            android:authorities="${applicationId}.devlog"
+            android:exported="true"
+            android:readPermission="dev.updater.permission.READ_DEVLOG" />
+
     </application>
 
 </manifest>

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

@@ -0,0 +1,193 @@
+package dev.iris.android.demo;
+
+import android.content.ContentProvider;
+import android.content.ContentValues;
+import android.content.UriMatcher;
+import android.database.Cursor;
+import android.database.MatrixCursor;
+import android.net.Uri;
+
+/**
+ * This app's own recent log, exposed on the device.
+ *
+ * Iris runs these builds on a phone with no {@code adb}, and Android
+ * forbids one app reading another's {@code logcat} -- so nothing outside
+ * this process can recover what it wrote. The process already keeps a
+ * bounded copy of its log (Rust: {@code client_core::log_ring}); this
+ * hands it to Dev Updater, which is on the same phone, so it needs no
+ * tunnel, no token and no second enrolment.
+ *
+ * <p>The shape is <em>Dev Updater's contract</em>, not something invented
+ * here -- see that project's {@code README.md}, "An app's own log". Any
+ * app it delivers can implement the same and get the same Runtime tab.
+ * Two paths:
+ *
+ * <ul>
+ *   <li>{@code lines?since=<seq>} -- every held line with a sequence at or
+ *       after {@code since}, oldest first.
+ *   <li>{@code status} -- one row: how many lines are held, how many the
+ *       ring's own bound has dropped, and the newest sequence ({@code -1}
+ *       for a log nothing has been written to, which is also how a reader
+ *       notices this process restarted).
+ * </ul>
+ *
+ * <p>Read-only: there is nothing here for anyone else to change, so the
+ * three writing methods throw rather than silently doing nothing.
+ *
+ * <p>The authority is {@code <applicationId>.devlog}, filled in from
+ * Gradle so the bench build and the ordinary one each get their own and
+ * neither can read the other's. Read access is guarded by
+ * {@code dev.updater.permission.READ_DEVLOG}, declared in the manifest.
+ *
+ * <p>No {@code notifyChange}: the ring is filled by a {@code log::Log}
+ * backend on whatever thread logged, and giving that a way to reach a
+ * provider would mean plumbing a callback through {@code client-core} for
+ * every platform. Dev Updater polls while its tab is open, which its
+ * contract says it does precisely so implementing this stays cheap.
+ */
+public final class DevLogProvider extends ContentProvider {
+    static {
+        // The provider is created before any activity, so it cannot rely
+        // on MainActivity's own load. Loading twice is a no-op.
+        System.loadLibrary("main");
+    }
+
+    /** Matches {@link #nativeLinesSince}'s flat answer. Both sides say it once. */
+    private static final int FIELDS_PER_LINE = 5;
+
+    private static final String[] LINE_COLUMNS = {"seq", "t_ms", "level", "target", "message"};
+    private static final String[] STATUS_COLUMNS = {"held", "dropped", "newest_seq"};
+
+    private static final int LINES = 1;
+    private static final int STATUS = 2;
+
+    private UriMatcher matcher;
+
+    /** Every held line, {@link #FIELDS_PER_LINE} strings each, oldest first. */
+    private static native String[] nativeLinesSince(long since);
+
+    /** Three strings: held, dropped, newest sequence. */
+    private static native String[] nativeStatus();
+
+    /**
+     * Tells the Rust side which authority this build registered under, so
+     * the diagnostics pane can name somewhere a reader can actually query
+     * -- and so "declared but never created" is a state it can say. Only
+     * the provider knows it was instantiated; Android creates one lazily.
+     *
+     * <p>The files directory goes with it because <em>this is usually the
+     * only thing running</em>: after the app has died, Dev Updater's query
+     * starts the process for the provider alone, with no activity, so
+     * {@code MainActivity.nativeSetFilesDir} is never called and the line
+     * the panic hook left on disk is never replayed into the ring. That is
+     * exactly the run whose log somebody wants.
+     */
+    private static native void nativeReady(String authority, String filesDir);
+
+    @Override
+    public boolean onCreate() {
+        // The authority is not a constant here: it is derived from this
+        // build's applicationId, so the bench package and the ordinary one
+        // do not share one. Read back from the manifest rather than
+        // recomposed, so there is one answer to what it is.
+        String authority = getContext().getPackageName() + ".devlog";
+        matcher = new UriMatcher(UriMatcher.NO_MATCH);
+        matcher.addURI(authority, "lines", LINES);
+        matcher.addURI(authority, "status", STATUS);
+        nativeReady(authority, getContext().getFilesDir().getAbsolutePath());
+        return true;
+    }
+
+    @Override
+    public Cursor query(
+            Uri uri,
+            String[] projection,
+            String selection,
+            String[] selectionArgs,
+            String sortOrder) {
+        switch (matcher.match(uri)) {
+            case LINES:
+                return lines(sinceOf(uri));
+            case STATUS:
+                return status();
+            default:
+                // Null rather than an exception: an unknown path is a
+                // reader asking for something this app does not have, and
+                // the contract's own answer for that is no cursor.
+                return null;
+        }
+    }
+
+    /**
+     * {@code ?since=} as a number, or 0 for a reader starting from the
+     * beginning. A value that is not a number is treated as 0 rather than
+     * refused -- what a caller wants from a malformed cursor is the log,
+     * not a stack trace about the query string.
+     */
+    private static long sinceOf(Uri uri) {
+        String since = uri.getQueryParameter("since");
+        if (since == null) {
+            return 0;
+        }
+        try {
+            return Long.parseLong(since);
+        } catch (NumberFormatException ignored) {
+            return 0;
+        }
+    }
+
+    private static Cursor lines(long since) {
+        String[] fields = nativeLinesSince(since);
+        if (fields == null) {
+            return null;
+        }
+        MatrixCursor cursor = new MatrixCursor(LINE_COLUMNS, fields.length / FIELDS_PER_LINE);
+        for (int at = 0; at + FIELDS_PER_LINE <= fields.length; at += FIELDS_PER_LINE) {
+            cursor.addRow(
+                    new Object[] {
+                        Long.parseLong(fields[at]),
+                        Long.parseLong(fields[at + 1]),
+                        fields[at + 2],
+                        fields[at + 3],
+                        fields[at + 4],
+                    });
+        }
+        return cursor;
+    }
+
+    private static Cursor status() {
+        String[] fields = nativeStatus();
+        if (fields == null || fields.length != STATUS_COLUMNS.length) {
+            return null;
+        }
+        MatrixCursor cursor = new MatrixCursor(STATUS_COLUMNS, 1);
+        cursor.addRow(
+                new Object[] {
+                    Long.parseLong(fields[0]), Long.parseLong(fields[1]), Long.parseLong(fields[2]),
+                });
+        return cursor;
+    }
+
+    @Override
+    public String getType(Uri uri) {
+        // A MIME type is for something meant to be handed to another app
+        // as data; these rows are read by one reader that knows the
+        // columns. Saying nothing is the honest answer, not a gap.
+        return null;
+    }
+
+    @Override
+    public Uri insert(Uri uri, ContentValues values) {
+        throw new UnsupportedOperationException("this app's log is read-only");
+    }
+
+    @Override
+    public int update(Uri uri, ContentValues values, String selection, String[] selectionArgs) {
+        throw new UnsupportedOperationException("this app's log is read-only");
+    }
+
+    @Override
+    public int delete(Uri uri, String selection, String[] selectionArgs) {
+        throw new UnsupportedOperationException("this app's log is read-only");
+    }
+}

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

@@ -1,6 +1,8 @@
 package dev.iris.android.demo;
 
 import android.app.Activity;
+import android.content.Intent;
+import android.net.Uri;
 import android.os.Build;
 import android.os.Bundle;
 import android.view.WindowInsets;
@@ -20,9 +22,28 @@ public final class MainActivity extends Activity {
         System.loadLibrary("main");
     }
 
+    /**
+     * The app's private directory, where the Rust side keeps its enrollment
+     * (`src/enrollment.rs`). Handed over before the view is built, because
+     * the client the view creates reads the enrollment as it starts.
+     */
+    private static native void nativeSetFilesDir(String path);
+
+    /**
+     * One `aiapp://enroll?host=&port=&token=&ca=` link, as Dev Updater's
+     * Enroll button opens it. Parsed and stored on the Rust side, which is
+     * where the enrollment lives for the desktop app too -- nothing about
+     * the link's format is known here.
+     */
+    private static native void nativeEnroll(String uri);
+
     @Override
     public void onCreate(Bundle state) {
         super.onCreate(state);
+        // Before the view: creating it starts the Rust client, which asks
+        // straight away which server it is enrolled with.
+        nativeSetFilesDir(getFilesDir().getAbsolutePath());
+        handleEnrollmentIntent(getIntent());
         IrisView view = new IrisView(this);
         view.setLayoutParams(new FrameLayout.LayoutParams(
                 FrameLayout.LayoutParams.MATCH_PARENT, FrameLayout.LayoutParams.MATCH_PARENT));
@@ -99,6 +120,35 @@ public final class MainActivity extends Activity {
         });
     }
 
+    /**
+     * A link that arrives while the activity is already up. `singleTop` is
+     * not set, so this is the resumed case only -- the fresh-launch case
+     * goes through `onCreate`'s `getIntent`. `setIntent` so a later
+     * `getIntent` reports the one actually being acted on rather than the
+     * one this activity started with.
+     */
+    @Override
+    protected void onNewIntent(Intent intent) {
+        super.onNewIntent(intent);
+        setIntent(intent);
+        handleEnrollmentIntent(intent);
+    }
+
+    /**
+     * Hands a VIEW intent's URI to the Rust side, which decides whether it
+     * is an enrollment link -- the scheme is checked here only so a launch
+     * intent (which carries no data) costs nothing.
+     */
+    private static void handleEnrollmentIntent(Intent intent) {
+        if (intent == null) {
+            return;
+        }
+        Uri data = intent.getData();
+        if (data != null) {
+            nativeEnroll(data.toString());
+        }
+    }
+
     /** Read one `WindowInsets` and hand it to the Rust side. The only
      * place that reads these fields, so the static dispatch and the
      * animation callback above cannot come to report different things. */

iris/android-app/build-apk.sh

@@ -12,19 +12,29 @@
 #   emulator stays on debug" rule -- pass `release` explicitly for a phone
 #   build). --abi defaults to arm64-v8a (a phone/real device); pass
 #   x86_64 for this checkout's own AVD. --features defaults to
-#   "transcript-screen bench" -- deliberately *without* `force-gles`, unlike
-#   an earlier version of this default. `force-gles` (`iris/Cargo.toml`'s
-#   own doc) exists only to force the emulator off its default software
-#   Vulkan and onto GLES for one specific measurement (RUST.md's I5, "Where
-#   iris's frame time goes") -- it was never meant to reach a real device,
-#   but this script's old default put it in every arm64 build regardless,
-#   so the P0 bench APK delivered to Iris's phone forced GLES there too.
-#   That is the named hypothesis in RUST.md's P0 box ("iris bench crash on
-#   the phone, 2026-09-06"): a real Vulkan driver is what a phone should
-#   run, and GLES is the backend the same box's own SwiftShader finding
-#   already flagged as the fragile one for this shader's storage buffers.
-#   Pass `--features "transcript-screen force-gles bench"` explicitly for
-#   an emulator backend-isolation run; never for a build meant for a phone.
+#   "transcript-screen bench" -- deliberately *without* `force-gles`, and
+#   nothing should add it back for the emulator's sake.
+#
+#   **The emulator does not need a GLES build, because it has no hardware
+#   Vulkan to be steered away from** (docs/RUST.md, "What the emulator
+#   gives a GPU app", 2026-09-08): its guest's only Vulkan is SwiftShader
+#   in software, its GLES is the host's real GPU through virgl, and iris's
+#   own runtime fallback -- `Backends::PRIMARY`, no adapter, rebuild on
+#   `Backends::GL` -- takes an ordinary build there by itself. So the
+#   emulator and the phone run the *same binary* and differ only in what
+#   that binary finds, which is the whole point: a build flag that changed
+#   the backend would mean the thing measured here is not the thing
+#   shipped.
+#
+#   `force-gles` (`iris/Cargo.toml`'s own doc) pins the backend at compile
+#   time for a backend-isolation measurement (RUST.md's I5, "Where iris's
+#   frame time goes"), and the desktop is the better place to run it now
+#   (`run-headless.sh ... --features iris/force-gles`). It was never meant
+#   to reach a real device, but this script's old default put it in every
+#   arm64 build regardless, so the P0 bench APK delivered to Iris's phone
+#   forced GLES there too -- the named hypothesis in RUST.md's P0 box
+#   ("iris bench crash on the phone, 2026-09-06"). Never pass it for a
+#   build meant for a phone.
 set -eu
 cd "$(dirname "$0")"
 
@@ -57,6 +67,18 @@ export ANDROID_NDK_HOME="$NDK_DIR"
 # finds -- a debug x86_64 emulator build left behind made an arm64 "release"
 # 339 MB on 2026-09-06.
 rm -rf app/src/main/jniLibs
+# ...and Gradle's own copy of them, which `rm -rf jniLibs` does not reach.
+# `mergeReleaseNativeLibs` is *up to date* against its cached inputs, so a
+# build that switches ABI packages the previous ABI: an `--abi x86_64`
+# release APK containing `lib/arm64-v8a/libmain.so` installed fine and
+# aborted at startup with `Could not get adapter!: NotFound {
+# active_backends: VULKAN }` under libndk_translation -- which reads
+# exactly like the phone's own Vulkan problem and is nothing of the kind.
+# Scoped to the merge task's directory rather than all of `app/build`, so
+# an ABI change costs the native merge and not the whole Gradle build.
+rm -rf app/build/intermediates/merged_native_libs \
+       app/build/intermediates/stripped_native_libs \
+       app/build/intermediates/merged_jni_libs
 echo "build-apk.sh: cargo ndk -t $ABI build ${BUILD_TYPE:+(${BUILD_TYPE})} --features \"$FEATURES\""
 if [ "$BUILD_TYPE" = "release" ]; then
     cargo ndk -t "$ABI" -P 29 -o app/src/main/jniLibs/ build --release --features "$FEATURES"

iris/android-app/build.rs

@@ -1,100 +0,0 @@
-// Only does anything under the `transcript-screen` feature (RUST.md's I5
-// Android integration) -- the plain tabs build (I2/I4) needs none of this
-// and stays untouched, same reasoning as the feature gate in Cargo.toml.
-//
-// Bakes the sandbox server's host, port, token and pinned CA in at build
-// time, the same way `app/androidApp/build.gradle.kts`'s
-// `GeneratePinnedCert` task bakes the CA for the Compose app -- see that
-// file's comment for why reading the machine's own certificate at build
-// time is the right trust boundary. This build additionally bakes the
-// host/port/token, which the Compose app does not: that app enrolls at
-// runtime from a scanned QR/deep link, and a from-scratch enrollment UI
-// (Keystore-sealed token storage, a QR/link scanner) is real, separate
-// scope this integration does not need to build to answer RUST.md's
-// question -- there is nothing here yet resembling `ServerConfig.kt`. So
-// this is a **deliberate simplification for this rig only**: an APK built
-// this way is good for exactly the emulator/server pair that built it, and
-// must never be treated as a template for a real enrollment flow. Recorded
-// in RUST.md's I5 box rather than left to be rediscovered.
-use std::path::PathBuf;
-
-fn main() {
-    if std::env::var_os("CARGO_FEATURE_TRANSCRIPT_SCREEN").is_none() {
-        return;
-    }
-    // P0's bench build (docs/RUST.md) opens the checked-in fixture with no
-    // server at all -- `bench_client.rs` never references the `pinned`
-    // module this generates, so requiring a live server's host/port/token/
-    // CA to build it (as plain `transcript-screen` does, below) would be a
-    // pointless requirement for a build that talks to nothing.
-    if std::env::var_os("CARGO_FEATURE_BENCH").is_some() {
-        return;
-    }
-    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_HOST");
-    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_PORT");
-    println!("cargo:rerun-if-env-changed=AI_APP_TRANSCRIPT_TOKEN");
-    println!("cargo:rerun-if-env-changed=AI_APP_CA");
-    println!("cargo:rerun-if-env-changed=XDG_CONFIG_HOME");
-
-    let host = require_env(
-        "AI_APP_TRANSCRIPT_HOST",
-        "the sandbox server's host as the emulator reaches it, e.g. 10.0.2.2",
-    );
-    let port = require_env(
-        "AI_APP_TRANSCRIPT_PORT",
-        "the sandbox server's port -- app/ui-sandbox.sh's start banner prints it",
-    );
-    let token = require_env(
-        "AI_APP_TRANSCRIPT_TOKEN",
-        "the bearer token -- ~/.config/ai-app/sandbox-token, or the start banner's enrollment link",
-    );
-
-    let ca_path = std::env::var_os("AI_APP_CA")
-        .map(PathBuf::from)
-        .unwrap_or_else(|| {
-            let base = std::env::var_os("XDG_CONFIG_HOME")
-                .map(PathBuf::from)
-                .unwrap_or_else(|| {
-                    let home = std::env::var_os("HOME").expect("HOME must be set");
-                    PathBuf::from(home).join(".config")
-                });
-            base.join("ai-app").join("certs").join("ca.pem")
-        });
-    let ca_pem = std::fs::read_to_string(&ca_path).unwrap_or_else(|e| {
-        panic!(
-            "no CA certificate at {} ({e}).\n\
-             Start ai-server (or app/ui-sandbox.sh) once on this machine first -- it \
-             generates the CA this build pins. Set AI_APP_CA=/path/to/ca.pem to build \
-             against a different one.",
-            ca_path.display()
-        )
-    });
-    let ca_pem = ca_pem.trim();
-    if !ca_pem.starts_with("-----BEGIN CERTIFICATE-----") {
-        panic!("{} is not a PEM certificate.", ca_path.display());
-    }
-
-    let out_dir = PathBuf::from(std::env::var_os("OUT_DIR").unwrap());
-    let generated = format!(
-        "// Generated by build.rs from {host}:{port} and {ca}. Do not edit.\n\
-         pub const HOST: &str = {host_lit:?};\n\
-         pub const PORT: u16 = {port};\n\
-         pub const TOKEN: &str = {token_lit:?};\n\
-         pub const CA_PEM: &str = {ca_lit:?};\n",
-        host = host,
-        port = port
-            .parse::<u16>()
-            .unwrap_or_else(|e| panic!("AI_APP_TRANSCRIPT_PORT={port:?} is not a u16: {e}")),
-        ca = ca_path.display(),
-        host_lit = host,
-        token_lit = token,
-        ca_lit = ca_pem,
-    );
-    std::fs::write(out_dir.join("pinned_config.rs"), generated).unwrap();
-}
-
-fn require_env(name: &str, what: &str) -> String {
-    std::env::var(name).unwrap_or_else(|_| {
-        panic!("{name} must be set to build the transcript-screen feature -- {what}")
-    })
-}

iris/android-app/run-bench.sh

@@ -46,6 +46,21 @@ 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
 
+# Which adapter drew, before any number is printed. The emulator is a GLES
+# machine -- its guest has no hardware Vulkan (docs/RUST.md, "What the
+# emulator gives a GPU app") -- so iris's runtime fallback lands on `Gl`,
+# and `Gl (... virgl ...)` is the host's real GPU while `Gl (...
+# SwiftShader ...)` is the CPU. Those two produce frame times an order of
+# magnitude apart and are otherwise indistinguishable in this report, so
+# the line is printed rather than left in logcat for somebody to think of.
+ADAPTER=$(adb -s "$SERIAL" logcat -d -s iris-android-app:I 2>/dev/null \
+    | sed -n 's/.*\(iris renderer: .*\)/\1/p' | tail -1)
+if [ -n "$ADAPTER" ]; then
+    echo "run-bench.sh: $ADAPTER"
+else
+    echo "run-bench.sh: no 'iris renderer:' line in logcat -- cannot say what drew this run" >&2
+fi
+
 # 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

iris/android-app/src/app_log.rs

@@ -0,0 +1,186 @@
+//! The platform half of this app's logging: what
+//! `client_core::log_ring` needs that only Android can supply, which is
+//! `android_logger` as the logger to forward to and nothing else.
+//!
+//! Everything general -- the ring, its bounds, the `log::Log` backend --
+//! is in `client-core`, shared with the desktop app (AGENTS.md's sharing
+//! rule).
+//!
+//! **Why an app carries its own log at all**: Iris tests these builds on a
+//! GrapheneOS phone with no `adb`, and Android forbids one app reading
+//! another's `logcat`. Nothing outside this process can recover what it
+//! wrote, so the process keeps a copy -- and hands it to Dev Updater on
+//! the same phone through `devlog`'s `ContentProvider`. See
+//! `docs/DECISIONS.md`, 2026-09-07.
+
+use client_core::log_ring::{self, LogRing};
+
+/// Installs the ring in front of `android_logger`, so `logcat` still sees
+/// exactly what it saw before and the ring sees it too.
+///
+/// Called once, from `JNI_OnLoad`. A second call is refused by `log`
+/// itself; the message says which caller, since two initialisation paths
+/// is a programmer error rather than something to recover from.
+pub fn install(max_level: log::LevelFilter) {
+    let inner = android_logger::AndroidLogger::new(
+        android_logger::Config::default()
+            .with_max_level(max_level)
+            .with_tag("iris-android-app"),
+    );
+    if log_ring::install_process_logger(
+        Box::new(inner),
+        max_level,
+        iris::diagnostics::trace_enabled,
+    )
+    .is_err()
+    {
+        // Not a panic: a logger already installed means logging works,
+        // just without the ring, and taking the app down over a
+        // diagnostic would be worse than the diagnostic being missing.
+        // The line goes through whatever logger did win.
+        log::warn!("iris app log: a logger was already installed, so there is no ring");
+    }
+    install_panic_hook();
+}
+
+/// The process's ring -- what `Copy report` appends, what the diagnostics
+/// pane counts, and what `devlog`'s provider hands to Dev Updater.
+pub fn ring() -> &'static LogRing {
+    log_ring::process_ring()
+}
+
+/// Only the bench build has a diagnostics pane to put this in; the
+/// transcript build's screen is the app's own and has no room for a
+/// readout. Gated rather than left dead so the build stays warning-clean.
+#[cfg(feature = "bench")]
+/// Two lines for the diagnostics pane: how much of this app's log is held,
+/// and where it can be read from.
+///
+/// The second names the provider's authority rather than saying "logging
+/// is on", so a screenshot of this pane is enough to tell whether the
+/// contract is live and which package's log it is -- the bench build and
+/// the ordinary one have different ones.
+pub fn diagnostics_line() -> String {
+    let where_to_read = match crate::devlog::authority() {
+        Some(authority) => format!("devlog provider: content://{authority}"),
+        // Not "off": Android creates a provider lazily, so this is what
+        // "nobody has asked for it yet" looks like, and it is a different
+        // thing from a build that does not have one.
+        None => "devlog provider: declared, not created yet".to_string(),
+    };
+    format!("{}\n{where_to_read}", ring().summary())
+}
+
+/// Where the panic hook leaves its report, under the app's private
+/// directory. Read back and dropped by [`set_crash_dir`] on the next
+/// start.
+const CRASH_FILE: &str = "last-panic.txt";
+
+/// How many of the dying run's own log lines the panic hook saves with
+/// the panic, and [`set_crash_dir`] replays.
+///
+/// The panic's message and location say *what* broke; these say what the
+/// app was doing on the way there, which is the half that is otherwise
+/// unrecoverable -- the ring is memory only, so an abort takes every line
+/// before the panic with it. Bounded rather than the whole ring because
+/// this is written by a hook on a process that is about to die, and
+/// because the replay pushes each line into the new run's ring, where an
+/// unbounded paste would evict the run that is actually being watched.
+const CRASH_CONTEXT_LINES: usize = 80;
+
+/// The target the replayed context lines carry, so a reader can tell a
+/// line from the run that died from one this run wrote. They keep their
+/// original timestamp and level inside the text, which is why the level
+/// they are re-pushed at is not meaningful and the target has to be.
+const PREVIOUS_RUN_TARGET: &str = "previous_run";
+
+static CRASH_PATH: std::sync::OnceLock<std::path::PathBuf> = std::sync::OnceLock::new();
+
+/// Installs a `log`-level panic hook, so a panic's message and location
+/// reach the ring and `logcat` rather than only the tombstone.
+///
+/// **Why this is needed at all**: these builds are `panic = "abort"`
+/// (`Cargo.toml`), and the default hook writes to `stderr` plus
+/// `android_set_abort_message` -- the crash report. Iris runs these on a
+/// phone with no `adb`, so the crash report is exactly the surface she
+/// cannot read, and an `assert!` that fired said nothing anywhere she
+/// could see it. Routing it through `log::error!` puts it in front of
+/// `android_logger` *and* in the ring `devlog`'s provider hands to Dev
+/// Updater.
+///
+/// The ring is memory only, so after an abort the process that holds it
+/// is gone -- hence the file half. [`set_crash_dir`] replays it.
+fn install_panic_hook() {
+    let previous = std::panic::take_hook();
+    std::panic::set_hook(Box::new(move |info| {
+        let where_at = match info.location() {
+            Some(at) => format!("{}:{}:{}", at.file(), at.line(), at.column()),
+            None => "an unknown location".to_string(),
+        };
+        // `info`'s own `Display` repeats the location and a newline;
+        // the payload alone keeps this to the one line the ring wants.
+        let message = info.payload_as_str().unwrap_or("Box<dyn Any>");
+        let line = format!("iris panic at {where_at}: {message}");
+        log::error!("{line}");
+        if let Some(path) = CRASH_PATH.get() {
+            // The panic line first, then what the app was doing before
+            // it: one file, split again on that first newline by
+            // `set_crash_dir`.
+            let context = ring()
+                .try_tail_text(CRASH_CONTEXT_LINES)
+                // Said rather than left empty, so "the ring was locked as
+                // we died" cannot be read as "nothing had been logged".
+                .unwrap_or_else(|| {
+                    "(the log ring was locked as this run died; no context)".to_string()
+                });
+            // Best effort by design: a panic is already the failure, and
+            // failing to record it must not become a second one.
+            let _ = std::fs::write(path, format!("{line}\n{context}"));
+        }
+        previous(info);
+    }));
+}
+
+/// Tells the panic hook where to leave its report, and replays the report
+/// a previous run left there into the ring before deleting it.
+///
+/// Called from **both** `MainActivity.nativeSetFilesDir` and
+/// `DevLogProvider.nativeReady` -- whichever of the two runs first in
+/// this process, since after a crash Dev Updater's query starts the
+/// process for the provider alone and no activity ever runs. Safe to call
+/// twice: the file is gone after the first, so the second finds nothing
+/// and says nothing. The panic itself is replayed at `error` level and
+/// says it is from the previous run, so a crash loop shows the reason it
+/// is looping in the Runtime tab of the run that is still up.
+pub fn set_crash_dir(dir: &std::path::Path) {
+    let path = dir.join(CRASH_FILE);
+    if let Ok(previous) = std::fs::read_to_string(&path) {
+        // Delete before replaying rather than after: a replay that itself
+        // panicked would otherwise leave the file to be replayed again on
+        // every start, and a crash loop nothing can get out of is worse
+        // than one report lost.
+        let _ = std::fs::remove_file(&path);
+        replay_crash(&previous);
+    }
+    let _ = CRASH_PATH.set(path);
+}
+
+/// Puts a previous run's report back in the ring: its context lines in
+/// the order they happened, then the panic itself.
+///
+/// Chronological, so the Runtime tab reads as one story -- the lines that
+/// led to the crash, then the crash, then this run. The context goes in
+/// through `LogRing::push` rather than through `log::info!` so it is not
+/// stamped with this run's clock: each line already carries the time and
+/// level it was written at, and [`PREVIOUS_RUN_TARGET`] is what says
+/// whose run it was.
+fn replay_crash(report: &str) {
+    let (panic_line, context) = report.split_once('\n').unwrap_or((report, ""));
+    for line in context.lines().filter(|line| !line.is_empty()) {
+        ring().push(log::Level::Info, PREVIOUS_RUN_TARGET, line.to_string());
+    }
+    log::error!(
+        "iris app log: the previous run died -- {}",
+        panic_line.trim()
+    );
+}

iris/android-app/src/bench_client.rs

@@ -72,13 +72,15 @@ const TYPE_CHAR_MS: u64 = 50;
 const KEYBOARD_CYCLES: usize = 5;
 const KEYBOARD_WAIT_MS: u64 = 1_000;
 
-/// One animation step's target cadence -- close enough to 60Hz that a
-/// 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;
+/// How often this file *asks a question of* the running app -- polls for
+/// a `ctx.update` closure's answer, or for a fling to have settled.
+///
+/// It is not an animation cadence and nothing on screen moves at this
+/// rate: the frame loop advances animations once per frame at the
+/// display's own refresh (`UiData::tick_animations`). It used to be both,
+/// and that is the defect Iris reported on 2026-09-08 -- see
+/// `wait_for_fling_settle`.
+const POLL_MS: u64 = 16;
 
 /// How much of the screen a *filled* benchmark report may take before it
 /// scrolls instead of growing -- roughly a third of a phone screen, the
@@ -259,7 +261,7 @@ impl AndroidAppState for BenchClient {
         let font = rsc.ui.text.font_diagnostics();
         log::info!(
             "iris fonts: {} families found, default={:?} mono={:?}, resolved regular={:?} \
-             bold={:?} italic={:?} mono={:?}",
+             bold={:?} italic={:?} mono={:?}, icons={:?}",
             font.families_found,
             font.default_family,
             font.default_mono_family,
@@ -267,6 +269,7 @@ impl AndroidAppState for BenchClient {
             font.bold_resolved,
             font.italic_resolved,
             font.mono_resolved,
+            font.icon_family,
         );
 
         let mut client = Self {
@@ -415,6 +418,29 @@ const KEYBOARD_DIAGNOSTICS_DELAY_MS: u64 = 500;
 
 type Rsc = AndroidRsc<BenchClient>;
 
+/// What a report says about the `iris::input`/`iris::frame` trace, from
+/// the flag read at the start of what is being reported and again at the
+/// end.
+///
+/// Three answers rather than two. Those lines are default-off and the
+/// switch that turns them on is on screen while a benchmark runs, so
+/// "somebody moved it half way through" is a state that actually happens
+/// -- and reported as either "on" or "off" it is a confident sentence
+/// about a log that only covers part of the run. The "on" wording also
+/// says what it costs, because a traced run fills the ring in seconds and
+/// a reader looking at a log with nothing else in it should know why.
+fn trace_line(at_start: bool, at_end: bool) -> String {
+    match (at_start, at_end) {
+        (true, true) => "input/frame trace: on (iris::input and iris::frame lines are in \
+                         the app log, and a traced run fills the ring in seconds)"
+            .to_string(),
+        (false, false) => "input/frame trace: off".to_string(),
+        _ => "input/frame trace: switched during this run, so those lines cover only part \
+              of it"
+            .to_string(),
+    }
+}
+
 /// The header row's own backdrop -- see `bench_controls`'s doc comment on
 /// why it needs one at all. A dark neutral rather than pure black
 /// (`android::render::CLEAR_COLOR`) so the row reads as a distinct panel
@@ -442,6 +468,25 @@ const HEADER_SURFACE: UiColor = UiColor::new(28, 28, 34, 255);
 /// pixels) to `dp(...)` (IRIS_TODO.md's density-independent length unit),
 /// so the row's reserved height in the outer `Span::DOWN`
 /// (`AndroidAppState::new`) matches what is actually painted.
+/// The size every label in the header row is drawn at.
+///
+/// One constant for all four rather than a number per button, because the
+/// whole row has to be sized together. Adding the trace switch made four
+/// controls too wide for one row at the size three had used (18), and an
+/// earlier pass shrank this constant to 13 to make them fit -- exactly
+/// what UI_RULES forbids ("never shrink text to make it fit": a label a
+/// different size from its neighbours elsewhere in the app for a reason
+/// the reader cannot see). The fix is [`bench_controls`]'s two rows
+/// instead, which leaves room to put this back. Whoever adds a fifth
+/// control reconsiders the row split, not this number.
+const HEADER_TEXT: f32 = 18.0;
+
+/// The height of one row of header controls, in dp. `bench_controls` now
+/// stacks two of these, so this is the one number to change if a control's
+/// own padding ever changes instead of `dp(56)` and `dp(112)` needing to
+/// be kept in sync by hand.
+const HEADER_ROW_HEIGHT_DP: f32 = 56.0;
+
 fn bench_controls(rsc: &mut Rsc, top_pad: f32) -> StrongWidget {
     let run_rect = rect(Color::rgb(40, 70, 40))
         .on(
@@ -453,7 +498,9 @@ fn bench_controls(rsc: &mut Rsc, top_pad: f32) -> StrongWidget {
         .label("Run benchmark");
     let run = (
         run_rect,
-        wtext("Run benchmark").size(18).text_align(Align::CENTER),
+        wtext("Run benchmark")
+            .size(HEADER_TEXT)
+            .text_align(Align::CENTER),
     )
         .stack()
         .pad(dp(8))
@@ -462,14 +509,16 @@ fn bench_controls(rsc: &mut Rsc, top_pad: f32) -> StrongWidget {
     let copy_rect = rect(Color::rgb(50, 50, 60))
         .on(
             CursorSense::click(),
-            |ctx: EventIdCtx<'_, Rsc, _, _>, _rsc: &mut Rsc| {
-                ctx.state.copy_report();
+            |ctx: EventIdCtx<'_, Rsc, _, _>, rsc: &mut Rsc| {
+                ctx.state.copy_report(rsc);
             },
         )
         .label("Copy report");
     let copy = (
         copy_rect,
-        wtext("Copy report").size(18).text_align(Align::CENTER),
+        wtext("Copy report")
+            .size(HEADER_TEXT)
+            .text_align(Align::CENTER),
     )
         .stack()
         .pad(dp(8))
@@ -485,17 +534,61 @@ fn bench_controls(rsc: &mut Rsc, top_pad: f32) -> StrongWidget {
         .label("Diagnostics");
     let diagnostics = (
         diag_rect,
-        wtext("Diagnostics").size(18).text_align(Align::CENTER),
+        wtext("Diagnostics")
+            .size(HEADER_TEXT)
+            .text_align(Align::CENTER),
     )
         .stack()
         .pad(dp(8))
         .add(rsc);
 
-    let buttons = (run, copy, diagnostics).span(Dir::RIGHT).add(rsc);
+    // A switch rather than a button, so its own appearance says which
+    // state it is in: the two `iris::input`/`iris::frame` targets are
+    // default-off (`iris::diagnostics`'s module doc) because a 120Hz
+    // session fills the 2000-line ring in seconds, so "is it on right
+    // now" is the question somebody has while looking at a log that is
+    // either full of trace or has none.
+    //
+    // The visible text carries the state and the accessibility label does
+    // not, deliberately: the label is also what `run-bench.sh` taps by
+    // name, and a control that renames itself when pressed is one no
+    // script can find twice.
+    let tracing = iris::diagnostics::trace_enabled();
+    let trace_rect = rect(if tracing {
+        Color::rgb(90, 70, 30)
+    } else {
+        Color::rgb(50, 50, 60)
+    })
+    .on(
+        CursorSense::click(),
+        |ctx: EventIdCtx<'_, Rsc, _, _>, rsc: &mut Rsc| {
+            ctx.state.toggle_trace(rsc);
+        },
+    )
+    .label("Trace input and frames");
+    let trace = (
+        trace_rect,
+        wtext(if tracing { "Trace on" } else { "Trace off" })
+            .size(HEADER_TEXT)
+            .text_align(Align::CENTER),
+    )
+        .stack()
+        .pad(dp(8))
+        .add(rsc);
+
+    // Two rows rather than one: four controls at the restored `HEADER_TEXT`
+    // no longer fit a 1080px-wide row (that was the shrink this replaces --
+    // see the constant's own doc). Grouped by what they act on: the first
+    // row starts a benchmark and copies its result; the second is the
+    // diagnostics pane and the switch that decides what it will contain
+    // next time.
+    let row1 = (run, copy).span(Dir::RIGHT).add(rsc);
+    let row2 = (diagnostics, trace).span(Dir::RIGHT).add(rsc);
+    let buttons = (row1, row2).span(Dir::DOWN).add(rsc);
 
     (rect(HEADER_SURFACE), buttons)
         .stack()
-        .height(dp(56))
+        .height(dp(2.0 * HEADER_ROW_HEIGHT_DP))
         .pad(Padding::top(top_pad))
         .add_strong(rsc)
         .any()
@@ -526,6 +619,25 @@ impl BenchClient {
         self.last_report = Some(report);
     }
 
+    /// Turns the `iris::input`/`iris::frame` trace on or off, redraws the
+    /// switch that says so, and shows the pane that now reports it.
+    ///
+    /// Showing the pane is the point rather than a convenience: this is a
+    /// control whose whole effect is on what a *later* report says, so
+    /// putting the state on screen at the moment of the press is the only
+    /// thing that distinguishes it from a button that did nothing.
+    fn toggle_trace(&mut self, rsc: &mut Rsc) {
+        let on = !iris::diagnostics::trace_enabled();
+        iris::diagnostics::set_trace(on);
+        log::info!(
+            "iris diagnostics: input/frame trace {}",
+            if on { "on" } else { "off" }
+        );
+        let controls = bench_controls(rsc, self.last_top_pad);
+        (self.top_bar)(rsc).set(controls);
+        self.show_diagnostics(rsc);
+    }
+
     /// The diagnostics report as text, with no side effect on what is on
     /// screen -- shared by the `Diagnostics` button (which shows it) and
     /// the keyboard-open capture (which only logs it), so the two can
@@ -544,7 +656,20 @@ impl BenchClient {
         // logcat on her phone, and "the keyboard does not push the
         // composer up" cannot be told from "the listener never fired"
         // without it (`AndroidUiState::insets_report`).
-        format!("{renderer}\n{}", self.android_state().insets_report())
+        format!(
+            "{renderer}\n{}\n{}\n{}\n{}",
+            trace_line(
+                iris::diagnostics::trace_enabled(),
+                iris::diagnostics::trace_enabled()
+            ),
+            self.android_state().insets_report(),
+            // Which server this build talks to, and what to do when the
+            // answer is "none" -- the bench itself opens a checked-in
+            // fixture and needs no server, so this pane is the only place
+            // an enrolment can be seen to have taken.
+            crate::enrollment::status_line(),
+            crate::app_log::diagnostics_line()
+        )
     }
 
     /// The keyboard's own diagnostics capture -- see `on_insets_changed`'s
@@ -565,16 +690,40 @@ impl BenchClient {
         log::info!("iris keyboard diagnostics:\n{report}");
     }
 
-    fn copy_report(&mut self) {
-        let Some(report) = &self.last_report else {
-            log::info!("iris bench report: nothing to copy -- run the benchmark first");
-            return;
-        };
+    /// Always copies something, and never depends on `Diagnostics` or
+    /// `Run benchmark` having been pressed first (docs/IRIS_TODO.md,
+    /// 2026-09-07 night: "the copy report button seemed impossible to hit
+    /// until I hit the diagnostics one" -- it was silently declining
+    /// instead of reporting where it had failed, the UI_RULES failure "a
+    /// failure is reported where it happened"). With no benchmark run yet,
+    /// it copies the diagnostics pane's own text instead, with a first
+    /// line saying so -- `diagnostics_text` needs no prior button press
+    /// either, so this is never actually empty-handed.
+    fn copy_report(&mut self, rsc: &mut Rsc) {
         let Some(platform) = &self.platform else {
             log::info!("iris bench report: no platform handle, can't reach the clipboard");
             return;
         };
-        if platform.copy_to_clipboard("iris bench report", report) {
+        let report = match self.last_report.clone() {
+            Some(report) => report,
+            None => format!(
+                "no benchmark has run yet -- these are the diagnostics instead:\n\n{}",
+                self.diagnostics_text(rsc)
+            ),
+        };
+        // The ring's tail goes on the clipboard, not the full ring, and
+        // not into the on-screen pane either: the full ring can be over a
+        // thousand lines with tracing on, and pasting that into a phone's
+        // message box was Iris's own "causes a lot of lag" report. The
+        // full ring is still reachable through Dev Updater's Runtime tab
+        // (`devlog`'s provider reads the same ring) -- this only bounds
+        // what gets inlined here.
+        let report = format!(
+            "{report}\n\n=== app log ({}) ===\n{}",
+            crate::app_log::ring().summary(),
+            crate::app_log::ring().tail_text(client_core::log_ring::COPY_REPORT_TAIL_LINES)
+        );
+        if platform.copy_to_clipboard("iris bench report", &report) {
             log::info!("iris bench report: copied to clipboard");
         } else {
             log::info!("iris bench report: clipboard copy failed");
@@ -603,6 +752,11 @@ impl BenchClient {
             .and_then(|p| p.refresh_rate_hz())
             .unwrap_or(60.0);
         let cpu_start = process_cpu_ms();
+        // Read at the start as well as the end, because the switch is on
+        // screen while a run is going: a report that only asked afterwards
+        // would say "on" about a run whose first half has no trace in it
+        // -- the inferred answer presented as the measured one.
+        let trace_at_start = iris::diagnostics::trace_enabled();
         let run_started_at = Instant::now();
 
         rsc.spawn_task(async move |mut ctx| {
@@ -700,9 +854,11 @@ impl BenchClient {
                     "  type: {} characters inserted then deleted, one per {TYPE_CHAR_MS}ms",
                     TYPE_TEXT.chars().count()
                 );
+                let traced = trace_line(trace_at_start, iris::diagnostics::trace_enabled());
                 let report = format!(
-                    "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}"
+                    "iris bench report\n{traced}\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);
@@ -727,7 +883,7 @@ impl BenchClient {
 /// 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
+/// Polls rather than assuming one `POLL_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>,
@@ -747,7 +903,7 @@ where
         if let Ok(value) = rx.try_recv() {
             return value;
         }
-        tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS)).await;
+        tokio::time::sleep(Duration::from_millis(POLL_MS)).await;
     }
 }
 
@@ -775,13 +931,14 @@ async fn run_fling_phase(
     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;
+    tokio::time::sleep(Duration::from_millis(POLL_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);
+                animate_list(screen.list, rsc);
             }
         });
         redraw.request_redraw();
@@ -794,6 +951,7 @@ async fn run_fling_phase(
         ctx.update(|state: &mut BenchClient, rsc| {
             if let Some(screen) = &state.screen {
                 (screen.list)(rsc).fling(FLING_VELOCITY_PX_S);
+                animate_list(screen.list, rsc);
             }
         });
         redraw.request_redraw();
@@ -802,7 +960,10 @@ async fn run_fling_phase(
     }
     let end = read_anchor_position(ctx, redraw).await;
 
-    format!("start={start} outward={outward} end={end}")
+    // Says how the fling was advanced, because that is what changed on
+    // 2026-09-08 and a report from before then is not comparable: the
+    // phase used to tick the fling itself at ~60Hz.
+    format!("start={start} outward={outward} end={end} ticked=frame-loop")
 }
 
 async fn read_anchor_position(
@@ -816,11 +977,31 @@ async fn read_anchor_position(
     .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.
+/// Register the list with the frame loop, exactly as a finger's own
+/// release does (`transcript_ui::Selection::drag`'s `Released` arm) --
+/// `List::fling` sets a velocity and drives nothing by itself.
+fn animate_list(list: iris::prelude::WeakWidget<iris::prelude::List>, rsc: &mut Rsc) {
+    let id = list.id();
+    rsc.ui_mut().animate(id);
+}
+
+/// Waits for the fling started above to settle, or for
+/// `FLING_SETTLE_CAP_MS` -- 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.
+///
+/// **It observes; it does not drive.** Until 2026-09-08 this loop called
+/// `List::tick_fling` itself every `POLL_MS`, which advanced the
+/// fling in 16ms steps -- so on Iris's 120Hz phone every second frame
+/// redrew the list at a position it had already drawn, and the benchmark
+/// looked distinctly less smooth than the same list under her finger.
+/// That is what she reported that day, and it was the rig rather than the
+/// renderer: a real fling is ticked once per frame by
+/// `UiData::tick_animations`, from the frame callback. So the bench now
+/// starts the fling the way a gesture does (`fling` + `UiData::animate`)
+/// and polls `is_scrolling` to know when it is over, which makes the
+/// phase measure the same path a finger takes. The poll interval is only
+/// how often the *question* is asked and has no bearing on the animation.
 async fn wait_for_fling_settle(
     ctx: &mut iris::task::TaskCtx<Rsc>,
     redraw: &Arc<dyn RequestRedraw>,
@@ -829,14 +1010,14 @@ async fn wait_for_fling_settle(
     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()),
+            Some(screen) => (screen.list)(rsc).is_scrolling(),
             None => false,
         })
         .await;
         if !still_scrolling {
             return;
         }
-        tokio::time::sleep(Duration::from_millis(ANIM_STEP_MS)).await;
+        tokio::time::sleep(Duration::from_millis(POLL_MS)).await;
     }
 }
 

iris/android-app/src/devlog.rs

@@ -0,0 +1,206 @@
+//! The JNI half of `DevLogProvider`: reading this process's own log ring
+//! for a `ContentProvider` that Dev Updater queries.
+//!
+//! **Why**: Iris runs these builds on a phone with no `adb`, and Android
+//! forbids one app reading another's `logcat`, so nothing outside this
+//! process can recover what it wrote. The app already keeps a bounded copy
+//! (`client_core::log_ring`); this is how the copy leaves the process. Dev
+//! Updater is on the same phone, so handing it over needs no tunnel, no
+//! token and no second enrolment -- and it is Dev Updater's own contract
+//! rather than something invented here, so any app it delivers can do the
+//! same (its `README.md`, "An app's own log").
+//!
+//! **Everything general stays in `client-core`** (AGENTS.md's sharing
+//! rule). What is here is only what Android forces: the JNI boundary and
+//! the Java class on the other side of it.
+//!
+//! Both entry points answer a **flat `String[]`** rather than a row of
+//! typed columns. That is the whole of the JNI, and it is one array type
+//! instead of three interleaved ones for a payload the provider is about
+//! to hand back over binder as a `MatrixCursor` anyway; `DevLogProvider`
+//! parses the two numeric fields. Kept flat rather than nested for the
+//! same reason -- an array of arrays is four more JNI calls per line.
+
+use android_view::jni::JNIEnv;
+use android_view::jni::objects::{JClass, JObject, JString};
+use android_view::jni::sys::{jlong, jobjectArray};
+use std::sync::OnceLock;
+
+/// How many `String`s each log line occupies in the flat answer:
+/// `seq`, `t_ms`, `level`, `target`, `message`, in that order. The Java
+/// side has the same constant, and the two are the one place the shape is
+/// written down on each side.
+const FIELDS_PER_LINE: usize = 5;
+
+/// The authority the provider registered itself under, once it has been
+/// created. `None` until then, which is a state worth being able to say:
+/// a provider Android never instantiated and one that is answering look
+/// the same from inside this process otherwise.
+static AUTHORITY: OnceLock<String> = OnceLock::new();
+
+/// Where this app's log can be read from, for the diagnostics pane.
+///
+/// The provider's own answer rather than one composed from the package
+/// name here: what makes the line worth showing is that it names an
+/// authority somebody can actually query, and only the provider knows it
+/// registered.
+#[cfg(feature = "bench")]
+pub fn authority() -> Option<&'static str> {
+    AUTHORITY.get().map(String::as_str)
+}
+
+/// `DevLogProvider.nativeReady` -- the provider announcing the authority
+/// it registered under and the app's private directory, from its own
+/// `onCreate`.
+///
+/// The directory is taken here as well as in
+/// `MainActivity.nativeSetFilesDir` because **the provider is often the
+/// only thing running**: once the app has died, Dev Updater's query
+/// starts the process for the provider alone, so no activity ever runs
+/// and the panic hook's file would never be replayed into the ring. That
+/// is precisely the run whose log is being asked for. Whichever of the
+/// two arrives first does the replay; `set_crash_dir` deletes the file,
+/// so the second finds nothing and says nothing.
+///
+/// # Safety
+/// Called by the JVM with the arguments its `native` declaration names.
+#[unsafe(no_mangle)]
+pub extern "system" fn Java_dev_iris_android_demo_DevLogProvider_nativeReady(
+    mut env: JNIEnv,
+    _class: JClass,
+    authority: JString,
+    files_dir: JString,
+) {
+    // Before the authority line, so the previous run's death is above the
+    // line announcing this one rather than buried under it.
+    #[cfg(feature = "transcript-screen")]
+    if let Some(dir) = string_arg(&mut env, &files_dir) {
+        crate::app_log::set_crash_dir(std::path::Path::new(&dir));
+    }
+    #[cfg(not(feature = "transcript-screen"))]
+    let _ = &files_dir;
+    let Some(authority) = string_arg(&mut env, &authority) else {
+        return;
+    };
+    log::info!("iris devlog: serving this app's log at content://{authority}");
+    let _ = AUTHORITY.set(authority);
+}
+
+/// One `String` argument, or `None` for a null or unreadable one.
+fn string_arg(env: &mut JNIEnv, value: &JString) -> Option<String> {
+    if value.is_null() {
+        return None;
+    }
+    env.get_string(value).ok().map(Into::into)
+}
+
+/// `DevLogProvider.nativeStatus` -- `held`, `dropped`, `newest_seq`, as
+/// three strings.
+///
+/// `newest_seq` is `-1` for a ring nothing has been written to, which is
+/// what tells a reader holding a cursor that this process **restarted**:
+/// the ring is in memory, so a new process starts again at zero and a
+/// stale cursor would otherwise skip everything silently.
+///
+/// Exported by name rather than registered, matching this crate's other
+/// activity-side natives: the mangled name is the whole of what a class
+/// this app owns needs.
+///
+/// # Safety
+/// Called by the JVM with the arguments its `native` declaration names.
+#[unsafe(no_mangle)]
+pub extern "system" fn Java_dev_iris_android_demo_DevLogProvider_nativeStatus(
+    mut env: JNIEnv,
+    _class: JClass,
+) -> jobjectArray {
+    string_array(&mut env, &status_fields())
+}
+
+/// `DevLogProvider.nativeLinesSince` -- every held line with a sequence at
+/// or after `since`, oldest first, [`FIELDS_PER_LINE`] strings each.
+///
+/// Inclusive of `since` because [`client_core::log_ring::LogRing::since`]
+/// is, and one definition of the cursor is what keeps the app's own
+/// uploaded report and this provider describing the same lines.
+///
+/// # Safety
+/// Called by the JVM with the arguments its `native` declaration names.
+#[unsafe(no_mangle)]
+pub extern "system" fn Java_dev_iris_android_demo_DevLogProvider_nativeLinesSince(
+    mut env: JNIEnv,
+    _class: JClass,
+    since: jlong,
+) -> jobjectArray {
+    // A negative cursor is a caller asking for everything, not an error to
+    // take the app down over: the provider is a diagnostic.
+    string_array(&mut env, &line_fields(since.max(0) as u64))
+}
+
+/// The three status numbers, as the provider's row.
+#[cfg(feature = "transcript-screen")]
+fn status_fields() -> Vec<String> {
+    let ring = client_core::log_ring::process_ring();
+    vec![
+        ring.len().to_string(),
+        ring.dropped().to_string(),
+        ring.newest_seq().map_or(-1, |seq| seq as i64).to_string(),
+    ]
+}
+
+/// The tabs demo links no `client-core` and keeps no ring, so it holds
+/// nothing and has never dropped anything -- which is the truth, not a
+/// stand-in. The natives are still exported there, because a `native`
+/// method Java declares and the library does not is an
+/// `UnsatisfiedLinkError` the moment the class loads.
+#[cfg(not(feature = "transcript-screen"))]
+fn status_fields() -> Vec<String> {
+    vec!["0".to_string(), "0".to_string(), "-1".to_string()]
+}
+
+#[cfg(feature = "transcript-screen")]
+fn line_fields(since: u64) -> Vec<String> {
+    let (lines, _next) = client_core::log_ring::process_ring().since(since);
+    let mut fields = Vec::with_capacity(lines.len() * FIELDS_PER_LINE);
+    for line in lines {
+        fields.push(line.seq.to_string());
+        fields.push(line.at_ms.to_string());
+        fields.push(line.level.to_string());
+        fields.push(line.target);
+        fields.push(line.message);
+    }
+    fields
+}
+
+#[cfg(not(feature = "transcript-screen"))]
+fn line_fields(_since: u64) -> Vec<String> {
+    Vec::new()
+}
+
+/// A Java `String[]` of those, or a null array if the JVM refused one.
+///
+/// Null rather than a panic across the JNI boundary: `DevLogProvider`
+/// reads it as "the provider could not answer" and returns no cursor,
+/// which Dev Updater already draws as a distinct state. Taking the app
+/// down to report that its diagnostic is unavailable would be worse than
+/// the diagnostic being unavailable.
+fn string_array(env: &mut JNIEnv, fields: &[String]) -> jobjectArray {
+    let null = std::ptr::null_mut();
+    let Ok(class) = env.find_class("java/lang/String") else {
+        return null;
+    };
+    let Ok(array) = env.new_object_array(fields.len() as i32, class, JObject::null()) else {
+        return null;
+    };
+    for (index, field) in fields.iter().enumerate() {
+        let Ok(value) = env.new_string(field) else {
+            return null;
+        };
+        if env
+            .set_object_array_element(&array, index as i32, value)
+            .is_err()
+        {
+            return null;
+        }
+    }
+    array.into_raw()
+}

iris/android-app/src/enrollment.rs

@@ -0,0 +1,133 @@
+//! Which `ai-server` this app talks to, and how it was told.
+//!
+//! The parsing, the file and its owner-only mode are
+//! `client_core::config` (`EnrolledServer`/`EnrollmentStore`), shared with
+//! the desktop app. What is genuinely this platform's, and all that is
+//! here, is the intent plumbing: Android hands an `aiapp://enroll?...`
+//! link to `MainActivity`, which passes it and the app's private files
+//! directory across JNI (see `lib.rs`'s two exported functions).
+//!
+//! **Why the app is told at runtime rather than at build time.** The APK
+//! is cross-compiled in a VM and run against the server on the host, whose
+//! CA and token are not this machine's -- so nothing about the destination
+//! can be baked in, and no token or CA may sit in a repo or a delivered
+//! artifact either way. The CA arrives with the link (`ca` parameter,
+//! `wg_app_link::enroll::ca_param`), which is what makes an APK built
+//! anywhere able to pin the server it is pointed at.
+//!
+//! The files directory is process-wide state, which this project otherwise
+//! avoids: it arrives from the activity, and `AndroidAppState::new` -- the
+//! first thing that wants the enrollment -- has no parameter it could come
+//! in through. Same shape, and the same reason, as
+//! `client_core::log_ring`'s process ring.
+
+#[cfg(not(feature = "bench"))]
+use client_core::api::UreqTransport;
+use client_core::config::{EnrolledServer, EnrollmentStore};
+use std::path::PathBuf;
+use std::sync::OnceLock;
+
+/// `Context.getFilesDir()`, handed over by `MainActivity` before it builds
+/// the view. Set once per process; a second call with a different path is
+/// a programmer error rather than something to recover from, and a second
+/// call with the same one is what a re-created activity does.
+static FILES_DIR: OnceLock<PathBuf> = OnceLock::new();
+
+pub fn set_files_dir(dir: PathBuf) {
+    if let Err(existing) = FILES_DIR.set(dir.clone()) {
+        assert_eq!(
+            existing, dir,
+            "the app's files directory was set twice with different paths"
+        );
+    }
+}
+
+/// `None` before `MainActivity` has handed the directory over -- which is
+/// **not** the same as "not enrolled", and is why [`status`] has a state
+/// for it (UI_RULES: design the unknown state first).
+fn store() -> Option<EnrollmentStore> {
+    FILES_DIR.get().map(EnrollmentStore::new)
+}
+
+/// What this app has been told, or why it has not been.
+pub enum Status {
+    Enrolled(EnrolledServer),
+    /// Nothing has been enrolled yet: the ordinary first-run state.
+    NotEnrolled,
+    /// The question could not be answered -- the activity never handed a
+    /// files directory over, or the file is there and unreadable. Kept
+    /// apart from `NotEnrolled` because the two want different actions
+    /// from whoever is looking.
+    Unknown(String),
+}
+
+pub fn status() -> Status {
+    let Some(store) = store() else {
+        return Status::Unknown("the activity never handed over a files directory".to_string());
+    };
+    match store.load() {
+        Ok(Some(server)) => Status::Enrolled(server),
+        Ok(None) => Status::NotEnrolled,
+        Err(error) => Status::Unknown(error.to_string()),
+    }
+}
+
+/// One line for the diagnostics pane. The three states read differently on
+/// purpose: "not enrolled" says what to do about it, and "couldn't tell"
+/// must not be mistaken for it.
+///
+/// Only the bench build has a pane to put this in -- same gate, and the
+/// same reason, as `app_log::diagnostics_line`. The transcript build says
+/// the same things where they matter to it, in the message
+/// [`transport`]'s error becomes on screen.
+#[cfg(feature = "bench")]
+pub fn status_line() -> String {
+    match status() {
+        Status::Enrolled(server) => format!("enrolled: {}:{}", server.host, server.port),
+        Status::NotEnrolled => "not enrolled -- open the enrol link from Dev Updater".to_string(),
+        Status::Unknown(why) => format!("enrolment unreadable: {why}"),
+    }
+}
+
+/// Parses an `aiapp://enroll?...` link and saves it, replacing whatever
+/// was enrolled before -- opening a link is how somebody says "this server
+/// now", including after the old one's token was rotated.
+///
+/// The returned `Err` is the message for a person: this is called from a
+/// tap on a link, and a link that did nothing with nothing said is the
+/// failure the UI rules are most insistent about.
+pub fn apply_link(uri: &str) -> Result<EnrolledServer, String> {
+    let server = EnrolledServer::parse_link(uri)?;
+    let store = store().ok_or("the app has no files directory to save an enrollment in")?;
+    store
+        .save(&server)
+        .map_err(|error| format!("couldn't save the enrollment: {error}"))?;
+    Ok(server)
+}
+
+/// A transport for the enrolled server, pinning the CA the link carried.
+///
+/// Gated to the same builds as `transcript_client`, its only caller: the
+/// bench build opens a checked-in fixture and reaches no server, so
+/// compiling this into it would be a warning about dead code that is
+/// dead on purpose.
+///
+/// Every failure here is a sentence a screen can show, because there is
+/// nowhere else for it to go: this app has no `logcat` on the phone it is
+/// built for.
+#[cfg(not(feature = "bench"))]
+pub fn transport() -> Result<UreqTransport, String> {
+    let server = match status() {
+        Status::Enrolled(server) => server,
+        Status::NotEnrolled => {
+            return Err("Not enrolled yet -- open the enrol link from Dev Updater.".to_string());
+        }
+        Status::Unknown(why) => return Err(format!("Couldn't read the enrollment: {why}")),
+    };
+    let ca_pem = server.ca_pem.as_ref().ok_or(
+        "The enrollment link carried no CA, so there is nothing to pin. \
+         Enrol again with a link minted by this server.",
+    )?;
+    UreqTransport::new(server.base_url(), &server.token, ca_pem.as_bytes())
+        .map_err(|error| error.message)
+}

iris/android-app/src/lib.rs

@@ -40,6 +40,7 @@ use android_view::{
     Context, View,
     jni::{
         JNIEnv, JavaVM,
+        objects::{JClass, JString},
         sys::{JNI_VERSION_1_6, JavaVM as RawJavaVM, jint, jlong},
     },
     register_view_class,
@@ -51,10 +52,25 @@ use iris::prelude::*;
 use log::LevelFilter;
 use std::ffi::c_void;
 
+/// The app's own log ring and its upload -- only where `client-core` is
+/// linked, which is every build that has a server to send to. The plain
+/// tabs demo keeps `android_logger` alone, as it always had.
+#[cfg(feature = "transcript-screen")]
+mod app_log;
 #[cfg(feature = "bench")]
 mod bench_client;
 #[cfg(feature = "bench")]
 mod bench_jni;
+/// This app's log ring, handed to Dev Updater on the phone through a
+/// `ContentProvider`. Declared in every build for the reason the module
+/// gives: the Java class is in the manifest either way, and a `native`
+/// method the library does not export fails the class load.
+mod devlog;
+/// Which server this app talks to, told to it at runtime by an
+/// `aiapp://enroll` link. Only where `client-core` is linked -- the plain
+/// tabs demo makes no network call and has nothing to enrol against.
+#[cfg(feature = "transcript-screen")]
+mod enrollment;
 #[cfg(all(feature = "transcript-screen", not(feature = "bench")))]
 mod transcript_client;
 
@@ -119,6 +135,13 @@ extern "system" fn new_view_peer<'local>(
 /// mirrors android-view's own demo, which carries the same comment.
 #[unsafe(no_mangle)]
 pub unsafe extern "system" fn JNI_OnLoad(vm: *mut RawJavaVM, _: *mut c_void) -> jint {
+    // The ring in front of `android_logger` where there is one (see
+    // `app_log`), and `android_logger` alone otherwise. Both install the
+    // same tag and level, so `logcat` cannot tell the two builds apart --
+    // the ring only adds a second reader.
+    #[cfg(feature = "transcript-screen")]
+    app_log::install(LevelFilter::Debug);
+    #[cfg(not(feature = "transcript-screen"))]
     android_logger::init_once(
         android_logger::Config::default()
             .with_max_level(LevelFilter::Debug)
@@ -130,3 +153,83 @@ pub unsafe extern "system" fn JNI_OnLoad(vm: *mut RawJavaVM, _: *mut c_void) ->
     iris::android::register_native_methods(&mut env, VIEW_CLASS);
     JNI_VERSION_1_6
 }
+
+/// `MainActivity.nativeSetFilesDir` -- the app's private directory, handed
+/// over before the view exists because that is where the enrollment is
+/// read from and written to (`enrollment`'s module doc).
+///
+/// Exported by name rather than registered through `RegisterNatives`: the
+/// view's methods are registered because `android-view` owns that class
+/// and hands out one function pointer, whereas these two are this app's
+/// own activity and the mangled name is the whole of what is needed.
+///
+/// Declared in every build, including the tabs demo that has no
+/// `client-core` to store anything -- a `native` method Java declares and
+/// the library does not export is an `UnsatisfiedLinkError` when the class
+/// loads, which would take down a build that merely shares the activity.
+///
+/// # Safety
+/// Called by the JVM with the arguments its `native` declaration names.
+#[unsafe(no_mangle)]
+pub extern "system" fn Java_dev_iris_android_demo_MainActivity_nativeSetFilesDir(
+    mut env: JNIEnv,
+    _class: JClass,
+    dir: JString,
+) {
+    let Some(dir) = jstring(&mut env, dir) else {
+        return;
+    };
+    #[cfg(feature = "transcript-screen")]
+    {
+        app_log::set_crash_dir(std::path::Path::new(&dir));
+        enrollment::set_files_dir(std::path::PathBuf::from(&dir));
+    }
+    log::debug!("iris app: files directory is {dir}");
+}
+
+/// `MainActivity.nativeEnroll` -- one `aiapp://enroll?...` link, from the
+/// VIEW intent that started or resumed the activity.
+///
+/// Logged either way rather than answered: the activity has nothing to do
+/// with the result, and where the enrollment shows up is the diagnostics
+/// pane (`enrollment::status_line`), which reads the stored answer rather
+/// than being told it.
+///
+/// # Safety
+/// Called by the JVM with the arguments its `native` declaration names.
+#[unsafe(no_mangle)]
+pub extern "system" fn Java_dev_iris_android_demo_MainActivity_nativeEnroll(
+    mut env: JNIEnv,
+    _class: JClass,
+    uri: JString,
+) {
+    let Some(uri) = jstring(&mut env, uri) else {
+        return;
+    };
+    #[cfg(feature = "transcript-screen")]
+    match enrollment::apply_link(&uri) {
+        // Never the token: `wg-app-link`'s enroll module forbids logging
+        // it, and this line would otherwise be the one place it leaked.
+        Ok(server) => log::info!("iris app: enrolled with {}:{}", server.host, server.port),
+        Err(error) => log::warn!("iris app: that enrolment link was refused -- {error}"),
+    }
+    #[cfg(not(feature = "transcript-screen"))]
+    log::warn!("iris app: {uri} arrived, but this build has no server to enrol with");
+}
+
+/// A `JString` as a Rust `String`, or `None` for a null or non-UTF-8 one --
+/// neither is worth taking the app down for, and both are logged where
+/// they happen.
+fn jstring(env: &mut JNIEnv, value: JString) -> Option<String> {
+    if value.is_null() {
+        log::warn!("iris app: the activity passed a null string across JNI");
+        return None;
+    }
+    match env.get_string(&value) {
+        Ok(value) => Some(value.into()),
+        Err(error) => {
+            log::warn!("iris app: couldn't read a string from the activity -- {error}");
+            None
+        }
+    }
+}

iris/android-app/src/transcript_client.rs

@@ -7,15 +7,15 @@
 //!
 //! **Deliberate simplification, recorded rather than left to be
 //! rediscovered (RUST.md's I5 box has the full account)**: there is no
-//! session list and no enrollment UI here. The server, port, token and
-//! pinned CA are baked in at build time (`build.rs`'s
-//! `AI_APP_TRANSCRIPT_HOST`/`_PORT`/`_TOKEN`/`AI_APP_CA`), and the first
-//! session `ApiClient::fetch_sessions` returns is opened automatically --
-//! there is nothing to tap to get there, which is what `transcript-bench.sh`
-//! and `ui-trace` need to land straight on the screen under test. A real
-//! app needs `desktop-app`'s `EnrolledServer`/QR-link flow or E3's
-//! Keystore-sealed `ServerConfig.kt`; building a second one of those was
-//! not this pass's job.
+//! session list here -- the first session `ApiClient::fetch_sessions`
+//! returns is opened automatically, since there is nothing to tap to get
+//! there, which is what `transcript-bench.sh` and `ui-trace` need to land
+//! straight on the screen under test.
+//!
+//! Which server it opens it against is no longer baked in: it is the
+//! enrollment an `aiapp://enroll` link left behind (`crate::enrollment`,
+//! and `desktop-app`'s identical `--link`), because an APK
+//! cross-compiled here cannot pin the CA of a server on the host.
 //!
 //! **Reuses `iris/desktop-app`'s `app.rs` shape almost exactly** --
 //! `fold_event`/`group_tool_runs`/`fold_page`/`raw_seq` from
@@ -46,10 +46,6 @@ use iris::prelude::*;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU64, Ordering};
 
-mod pinned {
-    include!(concat!(env!("OUT_DIR"), "/pinned_config.rs"));
-}
-
 pub struct TranscriptClient {
     ui_state: AndroidUiState,
     /// The screen's own content -- everything under the fixed
@@ -85,18 +81,16 @@ impl HasAndroidUiState for TranscriptClient {
     }
 }
 
-/// Builds one `UreqTransport` from the config `build.rs` baked in. Called
-/// twice per session load, same as `desktop-app`'s `build_transport`
-/// closure -- `ApiClient` and the live-stream follow each need their own,
-/// since `UreqTransport` holds its own `ureq::Agent`.
+/// Builds one `UreqTransport` from the stored enrollment. Called twice per
+/// session load, same as `desktop-app`'s `build_transport` closure --
+/// `ApiClient` and the live-stream follow each need their own, since
+/// `UreqTransport` holds its own `ureq::Agent`.
+///
+/// Read afresh each time rather than held: opening a new enrolment link
+/// while the app is running is how somebody points it at another server,
+/// and a cached transport would keep talking to the old one.
 fn build_transport() -> Result<UreqTransport, String> {
-    let base_url = format!("https://{}:{}", pinned::HOST, pinned::PORT);
-    UreqTransport::new(
-        base_url,
-        pinned::TOKEN.to_string(),
-        pinned::CA_PEM.as_bytes(),
-    )
-    .map_err(|e| e.to_string())
+    crate::enrollment::transport()
 }
 
 fn placeholder<Rsc: HasEvents>(rsc: &mut Rsc, message: &str) -> StrongWidget {

iris/benches/fling_spline_reference.py

@@ -130,7 +130,11 @@ if __name__ == "__main__":
     # 2.55 is Iris's Pixel 9 Pro XL (docs/bench/iris-phone-v2-2026-09-06.md);
     # 2.75 is this checkout's emulator.
     for density in (2.55, 2.75):
-        for velocity in (5000.0, 11064.0):
+        # 15250 is `transcript-fixture/touch/flick-120hz.touch`'s own
+        # release velocity (velocity_reference.py), so `phone_screen.rs`
+        # can bound the fling it produces from *here* rather than from the
+        # `FlingCalculator` under test (docs/REVIEW-2026-09-07.md's T1).
+        for velocity in (5000.0, 11064.0, 15250.0):
             dur = fling_duration_s(velocity, density)
             print(
                 f"density={density} v={velocity}: "

iris/benches/report_to_touch.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+"""Turns `iris::input` debug lines -- from a phone's diagnostics report, or
+from a report the layer-1 harness produced with tracing on
+(`iris::diagnostics::set_trace(true)`) -- back into a `TouchScript` file
+`iris::harness::Harness::replay` can play back at layer 1.
+
+Why this exists: `docs/RUST.md`'s "Three test layers" box says the cheapest
+layer that can answer a question wins, and a gesture that misbehaves on
+Iris's phone is otherwise only describable in words. `iris::sense::
+log_input_event`'s one line per platform event (Android's on_touch_event
+once per `MotionEvent`, with historical samples inline; winit's once per
+pointer `WindowEvent`; the harness's `touch`, once per script line) already
+carries everything a `.touch` file's `t_ms action x y` needs -- this just
+reads it back out and reconstructs the samples in order, expanding each
+event's inline historical samples into their own `move` lines first (they
+are always intermediate positions of a move, and Android documents them as
+oldest first, which is also the order they appear in the line).
+
+Usage:
+    report_to_touch.py < report.txt > replay.touch
+    report_to_touch.py report.txt > replay.touch
+
+Only lines containing "iris input: action=..." are read; everything else in
+the report (insets, frame timings, drag-release summaries) is ignored, so
+this can be pointed at Copy report's whole clipboard text directly.
+"""
+
+import re
+import sys
+
+# The message half of `sense::log_input_event`'s format string, prefix-
+# agnostic: a real report line also carries the ring's own
+# `HH:MM:SS.mmm LEVEL target:` header (`LogLine::format`) or, forwarded
+# through `ai_server::client_log`, a `[<source> <clock> #<seq>]` tag ahead
+# of that -- neither of which this needs to understand, since `search`
+# (not `match`) finds the marker wherever it starts.
+LINE_RE = re.compile(
+    r"iris input: action=(?P<action>\w+) x=(?P<x>-?[0-9.]+) y=(?P<y>-?[0-9.]+) "
+    r"t=(?P<t>[0-9]+)ms history=(?P<hist>[0-9]+)(?P<rest>.*)$"
+)
+# One historical sample inside `rest`: `t:x,y`, space-separated, oldest first
+# -- see `log_input_event`'s own doc for why order matters.
+HIST_RE = re.compile(r"(?P<t>[0-9]+):(?P<x>-?[0-9.]+),(?P<y>-?[0-9.]+)")
+
+
+def _fmt(value: float) -> str:
+    """The number as `TouchScript::parse`'s own `f32::parse` would round-trip
+    it -- an integer without a trailing `.0` where the source was one
+    (every coordinate here is a physical pixel), `{:g}` otherwise so a
+    fractional value from a real device is not silently truncated."""
+    if value == int(value):
+        return str(int(value))
+    return f"{value:g}"
+
+
+def convert(lines):
+    """Every `iris::input` line, oldest first, expanded to one `(t_ms,
+    action, x, y)` tuple per touch sample -- a historical sample is always
+    an intermediate `move`, and the event's own sample keeps its real
+    action (`down`/`move`/`up`/`cancel`)."""
+    rows = []
+    for line in lines:
+        m = LINE_RE.search(line)
+        if not m:
+            continue
+        hist_count = int(m.group("hist"))
+        hist_matches = list(HIST_RE.finditer(m.group("rest")))
+        if len(hist_matches) != hist_count:
+            print(
+                f"report_to_touch: {line.strip()!r} says history={hist_count} but "
+                f"holds {len(hist_matches)} samples -- skipped",
+                file=sys.stderr,
+            )
+            continue
+        for hm in hist_matches:
+            rows.append(
+                (int(hm.group("t")), "move", float(hm.group("x")), float(hm.group("y")))
+            )
+        rows.append(
+            (int(m.group("t")), m.group("action"), float(m.group("x")), float(m.group("y")))
+        )
+    return rows
+
+
+def main():
+    if len(sys.argv) > 2:
+        print("usage: report_to_touch.py [report.txt] < report.txt", file=sys.stderr)
+        return 2
+    text = open(sys.argv[1]) if len(sys.argv) == 2 else sys.stdin
+    for t_ms, action, x, y in convert(text):
+        print(f"{t_ms} {action} {_fmt(x)} {_fmt(y)}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

iris/benches/velocity_reference.py

@@ -0,0 +1,298 @@
+#!/usr/bin/env python3
+"""Compose's touch velocity tracker, transcribed independently of the Rust port.
+
+Same reason `fling_spline_reference.py` exists: the numbers checked into
+`sense.rs`'s velocity tests must not be numbers the Rust produced. The old
+estimator -- total motion over the sample span, an average -- passed every test
+it had, because every one of those tests asserted the average's own definition
+back at it. An average cannot tell an accelerating flick from a steady drag, and
+that is exactly what Iris reported from the phone on 2026-09-07: "flinging now
+actually works but is slower than Compose's immediately after releasing the
+flick".
+
+Transcribed by hand from, and only from, the `-sources.jar` of
+**androidx.compose.ui:ui-android:1.12.0** and
+**androidx.compose.foundation:foundation-android:1.12.0**
+(dl.google.com/dl/android/maven2), read 2026-09-07:
+
+  * `androidx/compose/ui/input/pointer/util/VelocityTracker.kt` --
+    `VelocityTracker1D.calculateVelocity`, `polyFitLeastSquares`,
+    `calculateImpulseVelocity`, `kineticEnergyToVelocity`, and the constants
+    `HistorySize = 20`, `HorizonMilliseconds = 100`,
+    `AssumePointerMoveStoppedMilliseconds = 40`.
+  * `androidx/compose/ui/input/pointer/util/PlatformVelocityTracker.kt` --
+    `Lsq2VelocityTracker`, which is what the 2D `VelocityTracker` delegates to.
+  * `androidx/compose/ui/input/pointer/util/PlatformVelocityTracker.android.kt`
+    -- the `AndroidComposeUiFlags.isFrameworkVelocityTrackerEnabled` fork.
+  * `androidx/compose/ui/AndroidComposeUiFlags.android.kt` -- that flag's
+    default, which is `false`.
+  * `androidx/compose/foundation/gestures/Draggable.kt` -- `sendDragStart` /
+    `sendDragEvent` / `sendDragStopped`, i.e. *which* samples a touch drag
+    feeds the tracker and where the maximum-velocity clamp is applied.
+  * `androidx/compose/foundation/gestures/DifferentialVelocityTracker.kt` and
+    `NonTouchScrollingLogic.kt` -- the Impulse strategy's only caller.
+  * `androidx/compose/foundation/gestures/Scrollable.kt` --
+    `DefaultFlingBehavior.performFling`, for the minimum-velocity question.
+
+**Which strategy a touch fling actually uses, since this was the surprise.**
+`Strategy.Impulse` is *not* it. `scrollable`/`draggable` release through
+`DragGestureNode.sendDragStopped`, which calls the 2D `VelocityTracker`; on
+Android that is `Lsq2VelocityTracker` (the framework-tracker flag defaults to
+false), which is two `VelocityTracker1D(strategy = Lsq2)` -- a degree-2
+least-squares fit over **absolute positions**, whose velocity is the fitted
+polynomial's derivative at the newest sample. Impulse is reached only through
+`DifferentialVelocityTracker`, whose sole caller is `NonTouchScrollingLogic`:
+mouse wheel and trackpad, never a finger. So this script transcribes Lsq2 and
+iris ports Lsq2. `calculate_impulse_velocity` is here anyway, unused by the
+printed points, because ruling it out by reading is cheaper than ruling it out
+again next time somebody remembers "Compose uses impulse".
+
+**Which samples a touch drag feeds it.** `sendDragStart` adds the DOWN change;
+every subsequent MOVE, historical samples included, is added by `sendDragEvent`.
+The **UP position is never added**: `Lsq2VelocityTracker.addPointerInputChange`
+wraps its two `addPosition` calls in `if (!event.changedToUpIgnoreConsumed())`,
+and all the UP branch does is reset the tracker when more than 40ms have passed
+since the last MOVE (b/238654963). So a finger that stops before lifting reads
+as a stop, not as a decelerating tail. Positions are the raw event positions,
+so the touch slop is inside the motion the tracker sees even though the list
+never scrolled by it.
+
+Two of Compose's samples iris does *not* reproduce, both noted rather than
+copied: pre-slop MOVEs (iris's `DragArbiter` is `Undecided` then too, so it
+feeds none either -- these agree), and the single MOVE that *crosses* the slop,
+which Compose drops because `sendDragStart` adds only the DOWN. iris feeds that
+one, since it is a real measured position and dropping it would be copying a
+quirk of where Compose happens to split its state machine.
+
+**The clamps.** Maximum: `sendDragStopped` passes
+`LocalViewConfiguration.maximumFlingVelocity`, which on Android is
+`ViewConfiguration.getScaledMaximumFlingVelocity()` -- 8000 dp/s. Minimum:
+there is **none** on this path. `ViewConfiguration.minimumFlingVelocity`
+exists in Compose's `ViewConfiguration` interface but its only use in either
+artifact is `NestedScrollInteropConnection`, for View interop.
+`DefaultFlingBehavior.performFling` guards with `abs(initialVelocity) > 1f`
+and says why in its own comment: "we need it since spline curve gives us
+NaNs". 1 px/s, not 50 dp/s.
+
+Run it with no arguments; it prints the sample sets and the velocities the
+Rust tests assert on.
+"""
+
+import math
+
+HISTORY_SIZE = 20
+HORIZON_MILLISECONDS = 100.0
+ASSUME_POINTER_MOVE_STOPPED_MILLISECONDS = 40.0
+MIN_SAMPLE_SIZE_LSQ2 = 3
+
+# ViewConfiguration.getScaledMaximumFlingVelocity(), in dp/s.
+MAXIMUM_FLING_VELOCITY_DP_S = 8000.0
+# DefaultFlingBehavior.performFling's own threshold, in the units of the
+# positions fed to the tracker -- pixels per second here.
+FLING_MINIMUM_PX_S = 1.0
+
+
+def poly_fit_least_squares(x, y, sample_count, degree):
+    """`polyFitLeastSquares`: Gram-Schmidt QR, coefficients low order first."""
+    if degree < 1:
+        raise ValueError("The degree must be at positive integer")
+    if sample_count == 0:
+        raise ValueError("At least one point must be provided")
+
+    truncated_degree = sample_count - 1 if degree >= sample_count else degree
+    m = sample_count
+    n = truncated_degree + 1
+
+    # a[i][h] = x[h]**i, pre-multiplied by the (always 1.0) weight.
+    a = [[0.0] * m for _ in range(n)]
+    for h in range(m):
+        a[0][h] = 1.0
+        for i in range(1, n):
+            a[i][h] = a[i - 1][h] * x[h]
+
+    q = [[0.0] * m for _ in range(n)]
+    r = [[0.0] * n for _ in range(n)]
+    for j in range(n):
+        w = q[j]
+        w[:] = a[j][:m]
+        for i in range(j):
+            z = q[i]
+            dot = sum(w[h] * z[h] for h in range(m))
+            for h in range(m):
+                w[h] -= dot * z[h]
+        norm = math.sqrt(sum(v * v for v in w))
+        inverse_norm = 1.0 / max(norm, 1e-6)
+        for h in range(m):
+            w[h] *= inverse_norm
+        for i in range(n):
+            r[j][i] = 0.0 if i < j else sum(w[h] * a[i][h] for h in range(m))
+
+    coefficients = [0.0] * n
+    for i in range(n - 1, -1, -1):
+        c = sum(q[i][h] * y[h] for h in range(m))
+        for j in range(n - 1, i, -1):
+            c -= r[i][j] * coefficients[j]
+        coefficients[i] = c / r[i][i]
+    return coefficients
+
+
+def kinetic_energy_to_velocity(kinetic_energy):
+    sign = 0.0 if kinetic_energy == 0.0 else math.copysign(1.0, kinetic_energy)
+    return sign * math.sqrt(2 * abs(kinetic_energy))
+
+
+def calculate_impulse_velocity(data_points, time, sample_count, is_data_differential):
+    """`calculateImpulseVelocity` -- not on the touch path; see the module doc."""
+    work = 0.0
+    start = sample_count - 1
+    next_time = time[start]
+    for i in range(start, 0, -1):
+        current_time = next_time
+        next_time = time[i - 1]
+        if current_time == next_time:
+            continue
+        if is_data_differential:
+            delta = -data_points[i - 1]
+        else:
+            delta = data_points[i] - data_points[i - 1]
+        v_curr = delta / (current_time - next_time)
+        v_prev = kinetic_energy_to_velocity(work)
+        work += (v_curr - v_prev) * abs(v_curr)
+        if i == start:
+            work = work * 0.5
+    return kinetic_energy_to_velocity(work)
+
+
+def calculate_velocity(samples):
+    """`VelocityTracker1D.calculateVelocity` with `Strategy.Lsq2`.
+
+    `samples` is `(time_millis, position)` oldest first, at most the last
+    `HISTORY_SIZE` of which the circular buffer would still be holding.
+    Returns units per second.
+    """
+    held = samples[-HISTORY_SIZE:]
+    if not held:
+        return 0.0
+
+    data_points = []
+    time = []
+    newest_time, _ = held[-1]
+    previous_time = newest_time
+    for sample_time, sample_position in reversed(held):
+        age = float(newest_time - sample_time)
+        delta = abs(float(sample_time - previous_time))
+        # Lsq2 walks back sample to sample; only the non-differential
+        # Impulse branch compares every sample against the newest one.
+        previous_time = sample_time
+        if age > HORIZON_MILLISECONDS or delta > ASSUME_POINTER_MOVE_STOPPED_MILLISECONDS:
+            break
+        data_points.append(sample_position)
+        time.append(-age)
+        if len(data_points) == HISTORY_SIZE:
+            break
+
+    if len(data_points) < MIN_SAMPLE_SIZE_LSQ2:
+        return 0.0
+    try:
+        coefficients = poly_fit_least_squares(time, data_points, len(data_points), 2)
+    except ValueError:
+        return 0.0
+    # The 2nd coefficient is the fitted polynomial's derivative at x = 0,
+    # which is the newest sample's timestamp. units/ms -> units/s.
+    return coefficients[1] * 1000.0
+
+
+def clamped(velocity, maximum):
+    """`VelocityTracker1D.calculateVelocity(maximumVelocity)`."""
+    if velocity == 0.0 or math.isnan(velocity):
+        return 0.0
+    return min(velocity, maximum) if velocity > 0 else max(velocity, -maximum)
+
+
+def average(samples):
+    """The estimator being replaced: total motion over the span."""
+    if len(samples) < 2:
+        return 0.0
+    span = (samples[-1][0] - samples[0][0]) / 1000.0
+    if span <= 0.0:
+        return 0.0
+    return (samples[-1][1] - samples[0][1]) / span
+
+
+# --- The three recorded sample sets the Rust tests assert on. ----------------
+
+# 1. `transcript-fixture/touch/flick-120hz.touch`, as `DragGesture` feeds it:
+#    the DOWN position, then one position per MOVE. The UP at t=20 adds no
+#    sample (see the module doc), which is why the finger sitting still for its
+#    last 4ms does not drag the estimate down. y only; the flick is vertical.
+FLICK_120HZ = [(0, 1000.0), (4, 1040.0), (8, 1086.0), (12, 1138.0), (16, 1196.0)]
+
+# 2. A steady drag: 5px every 10ms for 100ms. A constant-velocity fit and an
+#    average must agree here -- this is the case that cannot tell the two
+#    estimators apart, which is why it is not the only one.
+STEADY_DRAG = [(i * 10, float(i * 5)) for i in range(11)]
+
+# 3. A flick that accelerates into the release: 10ms apart, deltas doubling.
+#    This is the case the average gets wrong, and the negative control for
+#    the port -- reverting to the average must fail this test and only this
+#    kind of test.
+ACCELERATING_FLICK = [(0, 0.0), (10, 2.0), (20, 6.0), (30, 14.0), (40, 30.0), (50, 54.0)]
+
+# 4. The two edges of the sample walk, checked here so the Rust asserts
+#    Compose's answer rather than iris's own reading of the rule.
+#    (a) An old, fast burst outside the 100ms horizon, then a slow steady
+#        drag: the burst must not leak into the estimate.
+OLD_BURST_THEN_STEADY = [(0, 0.0)] + [(10 + i * 10, 1000.0 + i) for i in range(11)]
+#    (b) The finger stops for 48ms and then lifts. The gap exceeds
+#        AssumePointerMoveStopped, so the walk breaks after one sample and
+#        there is no fling -- what stops a "park it and let go" from
+#        flinging at whatever speed the finger arrived with.
+STOPPED_BEFORE_RELEASE = [(0, 0.0), (4, 40.0), (8, 90.0), (12, 150.0), (60, 152.0)]
+
+# 5. `sense.rs`'s own `drag_gesture_tests`: what `DragGesture` feeds for a
+#    press and two move frames, which is the fewest a fit can use.
+TWO_MOVE_FRAMES = [(0, 0.0), (8, 100.0), (16, 220.0)]
+#    ... and one move frame, which Compose cannot fit either.
+ONE_MOVE_FRAME = [(0, 0.0), (8, 100.0)]
+
+# The phone: 1080x2424 at content_scale 2.55.
+PHONE_DENSITY = 2.55
+
+
+def report(name, samples):
+    v = calculate_velocity(samples)
+    print(f"{name}:")
+    print(f"  samples (t_ms, position): {samples}")
+    print(f"  Lsq2 (Compose's touch path):    {v:.4f} px/s")
+    print(f"  average (the old estimator):    {average(samples):.4f} px/s")
+    print(f"  impulse (non-touch, for ref):   ", end="")
+    held = list(reversed(samples[-HISTORY_SIZE:]))
+    newest = held[0][0]
+    print(
+        f"{calculate_impulse_velocity([p for _, p in held], [-(newest - t) for t, _ in held], len(held), False) * 1000.0:.4f} px/s"
+    )
+    print()
+
+
+if __name__ == "__main__":
+    print("Compose 1.12.0 touch velocity: VelocityTracker1D, Strategy.Lsq2,")
+    print("non-differential (positions), HistorySize=20, Horizon=100ms,")
+    print("AssumePointerMoveStopped=40ms, minSampleSize=3.\n")
+    report("flick-120hz.touch", FLICK_120HZ)
+    report("steady drag (5px/10ms)", STEADY_DRAG)
+    report("accelerating flick (deltas 2,4,8,16,24 per 10ms)", ACCELERATING_FLICK)
+
+    report("old burst then steady 1px/10ms", OLD_BURST_THEN_STEADY)
+    report("stopped 48ms before release", STOPPED_BEFORE_RELEASE)
+    report("press and two move frames", TWO_MOVE_FRAMES)
+    report("press and one move frame", ONE_MOVE_FRAME)
+
+    print("Clamps:")
+    print(f"  maximum: {MAXIMUM_FLING_VELOCITY_DP_S} dp/s")
+    print(
+        f"           = {MAXIMUM_FLING_VELOCITY_DP_S * PHONE_DENSITY:.1f} px/s at the phone's density {PHONE_DENSITY}"
+    )
+    print(f"  minimum: none on the fling path; DefaultFlingBehavior skips |v| <= {FLING_MINIMUM_PX_S} px/s")
+    print()
+    print("Two samples only (a press and one move, the phone's 120Hz worst case):")
+    print(f"  Lsq2 needs 3 and answers {calculate_velocity(FLICK_120HZ[:2]):.4f} px/s")

iris/core/assets/fonts/NERD_FONTS_LICENSE.txt

@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2014 Ryan L McIntyre
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

iris/core/assets/fonts/OFL.txt

@@ -1,201 +0,0 @@
-                                 Apache License
-                           Version 2.0, January 2004
-                        http://www.apache.org/licenses/
-
-   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
-
-   1. Definitions.
-
-      "License" shall mean the terms and conditions for use, reproduction,
-      and distribution as defined by Sections 1 through 9 of this document.
-
-      "Licensor" shall mean the copyright owner or entity authorized by
-      the copyright owner that is granting the License.
-
-      "Legal Entity" shall mean the union of the acting entity and all
-      other entities that control, are controlled by, or are under common
-      control with that entity. For the purposes of this definition,
-      "control" means (i) the power, direct or indirect, to cause the
-      direction or management of such entity, whether by contract or
-      otherwise, or (ii) ownership of fifty percent (50%) or more of the
-      outstanding shares, or (iii) beneficial ownership of such entity.
-
-      "You" (or "Your") shall mean an individual or Legal Entity
-      exercising permissions granted by this License.
-
-      "Source" form shall mean the preferred form for making modifications,
-      including but not limited to software source code, documentation
-      source, and configuration files.
-
-      "Object" form shall mean any form resulting from mechanical
-      transformation or translation of a Source form, including but
-      not limited to compiled object code, generated documentation,
-      and conversions to other media types.
-
-      "Work" shall mean the work of authorship, whether in Source or
-      Object form, made available under the License, as indicated by a
-      copyright notice that is included in or attached to the work
-      (an example is provided in the Appendix below).
-
-      "Derivative Works" shall mean any work, whether in Source or Object
-      form, that is based on (or derived from) the Work and for which the
-      editorial revisions, annotations, elaborations, or other modifications
-      represent, as a whole, an original work of authorship. For the purposes
-      of this License, Derivative Works shall not include works that remain
-      separable from, or merely link (or bind by name) to the interfaces of,
-      the Work and Derivative Works thereof.
-
-      "Contribution" shall mean any work of authorship, including
-      the original version of the Work and any modifications or additions
-      to that Work or Derivative Works thereof, that is intentionally
-      submitted to Licensor for inclusion in the Work by the copyright owner
-      or by an individual or Legal Entity authorized to submit on behalf of
-      the copyright owner. For the purposes of this definition, "submitted"
-      means any form of electronic, verbal, or written communication sent
-      to the Licensor or its representatives, including but not limited to
-      communication on electronic mailing lists, source code control systems,
-      and issue tracking systems that are managed by, or on behalf of, the
-      Licensor for the purpose of discussing and improving the Work, but
-      excluding communication that is conspicuously marked or otherwise
-      designated in writing by the copyright owner as "Not a Contribution."
-
-      "Contributor" shall mean Licensor and any individual or Legal Entity
-      on behalf of whom a Contribution has been received by Licensor and
-      subsequently incorporated within the Work.
-
-   2. Grant of Copyright License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      copyright license to reproduce, prepare Derivative Works of,
-      publicly display, publicly perform, sublicense, and distribute the
-      Work and such Derivative Works in Source or Object form.
-
-   3. Grant of Patent License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      (except as stated in this section) patent license to make, have made,
-      use, offer to sell, sell, import, and otherwise transfer the Work,
-      where such license applies only to those patent claims licensable
-      by such Contributor that are necessarily infringed by their
-      Contribution(s) alone or by combination of their Contribution(s)
-      with the Work to which such Contribution(s) was submitted. If You
-      institute patent litigation against any entity (including a
-      cross-claim or counterclaim in a lawsuit) alleging that the Work
-      or a Contribution incorporated within the Work constitutes direct
-      or contributory patent infringement, then any patent licenses
-      granted to You under this License for that Work shall terminate
-      as of the date such litigation is filed.
-
-   4. Redistribution. You may reproduce and distribute copies of the
-      Work or Derivative Works thereof in any medium, with or without
-      modifications, and in Source or Object form, provided that You
-      meet the following conditions:
-
-      (a) You must give any other recipients of the Work or
-          Derivative Works a copy of this License; and
-
-      (b) You must cause any modified files to carry prominent notices
-          stating that You changed the files; and
-
-      (c) You must retain, in the Source form of any Derivative Works
-          that You distribute, all copyright, patent, trademark, and
-          attribution notices from the Source form of the Work,
-          excluding those notices that do not pertain to any part of
-          the Derivative Works; and
-
-      (d) If the Work includes a "NOTICE" text file as part of its
-          distribution, then any Derivative Works that You distribute must
-          include a readable copy of the attribution notices contained
-          within such NOTICE file, excluding those notices that do not
-          pertain to any part of the Derivative Works, in at least one
-          of the following places: within a NOTICE text file distributed
-          as part of the Derivative Works; within the Source form or
-          documentation, if provided along with the Derivative Works; or,
-          within a display generated by the Derivative Works, if and
-          wherever such third-party notices normally appear. The contents
-          of the NOTICE file are for informational purposes only and
-          do not modify the License. You may add Your own attribution
-          notices within Derivative Works that You distribute, alongside
-          or as an addendum to the NOTICE text from the Work, provided
-          that such additional attribution notices cannot be construed
-          as modifying the License.
-
-      You may add Your own copyright statement to Your modifications and
-      may provide additional or different license terms and conditions
-      for use, reproduction, or distribution of Your modifications, or
-      for any such Derivative Works as a whole, provided Your use,
-      reproduction, and distribution of the Work otherwise complies with
-      the conditions stated in this License.
-
-   5. Submission of Contributions. Unless You explicitly state otherwise,
-      any Contribution intentionally submitted for inclusion in the Work
-      by You to the Licensor shall be under the terms and conditions of
-      this License, without any additional terms or conditions.
-      Notwithstanding the above, nothing herein shall supersede or modify
-      the terms of any separate license agreement you may have executed
-      with Licensor regarding such Contributions.
-
-   6. Trademarks. This License does not grant permission to use the trade
-      names, trademarks, service marks, or product names of the Licensor,
-      except as required for reasonable and customary use in describing the
-      origin of the Work and reproducing the content of the NOTICE file.
-
-   7. Disclaimer of Warranty. Unless required by applicable law or
-      agreed to in writing, Licensor provides the Work (and each
-      Contributor provides its Contributions) on an "AS IS" BASIS,
-      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
-      implied, including, without limitation, any warranties or conditions
-      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
-      PARTICULAR PURPOSE. You are solely responsible for determining the
-      appropriateness of using or redistributing the Work and assume any
-      risks associated with Your exercise of permissions under this License.
-
-   8. Limitation of Liability. In no event and under no legal theory,
-      whether in tort (including negligence), contract, or otherwise,
-      unless required by applicable law (such as deliberate and grossly
-      negligent acts) or agreed to in writing, shall any Contributor be
-      liable to You for damages, including any direct, indirect, special,
-      incidental, or consequential damages of any character arising as a
-      result of this License or out of the use or inability to use the
-      Work (including but not limited to damages for loss of goodwill,
-      work stoppage, computer failure or malfunction, or any and all
-      other commercial damages or losses), even if such Contributor
-      has been advised of the possibility of such damages.
-
-   9. Accepting Warranty or Additional Liability. While redistributing
-      the Work or Derivative Works thereof, You may choose to offer,
-      and charge a fee for, acceptance of support, warranty, indemnity,
-      or other liability obligations and/or rights consistent with this
-      License. However, in accepting such obligations, You may act only
-      on Your own behalf and on Your sole responsibility, not on behalf
-      of any other Contributor, and only if You agree to indemnify,
-      defend, and hold each Contributor harmless for any liability
-      incurred by, or claims asserted against, such Contributor by reason
-      of your accepting any such warranty or additional liability.
-
-   END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "[]"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright [yyyy] [name of copyright owner]
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-       http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License.

iris/core/build-icon-font.sh

@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# Rebuilds iris/core/assets/fonts/nerd_icons.ttf.
+#
+# iris draws its icons as glyphs in a Nerd Fonts subset it ships, rather
+# than as ordinary Unicode out of whatever the platform resolved. Unicode's
+# own geometric shapes are what this replaced: `tool.rs` set its disclosure
+# mark with U+25B8/25BE/25B4, and once iris stopped bundling fonts
+# (2026-09-07) Iris's phone drew an empty box for them and this VM drew a
+# dot. UI_RULES: "don't rely on characters the platform might not have --
+# ship the glyph or the asset rather than hoping."
+#
+# The whole symbols font is 3 MB for the handful below, so what is
+# committed is a subset. Add a codepoint to GLYPHS *and* to
+# `iris/core/src/icon.rs` (the two lists have to agree -- a codepoint in
+# the Rust that this script did not subset is a glyph that silently isn't
+# there), then run this and commit the result.
+#
+# Needs python3 and network access; fontTools is fetched into a temporary
+# venv, so nothing has to be installed on the machine.
+#
+# The same arrangement as the Compose app's `app/build-icon-font.sh`, which
+# this is copied from -- including the Mono face and the Material Design
+# family, so an icon means the same thing in both apps. Copied rather than
+# shared because most of it is the GLYPHS list, which has to differ: the
+# point of subsetting is to ship only the codepoints one app draws.
+set -euo pipefail
+
+# Codepoint, then the Nerd Fonts glyph name it came from. Material Design
+# Icons, as in the Compose app.
+GLYPHS=(
+    U+F035D  # md-menu_down -- a card that is open
+    U+F035F  # md-menu_right -- a card that opens
+    U+F0360  # md-menu_up -- collapse this group again
+)
+
+url=https://github.com/ryanoasis/nerd-fonts/releases/latest/download/NerdFontsSymbolsOnly.zip
+here="$(cd "$(dirname "$0")" && pwd)"
+out="$here/assets/fonts/nerd_icons.ttf"
+work="$(mktemp -d)"
+trap 'rm -rf "$work"' EXIT
+
+echo "Fetching $url"
+curl -fsSL -o "$work/nf.zip" "$url"
+python3 -c 'import sys,zipfile; zipfile.ZipFile(sys.argv[1]).extractall(sys.argv[2])' "$work/nf.zip" "$work"
+
+python3 -m venv "$work/venv"
+"$work/venv/bin/pip" -q install fonttools
+
+unicodes="$(IFS=,; echo "${GLYPHS[*]}")"
+mkdir -p "$(dirname "$out")"
+# The Mono face, where every glyph is one em wide and one em tall, so two
+# icons at the same font size are the same size without either being given
+# one -- the same reason the Compose app's script takes it. It is also what
+# makes an icon's box predictable beside a line of text.
+"$work/venv/bin/pyftsubset" "$work/SymbolsNerdFontMono-Regular.ttf" \
+    --unicodes="$unicodes" \
+    --layout-features= \
+    --drop-tables+=DSIG \
+    --output-file="$out"
+cp "$work/LICENSE" "$here/assets/fonts/NERD_FONTS_LICENSE.txt"
+
+echo "Wrote $out ($(stat -c %s "$out") bytes) with ${#GLYPHS[@]} glyphs"

iris/core/src/event/manager.rs

@@ -79,6 +79,8 @@ type EventData<Rsc, E> = (E, Rc<dyn for<'a> EventFn<Rsc, <E as Event>::Data<'a>>
 pub struct TypeEventManager<Rsc: HasEvents, E: Event> {
     // TODO: reduce visiblity!!
     pub active: HashMap<LayerId, HashMap<WidgetId, E::State>>,
+    /// This event's own input-wide state -- see [`Event::Global`].
+    pub global: E::Global,
     map: HashMap<WidgetId, Vec<EventData<Rsc, E>>>,
 }
 
@@ -107,6 +109,7 @@ impl<Rsc: HasEvents, E: Event> Default for TypeEventManager<Rsc, E> {
     fn default() -> Self {
         Self {
             active: Default::default(),
+            global: Default::default(),
             map: Default::default(),
         }
     }

iris/core/src/event/mod.rs

@@ -9,6 +9,20 @@ pub use rsc::*;
 pub trait Event: Sized + 'static + Clone {
     type Data<'a>: Clone = ();
     type State: Default = ();
+    /// State this event owns that belongs to no single widget -- what the
+    /// thing dispatching the event knows about the *input*, rather than
+    /// about a listener. `()` for almost every event; the cursor's is
+    /// `iris::sense::PointerInput` (which widget holds pointer capture,
+    /// and who is tracking the press in flight).
+    ///
+    /// It lives here so that such state has one owner, reached by `&mut`
+    /// through the event manager, instead of being parked on whatever
+    /// structure a handler happens to be able to reach and guarded with a
+    /// lock. Iris asked for that on 2026-09-08, of the pointer capture
+    /// that used to sit in a `Mutex` on `UiRenderState`: "everything
+    /// global should be stored in the general input handler, not in
+    /// specific senses with locking stuff."
+    type Global: Default = ();
     #[allow(unused_variables)]
     fn should_run<'a>(&self, data: &Self::Data<'a>) -> Option<Self::Data<'a>> {
         Some(data.clone())

iris/core/src/icon.rs

@@ -0,0 +1,39 @@
+//! The icons iris draws, as codepoints in the Nerd Fonts subset it ships.
+//!
+//! **Why a bundled font rather than ordinary Unicode**: the disclosure
+//! mark used to be U+25B8/25BE/25B4 out of whatever face the platform
+//! resolved, and once iris stopped bundling fonts (DECISIONS.md,
+//! 2026-09-07) Iris's phone drew an empty box for them and this VM drew a
+//! dot. UI_RULES' answer is not to avoid glyphs but to ship them, which is
+//! also what the Compose app has always done for its icons
+//! (`app/build-icon-font.sh`, `NerdIcons.kt`) -- the same Material Design
+//! family, so an icon means the same thing in both apps.
+//!
+//! **Why not vector assets or drawn shapes**: an icon beside a line of
+//! text wants that line's size, colour and baseline, and text gets all
+//! three for free. This replaced `iris::widget::mark`, which drew the
+//! triangle into a texture: correct, but one shape, and every further icon
+//! would have been another bespoke rasteriser.
+//!
+//! Each constant here has to have a matching codepoint in
+//! `iris/core/build-icon-font.sh`'s `GLYPHS`; a codepoint here that the
+//! script did not subset is a glyph that silently isn't there. The subset
+//! is the font's **Mono** face, where every glyph is one em wide and one
+//! em tall, so two icons at one font size are one size without either
+//! being given one -- and why an icon looks smaller than text at the same
+//! size, since the glyph is drawn inside that em rather than filling it.
+//!
+//! Draw one with [`crate::Family::Icons`]:
+//!
+//! ```ignore
+//! text(icon::OPEN, 12.0, MUTED).family(Family::Icons)
+//! ```
+
+/// `md-menu_down` -- a filled triangle pointing down: this card is open.
+pub const OPEN: &str = "\u{F035D}";
+
+/// `md-menu_right` -- pointing right: this card opens.
+pub const CLOSED: &str = "\u{F035F}";
+
+/// `md-menu_up` -- pointing up: fold this group of cards away again.
+pub const COLLAPSE: &str = "\u{F0360}";

iris/core/src/lib.rs

@@ -19,6 +19,7 @@ mod render;
 mod ui;
 mod widget;
 
+pub mod icon;
 pub mod util;
 
 pub use attr::*;

iris/core/src/orientation/axis.rs

@@ -1,6 +1,6 @@
 use super::*;
 
-#[derive(Copy, Clone, Eq, PartialEq)]
+#[derive(Copy, Clone, Eq, PartialEq, Debug)]
 pub enum Axis {
     X,
     Y,

iris/core/src/primitive/layer.rs

@@ -1,10 +1,6 @@
 use std::ops::{Index, IndexMut};
 
-use crate::{
-    UiRegion, WidgetId,
-    render::{MaskIdx, MoveIdx, Primitive, PrimitiveHandle, PrimitiveInst, Primitives},
-    util::to_mut,
-};
+use crate::{render::LayerOrder, util::to_mut};
 
 pub type LayerId = usize;
 
@@ -40,7 +36,10 @@ struct Child {
     tail: usize,
 }
 
-pub type PrimitiveLayers = Layers<Primitives>;
+/// The draw order of every layer. The primitives themselves live in one
+/// arena beside this (`UiRenderState::primitives`); a layer names the
+/// slots it draws, which is what its vertex buffer is.
+pub type PrimitiveLayers = Layers<LayerOrder>;
 
 impl<T: Default> Layers<T> {
     pub fn new() -> Layers<T> {
@@ -120,32 +119,6 @@ impl<T: Default> Layers<T> {
     }
 }
 
-impl PrimitiveLayers {
-    pub fn write<P: Primitive>(
-        &mut self,
-        layer: LayerId,
-        info: PrimitiveInst<P>,
-    ) -> PrimitiveHandle {
-        self[layer].write(layer, info)
-    }
-
-    pub fn free(&mut self, h: &PrimitiveHandle) -> MaskIdx {
-        self[h.layer].free(h)
-    }
-
-    pub fn write_image(
-        &mut self,
-        layer: LayerId,
-        id: WidgetId,
-        texture_idx: u32,
-        region: UiRegion,
-        mask_idx: MaskIdx,
-        move_idx: MoveIdx,
-    ) -> PrimitiveHandle {
-        self[layer].write_image(layer, id, texture_idx, region, mask_idx, move_idx)
-    }
-}
-
 impl<T: Default> Default for Layers<T> {
     fn default() -> Self {
         Self::new()

iris/core/src/primitive/text.rs

@@ -2,7 +2,7 @@ use crate::{Align, GlyphAtlas, GlyphKey, PlacedGlyph, RegionAlign, Textures, UiC
 use parley::{
     Alignment, AlignmentOptions, FontContext, FontFamily, FontFamilyName, FontStyle, FontWeight,
     GenericFamily, Layout, LayoutContext, LineHeight, PositionedLayoutItem, StyleProperty,
-    fontique::{Blob, FamilyId},
+    fontique::Blob,
 };
 use std::ops::Range;
 use std::sync::Arc;
@@ -12,23 +12,16 @@ use swash::{
     zeno::{Format, Vector},
 };
 
-/// Bundled fonts, registered over the system collection rather than relied
-/// on alone -- see `TextData::register_bundled_fonts`'s doc comment for
-/// why. Static weight/style cuts, not a variable font: parley/fontique
-/// resolve a variable font's weight axis by picking normalized coordinates
-/// on whatever single face registers for the family, and a phone whose
-/// system "Roboto" is actually the variable "Roboto Flex" is exactly the
-/// device class this sidesteps, rather than depends on working correctly.
-/// Noto Sans, OFL-licensed (`assets/fonts/OFL.txt`), chosen for coverage
-/// breadth (a transcript's content is not known in advance) over a
-/// smaller-footprint alternative -- see the doc comment for the size this
-/// added.
-const NOTO_SANS_REGULAR: &[u8] = include_bytes!("../../assets/fonts/NotoSans-Regular.ttf");
-const NOTO_SANS_BOLD: &[u8] = include_bytes!("../../assets/fonts/NotoSans-Bold.ttf");
-const NOTO_SANS_ITALIC: &[u8] = include_bytes!("../../assets/fonts/NotoSans-Italic.ttf");
-const NOTO_SANS_BOLD_ITALIC: &[u8] = include_bytes!("../../assets/fonts/NotoSans-BoldItalic.ttf");
-const NOTO_SANS_MONO_REGULAR: &[u8] = include_bytes!("../../assets/fonts/NotoSansMono-Regular.ttf");
-const NOTO_SANS_MONO_BOLD: &[u8] = include_bytes!("../../assets/fonts/NotoSansMono-Bold.ttf");
+/// The icon font iris ships: the Nerd Fonts Symbols **Mono** subset built
+/// by `iris/core/build-icon-font.sh`, holding only the codepoints
+/// `crate::icon` names (992 bytes for three glyphs today).
+///
+/// This is the one font bundled here, and it is not a text font: body and
+/// monospace text still come from the platform's own collection
+/// (DECISIONS.md, 2026-09-07). An icon is the opposite case -- a small,
+/// closed set of codepoints no system font is guaranteed to have -- which
+/// is the same division the Compose app makes.
+const NERD_ICONS: &[u8] = include_bytes!("../../assets/fonts/nerd_icons.ttf");
 
 /// What starting up found about text rendering, for the on-screen
 /// Diagnostics page and the one startup log line (RUST.md's P0 box, "log
@@ -57,6 +50,11 @@ pub struct FontDiagnostics {
     pub bold_resolved: Option<String>,
     pub italic_resolved: Option<String>,
     pub mono_resolved: Option<String>,
+    /// The family the bundled icon font registered under, or `None` if
+    /// registering it failed. Reported rather than assumed: it is the one
+    /// font iris ships, so `None` is a broken build and must not look
+    /// like a device that happens to lack a face.
+    pub icon_family: Option<String>,
 }
 
 /// Everything text needs that outlives one string: the font collection, the
@@ -77,91 +75,175 @@ pub struct TextData {
     /// truth would mean carrying a `Painter` (or output size) into every
     /// input handler for the sake of one field.
     pub density: f32,
+    /// The family name [`NERD_ICONS`] registered under, which is what
+    /// [`Family::Icons`] resolves to. `None` only if registering the
+    /// bundled font failed, which is a broken build rather than a
+    /// platform difference -- said in the startup diagnostics rather than
+    /// silently drawn as tofu.
+    pub icon_family: Option<String>,
 }
 
 impl Default for TextData {
+    /// Text comes entirely from the platform's own font collection --
+    /// `FontContext::new()` builds a `fontique::Collection` with
+    /// `CollectionOptions::system_fonts` on by default, which is real
+    /// discovery on both targets this crate ships on: Android's backend
+    /// parses `/system/fonts` and `/system/etc/fonts.xml` and maps
+    /// `SansSerif`/`SystemUi` to `["Roboto Flex", "Roboto", "Noto Sans"]`
+    /// and `Monospace` to the platform's `"monospace"` alias; the desktop
+    /// build's backend is fontconfig. No font is bundled or registered
+    /// here -- see DECISIONS.md's 2026-09-07 entry for why (matching what
+    /// the Compose app does: it takes body/monospace text from
+    /// `FontFamily.Default`/`FontFamily.Monospace`, i.e. Android's Roboto
+    /// and its platform monospace face, and ships no text font of its own,
+    /// only its committed Nerd Fonts icon subset for fixed glyphs).
     fn default() -> Self {
-        let mut data = Self {
-            font_cx: FontContext::new(),
+        let mut font_cx = FontContext::new();
+        patch_android_monospace(&mut font_cx);
+        let icon_family = register_icon_font(&mut font_cx);
+        Self {
+            font_cx,
             layout_cx: LayoutContext::new(),
             scale_cx: ScaleContext::new(),
             atlas: GlyphAtlas::default(),
             density: 1.0,
-        };
-        data.register_bundled_fonts();
-        data
+            icon_family,
+        }
     }
 }
 
-impl TextData {
-    /// Registers Noto Sans (regular/bold/italic/bold-italic) and Noto Sans
-    /// Mono (regular/bold) as static faces, and puts them **first** in the
-    /// `SansSerif`/`Monospace` generic-family fallback lists -- ahead of,
-    /// not instead of, whatever the platform already found, so a script
-    /// Noto Sans lacks (CJK, emoji, ...) still falls through to the system
-    /// font the same as before this existed.
-    ///
-    /// Exists because text rendering must not depend on the platform's own
-    /// font enumeration succeeding or resolving weight/style the way this
-    /// crate assumes: RUST.md's P0 box found bold spans on a real phone
-    /// rendering as blank gaps of the correct advance width (the glyph
-    /// simply wasn't rasterised -- `TextData::place`'s `None` arm), while
-    /// the emulator's system fonts happened to resolve every style. A
-    /// bundled, static-per-style family removes fontique's Android font
-    /// scan (`fontique::backend::android::SystemFonts::new`, which parses
-    /// `/system/fonts` and `/system/etc/fonts.xml`) from the path a glyph
-    /// has to survive to reach the screen at all.
-    ///
-    /// Cost: six static `.ttf`s, ~3.6 MB uncompressed
-    /// (`iris/core/assets/fonts/`), landing in the APK compressed --
-    /// `build-apk.sh`'s own output is what says the delivered number, not
-    /// this comment.
-    fn register_bundled_fonts(&mut self) {
-        fn register(cx: &mut FontContext, bytes: &'static [u8]) -> Option<FamilyId> {
-            let blob = Blob::new(Arc::new(bytes));
-            cx.collection
-                .register_fonts(blob, None)
-                .into_iter()
-                .map(|(id, _)| id)
-                .next()
-        }
-        let sans_id = register(&mut self.font_cx, NOTO_SANS_REGULAR);
-        register(&mut self.font_cx, NOTO_SANS_BOLD);
-        register(&mut self.font_cx, NOTO_SANS_ITALIC);
-        register(&mut self.font_cx, NOTO_SANS_BOLD_ITALIC);
-        let mono_id = register(&mut self.font_cx, NOTO_SANS_MONO_REGULAR);
-        register(&mut self.font_cx, NOTO_SANS_MONO_BOLD);
+/// Registers the bundled icon font as an ordinary named family and
+/// answers the name it registered under -- read back from the collection
+/// rather than written down here, so the name cannot drift from the file
+/// (`build-icon-font.sh` takes whatever face the Nerd Fonts release
+/// ships).
+///
+/// A *named* family rather than a generic one: nothing should fall back
+/// to it for ordinary text, and nothing should fall back out of it for an
+/// icon -- a system face that happens to have one of these codepoints
+/// would draw somebody else's picture.
+fn register_icon_font(font_cx: &mut FontContext) -> Option<String> {
+    let blob = Blob::new(Arc::new(NERD_ICONS));
+    let id = font_cx
+        .collection
+        .register_fonts(blob, None)
+        .into_iter()
+        .map(|(id, _)| id)
+        .next()?;
+    font_cx.collection.family_name(id).map(str::to_string)
+}
 
-        if let Some(sans_id) = sans_id {
-            let existing: Vec<_> = self
-                .font_cx
+/// Works around `fontique` 0.11.1's Android backend never resolving
+/// `GenericFamily::Monospace` (confirmed against
+/// `fontique-0.11.1/src/backend/android.rs`'s `SystemFonts::new`, and still
+/// present on `linebender/parley`'s `main` as of 2026-09-07, so there is no
+/// released fix to bump to yet -- see DECISIONS.md's 2026-09-07 entry,
+/// "Platform fonts," for the full account). Two bugs stack, not one:
+/// `DEFAULT_GENERIC_FAMILIES` looks up the name `"monospace"` *before*
+/// `fonts.xml` is parsed into that same name map, and even after parsing,
+/// AOSP's `fonts.xml` names it with a `<family name="monospace">` element
+/// (not an `<alias>`) whose `<font>` children the backend's own parser
+/// does not read (a `TODO` in that match arm) -- so the name gets a
+/// `FamilyId` with no font data behind it, and `family_by_name("monospace")`
+/// comes back empty too. Confirmed on this checkout's emulator: `adb pull
+/// /system/etc/fonts.xml` shows
+/// `<family name="monospace"><font weight="400"
+/// style="normal">DroidSansMono.ttf</font></family>` with no matching
+/// alias.
+///
+/// So this reads `fonts.xml` itself (already on-device, already the
+/// authority Compose's own `Typeface.MONOSPACE` resolves through) for the
+/// filename that declaration names, then finds which of fontique's
+/// *actually* scanned families (from `/system/fonts`, which do carry real
+/// font data, just under whatever name the font's own metadata gives it --
+/// "Droid Sans Mono" here, but that name is never hardcoded) owns a font
+/// file with that name, and registers that family as the `Monospace`
+/// generic the way the backend itself would have if its parser had reified
+/// the declaration. A no-op if the family is somehow already resolved
+/// (future fontique) or nothing matches (no `fonts.xml`, e.g. a headless
+/// test, or a device that names it some other way).
+#[cfg(target_os = "android")]
+fn patch_android_monospace(font_cx: &mut FontContext) {
+    use parley::fontique::SourceKind;
+
+    let already_resolved = font_cx
+        .collection
+        .generic_families(GenericFamily::Monospace)
+        .next()
+        .is_some();
+    if already_resolved {
+        return;
+    }
+    let Some(target_file) = android_monospace_font_filename() else {
+        return;
+    };
+    let names: Vec<String> = font_cx
+        .collection
+        .family_names()
+        .map(str::to_string)
+        .collect();
+    for name in names {
+        let Some(id) = font_cx.collection.family_id(&name) else {
+            continue;
+        };
+        let Some(info) = font_cx.collection.family(id) else {
+            continue;
+        };
+        let Some(font) = info.default_font() else {
+            continue;
+        };
+        let SourceKind::Path(path) = font.source().kind() else {
+            continue;
+        };
+        if path.file_name().and_then(|f| f.to_str()) == Some(target_file.as_str()) {
+            font_cx
                 .collection
-                .generic_families(GenericFamily::SansSerif)
-                .collect();
-            self.font_cx.collection.set_generic_families(
-                GenericFamily::SansSerif,
-                std::iter::once(sans_id).chain(existing),
-            );
-            let existing: Vec<_> = self
-                .font_cx
-                .collection
-                .generic_families(GenericFamily::SystemUi)
-                .collect();
-            self.font_cx.collection.set_generic_families(
-                GenericFamily::SystemUi,
-                std::iter::once(sans_id).chain(existing),
-            );
+                .append_generic_families(GenericFamily::Monospace, std::iter::once(id));
+            return;
         }
-        if let Some(mono_id) = mono_id {
-            let existing: Vec<_> = self
-                .font_cx
-                .collection
-                .generic_families(GenericFamily::Monospace)
-                .collect();
-            self.font_cx.collection.set_generic_families(
-                GenericFamily::Monospace,
-                std::iter::once(mono_id).chain(existing),
-            );
+    }
+}
+
+/// Reads the font filename `fonts.xml` names for its `"monospace"` family
+/// (e.g. `"DroidSansMono.ttf"`), by plain substring search rather than a
+/// real XML parser -- a new dependency for one well-known, stable AOSP file
+/// whose structure fontique itself already parses with a full parser one
+/// module over. Not a general XML reader; assumes the file has exactly one
+/// `<family name="monospace">` element with at least one `<font>` child,
+/// which is the format on every AOSP `fonts.xml` this was checked against.
+#[cfg(target_os = "android")]
+fn android_monospace_font_filename() -> Option<String> {
+    let android_root = std::env::var("ANDROID_ROOT").unwrap_or_else(|_| "/system".to_string());
+    let xml =
+        std::fs::read_to_string(std::path::Path::new(&android_root).join("etc/fonts.xml")).ok()?;
+    let family_start = xml.find("<family name=\"monospace\">")?;
+    let block = &xml[family_start..];
+    let block = &block[..block.find("</family>")?];
+    let font_tag = block.find("<font")?;
+    let after_tag = &block[font_tag..];
+    let content_start = after_tag.find('>')? + 1;
+    let content = &after_tag[content_start..];
+    let filename = content[..content.find('<')?].trim();
+    (!filename.is_empty()).then(|| filename.to_string())
+}
+
+#[cfg(not(target_os = "android"))]
+fn patch_android_monospace(_font_cx: &mut FontContext) {}
+
+impl TextData {
+    /// [`Family::Icons`] as the name the bundled font actually registered
+    /// under; everything else unchanged.
+    ///
+    /// Cloned rather than borrowed because the caller needs it while the
+    /// layout builder holds `&mut self` -- a `String` per shaped icon run,
+    /// paid only when the layout is rebuilt.
+    pub fn resolve_family(&self, family: &Family) -> Family {
+        match family {
+            Family::Icons => self
+                .icon_family
+                .clone()
+                .map_or(Family::Icons, Family::Named),
+            other => other.clone(),
         }
     }
 
@@ -242,6 +324,7 @@ impl TextData {
             bold_resolved,
             italic_resolved,
             mono_resolved,
+            icon_family: self.icon_family.clone(),
         }
     }
 }
@@ -253,6 +336,11 @@ pub enum Family {
     SansSerif,
     Serif,
     Monospace,
+    /// The bundled icon font -- see [`crate::icon`] for what is in it.
+    /// Named as an intention rather than as a font name because only
+    /// [`TextData`] knows what the file registered as; it resolves this
+    /// during shaping ([`TextData::resolve_family`]).
+    Icons,
     Named(String),
 }
 
@@ -262,6 +350,11 @@ impl Family {
             Self::SansSerif => FontFamilyName::Generic(GenericFamily::SansSerif),
             Self::Serif => FontFamilyName::Generic(GenericFamily::Serif),
             Self::Monospace => FontFamilyName::Generic(GenericFamily::Monospace),
+            // Only reachable if `resolve_family` did not run, which no
+            // shaping path allows -- and sans-serif is the honest answer
+            // for a build whose icon font failed to register: the reader
+            // gets the platform's own tofu rather than a wrong picture.
+            Self::Icons => FontFamilyName::Generic(GenericFamily::SansSerif),
             Self::Named(name) => FontFamilyName::Named(name.as_str().into()),
         };
         FontFamily::Single(name)
@@ -460,21 +553,29 @@ impl TextBuffer {
         if self.shaped.as_ref() == Some(&(attrs.clone(), width, density)) {
             return;
         }
+        // Resolved before the builder borrows `data`: `Family::Icons`
+        // names an intention, and the name behind it lives on `TextData`.
+        let base_family = data.resolve_family(&attrs.family);
+        let span_families: Vec<Option<Family>> = self
+            .spans
+            .iter()
+            .map(|span| span.family.as_ref().map(|f| data.resolve_family(f)))
+            .collect();
         let mut builder = data
             .layout_cx
             .ranged_builder(&mut data.font_cx, &self.text, 1.0, true);
-        builder.push_default(StyleProperty::FontFamily(attrs.family.family()));
+        builder.push_default(StyleProperty::FontFamily(base_family.family()));
         builder.push_default(StyleProperty::FontSize(attrs.font_size * density));
         builder.push_default(StyleProperty::LineHeight(LineHeight::Absolute(
             attrs.line_height * density,
         )));
         builder.push_default(StyleProperty::Brush(attrs.color));
-        for span in &self.spans {
+        for (span, family) in self.spans.iter().zip(&span_families) {
             let range = span.range.clone();
             if let Some(color) = span.color {
                 builder.push(StyleProperty::Brush(color), range.clone());
             }
-            if let Some(family) = &span.family {
+            if let Some(family) = family {
                 builder.push(StyleProperty::FontFamily(family.family()), range.clone());
             }
             if let Some(size) = span.font_size {
@@ -628,3 +729,49 @@ impl TextData {
         }
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::icon;
+
+    /// Every codepoint `icon` names is actually in the subset the script
+    /// built. This is the failure `build-icon-font.sh`'s own comment warns
+    /// about -- a constant added on one side and not the other is a glyph
+    /// that silently isn't there -- and it is invisible at runtime,
+    /// because a missing glyph draws as nothing rather than as an error.
+    #[test]
+    fn every_icon_is_in_the_bundled_font() {
+        let font = FontRef::from_index(NERD_ICONS, 0).expect("the bundled icon font parses");
+        let charmap = font.charmap();
+        for (name, glyph) in [
+            ("OPEN", icon::OPEN),
+            ("CLOSED", icon::CLOSED),
+            ("COLLAPSE", icon::COLLAPSE),
+        ] {
+            let mut chars = glyph.chars();
+            let ch = chars.next().expect("an icon is one character");
+            assert!(chars.next().is_none(), "{name} is more than one character");
+            assert_ne!(
+                charmap.map(ch),
+                0,
+                "{name} (U+{:04X}) is not in nerd_icons.ttf -- add it to \
+                 build-icon-font.sh's GLYPHS and rerun the script",
+                ch as u32
+            );
+        }
+    }
+
+    /// The font registers, so `Family::Icons` resolves to a real family
+    /// rather than falling through to sans-serif and drawing tofu.
+    #[test]
+    fn the_icon_family_registers_and_resolves() {
+        let data = TextData::default();
+        let family = data.resolve_family(&Family::Icons);
+        assert!(
+            matches!(family, Family::Named(_)),
+            "the bundled icon font did not register: {:?}",
+            data.icon_family
+        );
+    }
+}

iris/core/src/primitive/texture.rs

@@ -1,6 +1,7 @@
 use crate::util::{RefCounter, Vec2};
 use image::{DynamicImage, GenericImageView};
 use std::{
+    collections::HashMap,
     ops::Index,
     sync::mpsc::{Receiver, Sender, channel},
 };
@@ -21,6 +22,16 @@ pub enum TextureKind {
     },
 }
 
+/// What a [`Textures::shared`] texture is a picture of -- exactly, not by
+/// hash: `owner` names the widget kind whose description it is, and `id`
+/// packs that description's own fields, so two owners cannot collide and
+/// a debugger shows which picture a slot holds.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub struct SharedTextureKey {
+    pub owner: &'static str,
+    pub id: u64,
+}
+
 #[derive(Debug, Clone)]
 pub struct TextureHandle {
     slot: u32,
@@ -35,6 +46,14 @@ pub struct TextureHandle {
 pub struct Textures {
     free: Vec<u32>,
     images: Vec<Option<DynamicImage>>,
+    /// What each slot is, kept beside the image so a slot can be pushed
+    /// again without the handle that knows -- see [`Textures::reupload`].
+    kinds: Vec<TextureKind>,
+    /// Textures built from a description rather than from a file, one per
+    /// distinct description: see [`Textures::shared`]. The map holds a
+    /// reference of its own, so a shared texture outlives every widget
+    /// drawing it and its slot is never recycled underneath one.
+    shared: HashMap<SharedTextureKey, TextureHandle>,
     /// Next layer to hand out to an atlas page. Pages are never freed (no
     /// atlas eviction), so this only grows and `free` never holds one.
     next_page_layer: u32,
@@ -77,6 +96,8 @@ impl Textures {
         Self {
             free: Vec::new(),
             images: Vec::new(),
+            kinds: Vec::new(),
+            shared: HashMap::new(),
             next_page_layer: 0,
             updates: Vec::new(),
             send,
@@ -119,16 +140,46 @@ impl Textures {
     fn push(&mut self, kind: TextureKind, image: DynamicImage) -> u32 {
         if let Some(i) = self.free.pop() {
             self.images[i as usize] = Some(image);
+            self.kinds[i as usize] = kind;
             self.updates.push(Update::Set(kind, i));
             i
         } else {
             let i = self.images.len() as u32;
             self.images.push(Some(image));
+            self.kinds.push(kind);
             self.updates.push(Update::Push(kind, i));
             i
         }
     }
 
+    /// The one texture for `key`, building it on the first ask and handing
+    /// out a further reference to it every time after.
+    ///
+    /// **Why this exists**: a texture rasterised from a *description* --
+    /// `widget::mark`'s triangle, from a direction and a colour -- has as
+    /// many copies as there are widgets asking for it, and each copy is
+    /// its own GPU texture, its own bind group and its own draw call. A
+    /// transcript screen with a folded card per tool call built one per
+    /// card: hundreds of 48x48 textures of three distinct pictures,
+    /// created and freed again as rows recycled. `make` is not called when
+    /// the key is already known, so the rasterising is paid once too.
+    ///
+    /// The map keeps its own reference for the life of the `Textures`, so
+    /// a shared slot is never freed and never reused for something else --
+    /// which is what makes a handle held by a long-lived widget safe.
+    pub fn shared(
+        &mut self,
+        key: SharedTextureKey,
+        make: impl FnOnce() -> DynamicImage,
+    ) -> TextureHandle {
+        if let Some(handle) = self.shared.get(&key) {
+            return handle.clone();
+        }
+        let handle = self.add(make());
+        self.shared.insert(key, handle.clone());
+        handle
+    }
+
     /// The stored image for a handle, to be written into before `patch`.
     pub fn image_mut(&mut self, handle: &TextureHandle) -> &mut DynamicImage {
         self.images[handle.slot as usize]
@@ -141,25 +192,35 @@ 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();
+    /// Queue every live slot for upload again, in slot order -- what a
+    /// genuinely new GPU device needs, in place of forgetting everything.
+    ///
+    /// A new device starts with no textures, and the renderer-side mirror
+    /// of these slots (`render::texture::GpuTextures`) starts empty with
+    /// it. What it must not do is start empty while the handles widgets
+    /// are still holding name slots by *index*: `Textures::reset` used to
+    /// throw this bookkeeping away, which left every live `TextureHandle`
+    /// -- one per `widget::mark`, hundreds on a transcript screen --
+    /// pointing at a slot nothing recognised, and the first frame after an
+    /// Android surface rebuild panicked in `image_bind_group` ("texture
+    /// slot 89 is not a live standalone image: None"). Re-uploading
+    /// instead keeps every index meaning what it meant, because this side
+    /// still holds the images: the slot list is rebuilt identically,
+    /// including the empty slots, which go across as `PushFree` so the
+    /// ones after them still land where they were.
+    ///
+    /// The glyph atlas comes back with it and is deliberately *not*
+    /// cleared any more: its pages are slots here, this side holds their
+    /// pixels, and re-uploading them restores exactly the atlas that was
+    /// there -- so an app switch no longer costs a re-rasterisation of
+    /// every glyph on screen either.
+    ///
+    /// Pending updates are dropped rather than kept: each is either a push
+    /// or a patch of a slot this replays in full.
+    pub fn reupload(&mut self) {
+        self.updates.clear();
+        self.updates
+            .extend((0..self.images.len() as u32).map(|i| Update::Push(self.kinds[i as usize], i)));
     }
 
     pub fn free(&mut self) {
@@ -245,3 +306,90 @@ impl Default for Textures {
         Self::new()
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use image::RgbaImage;
+
+    fn image(n: u32) -> DynamicImage {
+        RgbaImage::new(n, n).into()
+    }
+
+    fn key(id: u64) -> SharedTextureKey {
+        SharedTextureKey { owner: "test", id }
+    }
+
+    /// What `widget::mark` needs: one texture per description, however
+    /// many widgets ask for it, and a different description is a
+    /// different texture.
+    #[test]
+    fn a_shared_texture_is_built_once_and_handed_out_again() {
+        let mut textures = Textures::new();
+        let built = std::cell::Cell::new(0);
+        let make = |textures: &mut Textures, id: u64| {
+            textures.shared(key(id), || {
+                built.set(built.get() + 1);
+                image(4)
+            })
+        };
+        let first = make(&mut textures, 1);
+        let again = make(&mut textures, 1);
+        let other = make(&mut textures, 2);
+        assert_eq!(built.get(), 2, "the second ask for key 1 rasterised again");
+        assert_eq!(first.image_index(), again.image_index());
+        assert_ne!(first.image_index(), other.image_index());
+    }
+
+    /// The map's own reference is what keeps a shared slot alive: every
+    /// widget holding one can go away and the slot must not be recycled,
+    /// because the next widget to ask gets that same index back.
+    #[test]
+    fn a_shared_slot_is_not_freed_when_the_last_widget_drops_it() {
+        let mut textures = Textures::new();
+        let slot = textures.shared(key(1), || image(4)).image_index();
+        textures.free();
+        let plain = textures.add(image(4));
+        assert_ne!(
+            plain.image_index(),
+            slot,
+            "an ordinary texture was handed the shared mark's slot"
+        );
+    }
+
+    /// A new GPU device gets the same slot numbering back, so a handle a
+    /// widget has been holding all along still names its own texture --
+    /// the crash `reupload` replaced `reset` to fix.
+    #[test]
+    fn reupload_replays_every_slot_in_order_including_the_empty_ones() {
+        let mut textures = Textures::new();
+        let keep_a = textures.add(image(4));
+        let dropped = textures.add(image(4));
+        let keep_b = textures.add(image(4));
+        let (a, gone, b) = (
+            keep_a.image_index(),
+            dropped.image_index(),
+            keep_b.image_index(),
+        );
+        drop(dropped);
+        textures.free();
+        // Drain the updates so far, the way a frame does.
+        assert!(textures.updates().count() > 0);
+
+        textures.reupload();
+        let kinds: Vec<String> = textures
+            .updates()
+            .map(|u| match u {
+                TextureUpdate::Push(..) => "push".to_string(),
+                TextureUpdate::PushFree(..) => "push-free".to_string(),
+                _ => "other".to_string(),
+            })
+            .collect();
+        assert_eq!(
+            kinds,
+            ["push", "push-free", "push"],
+            "slots {a}, {gone} (freed) and {b} must replay in order, so the \
+             indices after a hole still land where they were"
+        );
+    }
+}

iris/core/src/render/data.rs

@@ -8,6 +8,15 @@ pub struct WindowUniform {
     pub height: f32,
 }
 
+/// One primitive's placement and what to draw there, in the one arena
+/// every layer shares (`Primitives`). Read from a storage buffer by
+/// **both** shader stages: the vertex stage for the corners of the
+/// primitive it is drawing, the fragment stage for the corners of a
+/// *mask's* primitive, which is generally a different one and often in
+/// another layer. A layer's vertex buffer carries only the slot
+/// ([`instance_slot_layout`]), so there is exactly one copy of a
+/// placement and a mask cannot disagree with what was drawn. See
+/// LAYOUT.md's "Masks with a shape".
 #[repr(C)]
 #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct PrimitiveInstance {
@@ -18,24 +27,17 @@ pub struct PrimitiveInstance {
     pub move_idx: MoveIdx,
 }
 
-impl PrimitiveInstance {
-    const ATTRIBS: [VertexAttribute; 8] = vertex_attr_array![
-        0 => Float32x2,
-        1 => Float32x2,
-        2 => Float32x2,
-        3 => Float32x2,
-        4 => Uint32,
-        5 => Uint32,
-        6 => Uint32,
-        7 => Uint32,
-    ];
-
-    pub fn desc() -> VertexBufferLayout<'static> {
-        VertexBufferLayout {
-            array_stride: std::mem::size_of::<Self>() as BufferAddress,
-            step_mode: VertexStepMode::Instance,
-            attributes: &Self::ATTRIBS,
-        }
+/// The vertex layout of a layer's draw order: one `u32` slot into the
+/// global instance arena per instance, stepped per instance. Everything a
+/// primitive is made of used to be here as eight vertex attributes; it
+/// moved into the storage buffer above so the fragment stage can read it
+/// too.
+pub fn instance_slot_layout() -> VertexBufferLayout<'static> {
+    const ATTRIBS: [VertexAttribute; 1] = vertex_attr_array![0 => Uint32];
+    VertexBufferLayout {
+        array_stride: std::mem::size_of::<u32>() as BufferAddress,
+        step_mode: VertexStepMode::Instance,
+        attributes: &ATTRIBS,
     }
 }
 
@@ -47,15 +49,53 @@ impl MaskIdx {
 
 pub type MoveIdx = Id<u32>;
 
+/// A clip, as a reference to a primitive already written plus the mask it
+/// nests inside. The fragment stage evaluates that primitive's coverage
+/// *at the masked pixel* -- for a rect, the same `rounded_rect_coverage`
+/// from the same SDF the rect itself is drawn with -- and multiplies it
+/// into the pixel's alpha, so a rounded container's corner and its
+/// children's clipped corner are the same arithmetic and cannot disagree.
+/// See LAYOUT.md's "Masks with a shape".
+///
+/// **No `kind` and no `flags`**, which the design sketched: the referenced
+/// instance already carries its own `binding`, and a copy of it here is a
+/// second thing to keep in step; alpha-only is the only mode there is, so
+/// there is nothing to select. Both are a field away if a second mode
+/// appears.
 #[repr(C)]
 #[derive(Debug, Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct Mask {
-    pub region: UiRegion,
-    /// The mask-owning widget's own move slot -- resolved in the fragment
-    /// shader against the same chain the vertex shader walks for a
-    /// primitive's own corners, so a mask and the content clipped by it
-    /// can move independently. See LAYOUT.md section 2b.
-    pub move_idx: MoveIdx,
+    /// The slot in `UiRenderState::primitives` of the primitive whose
+    /// coverage this mask is. Today always a `RectPrimitive`: a glyph or
+    /// a standalone image would need, respectively, a CPU-side alpha
+    /// plane for the hit test to agree with the shader, and a bind-group
+    /// switch the fragment stage cannot make -- `Painter::set_mask`
+    /// rejects both by name rather than leaving the shader to read a rect
+    /// that is not there.
+    ///
+    /// Who owns it depends on which way the mask was set. A plain
+    /// `.masked()` writes its own undrawn rect, so the primitive is in
+    /// the masking widget's `ActiveData::primitives` and lives exactly as
+    /// long as the mask. `.masked_by(shape)` points at a *child's*
+    /// primitive, which that child can free on any redraw of its own --
+    /// so `UiRenderState::remask_shape_users` marks the mask's owner for
+    /// redraw whenever a referenced slot is freed, since that widget's
+    /// own `set_mask` is the only thing that resolves the slot again.
+    pub primitive: u32,
+    /// The mask this one was set *inside* (`MaskIdx::NONE` at the top), so
+    /// clipping nests: the fragment stage walks the chain and multiplies
+    /// every coverage on it, which is what makes a pixel inside two
+    /// feathered corners dimmed by both. Chained rather than intersected
+    /// on the CPU because each mask moves with its own widget -- a code
+    /// fence inside a transcript row carries the row's scroll, the list's
+    /// own box does not, and one region resolved when the fence was last
+    /// drawn gets the second of those wrong as soon as the row moves.
+    ///
+    /// A child holds one ref on its parent's slot (`Painter::set_mask`),
+    /// released when the child's own slot goes
+    /// (`UiRenderState::remove`), so the chain cannot outlive what it
+    /// points at.
+    pub parent: MaskIdx,
 }
 
 /// One widget's cumulative on-screen translation, and the slot of the

iris/core/src/render/mod.rs

@@ -1,6 +1,10 @@
 use crate::{
     UiData, UiRenderState,
-    render::{data::PrimitiveInstance, texture::GpuTextures, util::ArrBuf},
+    render::{
+        data::{PrimitiveInstance, instance_slot_layout},
+        texture::GpuTextures,
+        util::ArrBuf,
+    },
     util::{HashMap, Vec2},
 };
 use data::WindowUniform;
@@ -14,6 +18,7 @@ mod atlas;
 mod data;
 mod frame_report;
 mod primitive;
+mod sdf;
 mod texture;
 mod util;
 
@@ -21,8 +26,14 @@ pub use atlas::*;
 pub use data::{Mask, MaskIdx, MoveIdx, MoveOffset};
 pub use frame_report::{FrameReport, FrameStats, JANK_THRESHOLD};
 pub use primitive::*;
+pub use sdf::{distance_from_rect, rounded_rect_coverage};
 
-const SHAPE_SHADER: &str = include_str!("./shader.wgsl");
+/// The one shader every primitive is drawn with. Public so a test can run
+/// a function out of it against the CPU transliteration in [`sdf`] --
+/// `iris/tests/mask_sdf.rs`, which LAYOUT.md's "Masks with a shape" turns
+/// on: a masked corner that cannot be tapped and a masked corner that is
+/// not drawn are only the same corner while the two agree.
+pub const SHAPE_SHADER: &str = include_str!("./shader.wgsl");
 
 /// The `wgpu::Limits` both platform backends (`android::render::
 /// AndroidRenderer::new`, `default::render::UiRenderer::new`) ask
@@ -120,6 +131,11 @@ impl WgpuErrorLog {
 pub struct UiRenderNode {
     uniform_group: BindGroup,
     primitive_layout: BindGroupLayout,
+    /// Group 1: `rects` and `glyphs`. Global and bound once per frame,
+    /// not per layer -- a mask referencing a rect drawn in another layer
+    /// has to be able to read it (see `Primitives`).
+    primitives: PrimitiveBuffers,
+    primitive_group: BindGroup,
     rsc_layout: BindGroupLayout,
     rsc_group: BindGroup,
 
@@ -129,6 +145,9 @@ pub struct UiRenderNode {
     active: Vec<usize>,
     window_buffer: Buffer,
     textures: GpuTextures,
+    /// Every primitive's placement, read by the vertex stage for the
+    /// primitive being drawn and by the fragment stage for a mask's.
+    instances: ArrBuf<PrimitiveInstance>,
     masks: ArrBuf<Mask>,
     move_offsets: ArrBuf<MoveOffset>,
     /// Group 3: the masks and move-offsets storage buffers, on their own --
@@ -146,16 +165,16 @@ pub struct UiRenderNode {
     masks_group: BindGroup,
 }
 
+/// One layer's vertex buffers: the slots it draws, in order. The
+/// primitives themselves are in `UiRenderNode::instances`.
 struct RenderLayer {
-    instance: ArrBuf<PrimitiveInstance>,
-    primitives: PrimitiveBuffers,
-    primitive_group: BindGroup,
-    /// A standalone image's instances, kept apart from `instance` because
-    /// each one draws with its own bind group -- see `UiRenderNode::draw`.
-    image_instance: ArrBuf<PrimitiveInstance>,
-    /// The texture slot each entry of `image_instance` draws with, in the
-    /// same order, refreshed alongside it. Not stored in the vertex buffer
-    /// itself because it names a bind group, not shader data.
+    order: ArrBuf<u32>,
+    /// A standalone image's slots, kept apart from `order` because each
+    /// one draws with its own bind group -- see `UiRenderNode::draw`.
+    images: ArrBuf<u32>,
+    /// The texture slot each entry of `images` draws with, in the same
+    /// order, refreshed alongside it. Not in the vertex buffer itself
+    /// because it names a bind group, not shader data.
     image_tex_indices: Vec<u32>,
 }
 
@@ -163,6 +182,8 @@ impl UiRenderNode {
     pub fn draw<'a>(&'a self, pass: &mut RenderPass<'a>) {
         pass.set_pipeline(&self.pipeline);
         pass.set_bind_group(0, &self.uniform_group, &[]);
+        // Group 1 is global now, so it is set here rather than per layer.
+        pass.set_bind_group(1, &self.primitive_group, &[]);
         // Set once, not per layer or per image: masks/move_offsets are read
         // by every primitive and every standalone image alike, and living
         // in their own group (rather than folded into group 2 alongside the
@@ -172,14 +193,13 @@ impl UiRenderNode {
         pass.set_bind_group(3, &self.masks_group, &[]);
         for i in &self.active {
             let layer = &self.layers[i];
-            if layer.instance.len() == 0 && layer.image_instance.len() == 0 {
+            if layer.order.len() == 0 && layer.images.len() == 0 {
                 continue;
             }
-            pass.set_bind_group(1, &layer.primitive_group, &[]);
-            if layer.instance.len() > 0 {
+            if layer.order.len() > 0 {
                 pass.set_bind_group(2, &self.rsc_group, &[]);
-                pass.set_vertex_buffer(0, layer.instance.buffer.slice(..));
-                pass.draw(0..4, 0..layer.instance.len() as u32);
+                pass.set_vertex_buffer(0, layer.order.buffer.slice(..));
+                pass.draw(0..4, 0..layer.order.len() as u32);
             }
             // Images draw after this layer's rects and glyphs, one draw call
             // each with its own bind group. That draws every image "on top"
@@ -188,8 +208,8 @@ impl UiRenderNode {
             // draw order was already undefined before images had their own
             // list -- nothing before this relied on interleaving a rect
             // between two images at a particular position.
-            if layer.image_instance.len() > 0 {
-                pass.set_vertex_buffer(0, layer.image_instance.buffer.slice(..));
+            if layer.images.len() > 0 {
+                pass.set_vertex_buffer(0, layer.images.buffer.slice(..));
                 for (k, &tex_idx) in layer.image_tex_indices.iter().enumerate() {
                     pass.set_bind_group(2, self.textures.image_bind_group(tex_idx), &[]);
                     pass.draw(0..4, k as u32..k as u32 + 1);
@@ -206,67 +226,45 @@ impl UiRenderNode {
         ui_render: &mut UiRenderState,
     ) -> FrameUpdateStats {
         self.active.clear();
-        for (i, primitives) in ui_render.layers.iter_mut() {
+        for (i, order) in ui_render.layers.iter_mut() {
             self.active.push(i);
-            for change in primitives.apply_free() {
-                if let Some(inst) = ui_render.active.get_mut(&change.id) {
-                    for h in &mut inst.primitives {
-                        // `is_image` disambiguates: `instances` and `images`
-                        // are separate lists with independent indices, so
-                        // without it a rect's renumbering could be applied to
-                        // an image handle that happened to share the same
-                        // (layer, inst_idx).
-                        if h.layer == i
-                            && h.inst_idx == change.old
-                            && (h.binding == IMAGE_BINDING) == change.is_image
-                        {
-                            h.inst_idx = change.new;
-                            break;
-                        }
-                    }
-                }
-            }
-            let rlayer = self.layers.entry(i).or_insert_with(|| {
-                let primitives = PrimitiveBuffers::new(device);
-                let primitive_group =
-                    Self::primitive_group(device, &self.primitive_layout, primitives.buffers());
-                RenderLayer {
-                    instance: ArrBuf::new(
-                        device,
-                        BufferUsages::VERTEX | BufferUsages::COPY_DST,
-                        "instance",
-                    ),
-                    primitives,
-                    primitive_group,
-                    image_instance: ArrBuf::new(
-                        device,
-                        BufferUsages::VERTEX | BufferUsages::COPY_DST,
-                        "image instance",
-                    ),
-                    image_tex_indices: Vec::new(),
-                }
-            });
-            if primitives.updated {
-                rlayer
-                    .instance
-                    .update(device, queue, primitives.instances());
-                rlayer.primitives.update(device, queue, primitives.data());
-                rlayer.primitive_group = Self::primitive_group(
+            let rlayer = self.layers.entry(i).or_insert_with(|| RenderLayer {
+                order: ArrBuf::new(
                     device,
-                    &self.primitive_layout,
-                    rlayer.primitives.buffers(),
-                );
-                rlayer
-                    .image_instance
-                    .update(device, queue, primitives.image_instances());
-                rlayer.image_tex_indices = primitives
-                    .image_instances()
+                    BufferUsages::VERTEX | BufferUsages::COPY_DST,
+                    "layer order",
+                ),
+                images: ArrBuf::new(
+                    device,
+                    BufferUsages::VERTEX | BufferUsages::COPY_DST,
+                    "layer image order",
+                ),
+                image_tex_indices: Vec::new(),
+            });
+            if order.updated {
+                rlayer.order.update(device, queue, order.order());
+                rlayer.images.update(device, queue, order.images());
+                rlayer.image_tex_indices = order
+                    .images()
                     .iter()
-                    .map(|inst| inst.idx)
+                    .map(|&slot| ui_render.primitives.instance(slot).idx)
                     .collect();
-                primitives.updated = false;
+                order.updated = false;
             }
         }
+        let instances_resized = if ui_render.primitives.updated {
+            ui_render.primitives.updated = false;
+            let resized = self
+                .instances
+                .update(device, queue, ui_render.primitives.instances());
+            self.primitives
+                .update(device, queue, ui_render.primitives.data());
+            self.primitive_group =
+                Self::primitive_group(device, &self.primitive_layout, self.primitives.buffers());
+            resized
+        } else {
+            false
+        };
         let masks_resized = if ui.masks.changed {
             ui.masks.changed = false;
             self.masks.update(device, queue, &ui.masks[..])
@@ -280,9 +278,14 @@ impl UiRenderNode {
         } else {
             false
         };
-        if masks_resized || moves_resized {
-            self.masks_group =
-                Self::masks_group(device, &self.masks_layout, &self.masks, &self.move_offsets);
+        if masks_resized || moves_resized || instances_resized {
+            self.masks_group = Self::masks_group(
+                device,
+                &self.masks_layout,
+                &self.masks,
+                &self.move_offsets,
+                &self.instances,
+            );
         }
         let rebuild_main = self.textures.update(&mut ui.textures, &self.rsc_layout);
         if rebuild_main {
@@ -408,6 +411,14 @@ impl UiRenderNode {
         });
 
         let tex_manager = GpuTextures::new(device, queue);
+        let primitives = PrimitiveBuffers::new(device);
+        let primitive_group =
+            Self::primitive_group(device, &primitive_layout, primitives.buffers());
+        let instances = ArrBuf::new(
+            device,
+            BufferUsages::STORAGE | BufferUsages::COPY_DST,
+            "ui instances",
+        );
         let masks = ArrBuf::new(
             device,
             BufferUsages::STORAGE | BufferUsages::COPY_DST,
@@ -422,7 +433,8 @@ impl UiRenderNode {
         let rsc_layout = Self::rsc_layout(device);
         let rsc_group = Self::rsc_group(device, &rsc_layout, &tex_manager);
         let masks_layout = Self::masks_layout(device);
-        let masks_group = Self::masks_group(device, &masks_layout, &masks, &move_offsets);
+        let masks_group =
+            Self::masks_group(device, &masks_layout, &masks, &move_offsets, &instances);
 
         let pipeline_layout = device.create_pipeline_layout(&PipelineLayoutDescriptor {
             label: Some("UI Shape Pipeline Layout"),
@@ -440,7 +452,7 @@ impl UiRenderNode {
             vertex: VertexState {
                 module: &shader,
                 entry_point: Some("vs_main"),
-                buffers: &[PrimitiveInstance::desc()],
+                buffers: &[instance_slot_layout()],
                 compilation_options: Default::default(),
             },
             fragment: Some(FragmentState {
@@ -486,6 +498,8 @@ impl UiRenderNode {
         Ok(Self {
             uniform_group,
             primitive_layout,
+            primitives,
+            primitive_group,
             rsc_layout,
             rsc_group,
             pipeline,
@@ -493,6 +507,7 @@ impl UiRenderNode {
             layers: HashMap::default(),
             active: Vec::new(),
             textures: tex_manager,
+            instances,
             masks,
             move_offsets,
             masks_layout,
@@ -627,6 +642,16 @@ impl UiRenderNode {
                     },
                     count: None,
                 },
+                BindGroupLayoutEntry {
+                    binding: 2,
+                    visibility: ShaderStages::VERTEX | ShaderStages::FRAGMENT,
+                    ty: BindingType::Buffer {
+                        ty: BufferBindingType::Storage { read_only: true },
+                        has_dynamic_offset: false,
+                        min_binding_size: None,
+                    },
+                    count: None,
+                },
             ],
             label: Some("ui masks"),
         })
@@ -637,6 +662,7 @@ impl UiRenderNode {
         layout: &BindGroupLayout,
         masks: &ArrBuf<Mask>,
         move_offsets: &ArrBuf<MoveOffset>,
+        instances: &ArrBuf<PrimitiveInstance>,
     ) -> BindGroup {
         device.create_bind_group(&BindGroupDescriptor {
             layout,
@@ -649,6 +675,10 @@ impl UiRenderNode {
                     binding: 1,
                     resource: move_offsets.buffer.as_entire_binding(),
                 },
+                BindGroupEntry {
+                    binding: 2,
+                    resource: instances.buffer.as_entire_binding(),
+                },
             ],
             label: Some("ui masks"),
         })

iris/core/src/render/primitive.rs

@@ -11,51 +11,20 @@ use crate::{
 use bytemuck::Pod;
 use wgpu::*;
 
-pub struct Primitives {
-    instances: Vec<PrimitiveInstance>,
-    assoc: Vec<WidgetId>,
-    data: PrimitiveData,
-    free: Vec<usize>,
-
-    /// Standalone images, kept apart from `instances` because each one draws
-    /// with its own bind group rather than sharing the layer's one instanced
-    /// draw -- see TEXTURES.md's "Recommended shape". `idx` on each
-    /// `PrimitiveInstance` here is the texture's slot in `Textures`/
-    /// `GpuTextures`, not an index into `data`; there is no per-image entry
-    /// in `data` because a bind group already picks the texture; nothing
-    /// left to look up per-instance.
-    images: Vec<PrimitiveInstance>,
-    image_assoc: Vec<WidgetId>,
-    image_free: Vec<usize>,
-
-    pub updated: bool,
-}
-
-impl Default for Primitives {
-    fn default() -> Self {
-        Self {
-            instances: Default::default(),
-            assoc: Default::default(),
-            data: Default::default(),
-            free: Vec::new(),
-            images: Default::default(),
-            image_assoc: Default::default(),
-            image_free: Vec::new(),
-            updated: true,
-        }
-    }
-}
-
 /// The `binding` tag `Painter` writes on an image instance. Distinct from any
 /// `Primitive::BINDING` because images have no `PrimitiveData` entry to key
 /// one from -- a bind group already selects the texture -- so this only ever
-/// has to match the shader's `TEXTURE` constant and flag "this instance lives
-/// in `Primitives::images`, not `Primitives::instances`" to the code below.
+/// has to match the shader's `TEXTURE` constant and flag "this instance is
+/// drawn with its own bind group" to the code below.
 pub const IMAGE_BINDING: u32 = 1;
 
 pub trait Primitive: Pod {
     const BINDING: u32;
     fn vec(data: &mut PrimitiveData) -> &mut PrimitiveVec<Self>;
+    /// The read-only half of [`Self::vec`], for a caller that wants to
+    /// look one entry up rather than write one -- a mask reading the
+    /// radius of the rect it clips to ([`Primitives::data`]).
+    fn vec_ref(data: &PrimitiveData) -> &PrimitiveVec<Self>;
 }
 
 macro_rules! primitives {
@@ -121,6 +90,9 @@ macro_rules! primitives {
                 fn vec(data: &mut PrimitiveData) -> &mut PrimitiveVec<Self> {
                     &mut data.$name
                 }
+                fn vec_ref(data: &PrimitiveData) -> &PrimitiveVec<Self> {
+                    &data.$name
+                }
             }
         )*
     };
@@ -134,18 +106,61 @@ macro_rules! primitives {
     (@count $t:tt) => { 1 };
 }
 
-pub struct PrimitiveInst<P> {
-    pub id: WidgetId,
-    pub primitive: P,
-    pub region: UiRegion,
-    pub mask_idx: MaskIdx,
-    pub move_idx: MoveIdx,
+/// Every primitive instance in the tree, in one arena that all layers
+/// share, plus the per-primitive data (`rects`, `glyphs`) they index.
+///
+/// **Why one arena rather than one per layer**, which is what this was:
+/// the fragment stage evaluates a *mask's* primitive at the masked pixel
+/// (LAYOUT.md's "Masks with a shape"), and the widget that owns a mask is
+/// routinely in a different layer from the content it clips -- a rounded
+/// container in one layer, a `Stack`'s child content in the layer below.
+/// A per-layer buffer cannot answer that lookup at all: only one layer's
+/// group is bound at a time, so the mask would silently read another
+/// layer's rect. Both buffers are therefore global and bound once per
+/// frame, and a layer keeps only its draw *order* ([`LayerOrder`]).
+///
+/// Slots are stable for a primitive's whole life: nothing here is
+/// compacted, so a `Mask` can hold a slot across frames.
+pub struct Primitives {
+    instances: Vec<PrimitiveInstance>,
+    assoc: Vec<WidgetId>,
+    /// Slots freed since the last [`Self::apply_free`]. Deliberately not
+    /// reusable yet: the layer that drew one still names it in its draw
+    /// order until that call compacts the order, so handing it out again
+    /// first would draw the new primitive twice -- once through the stale
+    /// order entry and once through the new one.
+    freed: Vec<usize>,
+    /// Slots [`Self::apply_free`] released, which is what [`Self::alloc`]
+    /// hands out.
+    reusable: Vec<usize>,
+    data: PrimitiveData,
+    /// Whether the instance arena or the per-primitive data changed since
+    /// the last upload -- one flag for both, since they are uploaded
+    /// together.
+    pub updated: bool,
+}
+
+impl Default for Primitives {
+    fn default() -> Self {
+        Self {
+            instances: Default::default(),
+            assoc: Default::default(),
+            freed: Vec::new(),
+            reusable: Vec::new(),
+            data: Default::default(),
+            updated: true,
+        }
+    }
 }
 
 impl Primitives {
-    pub fn write<P: Primitive>(
+    /// Writes a primitive into the arena and hands back its slot and its
+    /// entry in the per-primitive data. The caller (`UiRenderState`) puts
+    /// the slot into a layer's draw order -- an instance that no layer
+    /// names is never rasterized, which is what a mask shape drawn only to
+    /// be *referenced* uses.
+    pub fn alloc<P: Primitive>(
         &mut self,
-        layer: usize,
         PrimitiveInst {
             id,
             primitive,
@@ -153,154 +168,118 @@ impl Primitives {
             mask_idx,
             move_idx,
         }: PrimitiveInst<P>,
-    ) -> PrimitiveHandle {
-        self.updated = true;
-        let vec = P::vec(&mut self.data);
-        let i = vec.add(primitive);
-        let inst = PrimitiveInstance {
-            region,
-            idx: i as u32,
-            mask_idx,
-            move_idx,
-            binding: P::BINDING,
-        };
-        let inst_i = if let Some(i) = self.free.pop() {
-            self.instances[i] = inst;
-            self.assoc[i] = id;
-            i
-        } else {
-            let i = self.instances.len();
-            self.instances.push(inst);
-            self.assoc.push(id);
-            i
-        };
-        PrimitiveHandle::new::<P>(layer, inst_i, i)
+    ) -> (u32, usize) {
+        let data_idx = P::vec(&mut self.data).add(primitive);
+        let slot = self.push(
+            PrimitiveInstance {
+                region,
+                idx: data_idx as u32,
+                mask_idx,
+                move_idx,
+                binding: P::BINDING,
+            },
+            id,
+        );
+        (slot, data_idx)
     }
 
-    /// Writes an image instance directly -- there is no `Primitive` impl for
-    /// it to go through `write`, since it has nowhere in `PrimitiveData` to
-    /// put a per-instance entry. `texture_idx` is the slot the bind group at
-    /// draw time is chosen from, carried in the otherwise-unused `idx` field.
-    pub fn write_image(
+    /// A standalone image, which has no `PrimitiveData` entry to allocate
+    /// -- its bind group already picks the texture, so `texture_idx` rides
+    /// in the otherwise-unused `idx` field and names the bind group the
+    /// draw call selects.
+    pub fn alloc_image(
         &mut self,
-        layer: usize,
         id: WidgetId,
         texture_idx: u32,
         region: UiRegion,
         mask_idx: MaskIdx,
         move_idx: MoveIdx,
-    ) -> PrimitiveHandle {
+    ) -> u32 {
+        self.push(
+            PrimitiveInstance {
+                region,
+                idx: texture_idx,
+                mask_idx,
+                move_idx,
+                binding: IMAGE_BINDING,
+            },
+            id,
+        )
+    }
+
+    fn push(&mut self, inst: PrimitiveInstance, id: WidgetId) -> u32 {
         self.updated = true;
-        let inst = PrimitiveInstance {
-            region,
-            idx: texture_idx,
-            mask_idx,
-            move_idx,
-            binding: IMAGE_BINDING,
-        };
-        let inst_i = if let Some(i) = self.image_free.pop() {
-            self.images[i] = inst;
-            self.image_assoc[i] = id;
+        let slot = if let Some(i) = self.reusable.pop() {
+            self.instances[i] = inst;
+            self.assoc[i] = id;
             i
         } else {
-            let i = self.images.len();
-            self.images.push(inst);
-            self.image_assoc.push(id);
-            i
+            self.instances.push(inst);
+            self.assoc.push(id);
+            self.instances.len() - 1
         };
-        PrimitiveHandle {
-            layer,
-            inst_idx: inst_i,
-            data_idx: 0,
-            binding: IMAGE_BINDING,
-        }
-    }
-
-    pub fn image_instances(&self) -> &Vec<PrimitiveInstance> {
-        &self.images
-    }
-
-    /// returns (old index, new index) for both lists this layer keeps --
-    /// `PrimitiveChange::is_image` says which, since the two have separate
-    /// index spaces and `old`/`new` alone would collide between them.
-    ///
-    /// Both lists free with `swap_remove`, so a layer's draw order was
-    /// already undefined before images existed: nothing here may assume one
-    /// primitive stays adjacent to another once anything in the layer has
-    /// been freed.
-    pub fn apply_free(&mut self) -> Vec<PrimitiveChange> {
-        let mut changes =
-            Self::apply_free_list(&mut self.free, &mut self.instances, &mut self.assoc, false);
-        changes.extend(Self::apply_free_list(
-            &mut self.image_free,
-            &mut self.images,
-            &mut self.image_assoc,
-            true,
-        ));
-        changes
-    }
-
-    fn apply_free_list(
-        free: &mut Vec<usize>,
-        instances: &mut Vec<PrimitiveInstance>,
-        assoc: &mut Vec<WidgetId>,
-        is_image: bool,
-    ) -> Vec<PrimitiveChange> {
-        free.sort_by(|a, b| b.cmp(a));
-        free.drain(..)
-            .filter_map(|i| {
-                instances.swap_remove(i);
-                assoc.swap_remove(i);
-                if i == instances.len() {
-                    return None;
-                }
-                let id = assoc[i];
-                let old = instances.len();
-                Some(PrimitiveChange {
-                    id,
-                    is_image,
-                    old,
-                    new: i,
-                })
-            })
-            .collect()
+        slot as u32
     }
 
+    /// Retires a slot, answering the mask it was drawn under so the caller
+    /// can drop that mask's ref. The slot itself only becomes reusable at
+    /// the next [`Self::apply_free`] -- see `freed`.
     pub fn free(&mut self, h: &PrimitiveHandle) -> MaskIdx {
         self.updated = true;
-        if h.binding == IMAGE_BINDING {
-            self.image_free.push(h.inst_idx);
-            self.images[h.inst_idx].mask_idx
-        } else {
+        let slot = h.slot as usize;
+        if h.binding != IMAGE_BINDING {
             self.data.free(h.binding, h.data_idx);
-            self.free.push(h.inst_idx);
-            self.instances[h.inst_idx].mask_idx
         }
+        self.freed.push(slot);
+        self.instances[slot].mask_idx
     }
 
-    /// How many instances are still bound for the GPU -- the O(1) half of
-    /// the orphan check, so the O(primitives) walk below only runs on a
-    /// frame that already looks wrong. See
+    /// Hands this frame's freed slots back for reuse. Called once per
+    /// frame from `UiRenderState::update`, **after** every layer has
+    /// compacted its draw order, since that order is the only thing still
+    /// naming them.
+    pub fn release_freed(&mut self) {
+        self.reusable.append(&mut self.freed);
+    }
+
+    /// Which widget drew the primitive in `slot` -- how a draw-order
+    /// change finds the handle it has to renumber.
+    pub fn owner(&self, slot: u32) -> WidgetId {
+        self.assoc[slot as usize]
+    }
+
+    pub fn clear(&mut self) {
+        self.updated = true;
+        self.instances.clear();
+        self.assoc.clear();
+        self.freed.clear();
+        self.reusable.clear();
+        self.data.clear();
+    }
+
+    /// How many instances are still live -- the O(1) half of the orphan
+    /// check, so the O(primitives) walk below only runs on a frame that
+    /// already looks wrong. See
     /// [`crate::UiRenderState::orphaned_primitives`].
     pub fn live_count(&self) -> usize {
-        (self.instances.len() - self.free.len()) + (self.images.len() - self.image_free.len())
+        self.instances.len() - self.freed.len() - self.reusable.len()
     }
 
-    /// Every instance that is still bound for the GPU, as `(inst_idx,
-    /// owner, is_image)` -- everything except the slots already handed to
-    /// [`Self::free`] and waiting for [`Self::apply_free`] to compact them
-    /// away. Only [`crate::UiRenderState::orphaned_primitives`] uses this,
-    /// to check that every drawn primitive still belongs to a live widget.
-    pub fn live_instances(&self) -> impl Iterator<Item = (usize, WidgetId, bool)> + '_ {
-        let free: HashSet<usize> = self.free.iter().copied().collect();
-        let image_free: HashSet<usize> = self.image_free.iter().copied().collect();
-        let rects = (0..self.instances.len())
-            .filter(move |i| !free.contains(i))
-            .map(|i| (i, self.assoc[i], false));
-        let images = (0..self.images.len())
-            .filter(move |i| !image_free.contains(i))
-            .map(|i| (i, self.image_assoc[i], true));
-        rects.chain(images)
+    /// Every live instance as `(slot, owner, is_image)` -- everything
+    /// except the freed and the reusable. Only
+    /// [`crate::UiRenderState::orphaned_primitives`] uses this, to check
+    /// that every live primitive still belongs to a live widget.
+    pub fn live_instances(&self) -> impl Iterator<Item = (u32, WidgetId, bool)> + '_ {
+        let dead: HashSet<usize> = self.freed.iter().chain(&self.reusable).copied().collect();
+        (0..self.instances.len())
+            .filter(move |i| !dead.contains(i))
+            .map(|i| {
+                (
+                    i as u32,
+                    self.assoc[i],
+                    self.instances[i].binding == IMAGE_BINDING,
+                )
+            })
     }
 
     pub fn data(&self) -> &PrimitiveData {
@@ -311,46 +290,163 @@ impl Primitives {
         &self.instances
     }
 
+    pub fn instance(&self, slot: u32) -> &PrimitiveInstance {
+        &self.instances[slot as usize]
+    }
+
+    /// The per-primitive data behind `slot`, or `None` if that slot holds
+    /// a different kind of primitive -- the `binding` check is the same
+    /// one the shader's dispatch switch makes, and it is what stops a
+    /// caller reading a glyph's index into the rect table.
+    pub fn primitive_data<P: Primitive>(&self, slot: u32) -> Option<&P> {
+        let inst = self.instance(slot);
+        (inst.binding == P::BINDING).then(|| &P::vec_ref(&self.data)[inst.idx as usize])
+    }
+
     pub fn region_mut(&mut self, h: &PrimitiveHandle) -> &mut UiRegion {
         self.updated = true;
-        if h.binding == IMAGE_BINDING {
-            &mut self.images[h.inst_idx].region
-        } else {
-            &mut self.instances[h.inst_idx].region
-        }
+        &mut self.instances[h.slot as usize].region
     }
 }
 
-pub struct PrimitiveChange {
-    pub id: WidgetId,
-    /// Which of `Primitives::instances`/`Primitives::images` this change
-    /// belongs to -- their `old`/`new` indices are independent, so a
-    /// consumer matching only on `(layer, inst_idx)` could apply an image's
-    /// renumbering to a rect's handle that happens to share the same index.
-    pub is_image: bool,
-    pub old: usize,
-    pub new: usize,
+/// One layer's draw order: the slots of the global arena it draws, in the
+/// order they were written. The vertex buffer of a layer is exactly this.
+///
+/// Both lists free with `swap_remove`, so a layer's draw order was already
+/// undefined before this split: nothing here may assume one primitive
+/// stays adjacent to another once anything in the layer has been freed.
+#[derive(Default)]
+pub struct LayerOrder {
+    order: Vec<u32>,
+    /// Standalone images, kept apart because each draws with its own bind
+    /// group rather than sharing the layer's one instanced draw -- see
+    /// `UiRenderNode::draw`.
+    images: Vec<u32>,
+    free: Vec<usize>,
+    image_free: Vec<usize>,
+    pub updated: bool,
 }
 
+impl LayerOrder {
+    pub fn push(&mut self, slot: u32, is_image: bool) -> usize {
+        self.updated = true;
+        let list = if is_image {
+            &mut self.images
+        } else {
+            &mut self.order
+        };
+        list.push(slot);
+        list.len() - 1
+    }
+
+    /// Marks a position for removal. Deferred to [`Self::apply_free`] like
+    /// the arena's own, so that a position is only renumbered once per
+    /// frame however many were dropped.
+    pub fn free(&mut self, pos: usize, is_image: bool) {
+        self.updated = true;
+        if is_image {
+            self.image_free.push(pos);
+        } else {
+            self.free.push(pos);
+        }
+    }
+
+    /// Compacts both lists, answering every primitive whose position
+    /// moved so its handle can be corrected.
+    pub fn apply_free(&mut self) -> Vec<OrderChange> {
+        let mut changes = Self::apply_free_list(&mut self.free, &mut self.order, false);
+        changes.extend(Self::apply_free_list(
+            &mut self.image_free,
+            &mut self.images,
+            true,
+        ));
+        changes
+    }
+
+    fn apply_free_list(
+        free: &mut Vec<usize>,
+        list: &mut Vec<u32>,
+        is_image: bool,
+    ) -> Vec<OrderChange> {
+        // Descending, so removing a contiguous tail costs no renumbering
+        // at all -- which is what freeing one widget's primitives is.
+        free.sort_by(|a, b| b.cmp(a));
+        free.drain(..)
+            .filter_map(|pos| {
+                list.swap_remove(pos);
+                if pos == list.len() {
+                    return None;
+                }
+                Some(OrderChange {
+                    slot: list[pos],
+                    is_image,
+                    pos,
+                })
+            })
+            .collect()
+    }
+
+    pub fn order(&self) -> &Vec<u32> {
+        &self.order
+    }
+
+    pub fn images(&self) -> &Vec<u32> {
+        &self.images
+    }
+}
+
+/// A primitive whose position in a layer's draw order moved when
+/// something before it was freed -- `slot` names which primitive, so its
+/// owner's handle can be found and pointed at `pos`.
+pub struct OrderChange {
+    pub slot: u32,
+    /// Which of the layer's two lists moved: their positions are
+    /// independent index spaces, so a handle matching on position alone
+    /// could take an image's renumbering for a rect's.
+    pub is_image: bool,
+    pub pos: usize,
+}
+
+/// Whether a primitive goes into its layer's draw order. [`Drawn::No`] is
+/// a primitive written only to be *referenced* -- a mask's shape
+/// (LAYOUT.md's "Masks with a shape"). It is owned, moved, resized and
+/// freed exactly like any other; it is simply never rasterized.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum Drawn {
+    Yes,
+    No,
+}
+
+/// The `pos` of a [`Drawn::No`] primitive: it is in no layer's order, so
+/// there is no position to renumber or free.
+pub const NOT_DRAWN: usize = usize::MAX;
+
+/// Where one primitive lives: its stable slot in the global arena, and
+/// where in a layer's draw order it currently sits ([`NOT_DRAWN`] if it is
+/// only referenced).
 #[derive(Debug)]
 pub struct PrimitiveHandle {
     pub layer: usize,
-    pub inst_idx: usize,
+    pub pos: usize,
+    pub slot: u32,
     pub data_idx: usize,
     pub binding: u32,
 }
 
 impl PrimitiveHandle {
-    fn new<P: Primitive>(layer: usize, inst_idx: usize, data_idx: usize) -> Self {
-        Self {
-            layer,
-            inst_idx,
-            data_idx,
-            binding: P::BINDING,
-        }
+    pub fn is_image(&self) -> bool {
+        self.binding == IMAGE_BINDING
     }
 }
 
+pub struct PrimitiveInst<P> {
+    pub id: WidgetId,
+    pub primitive: P,
+    pub region: UiRegion,
+    pub mask_idx: MaskIdx,
+    pub move_idx: MoveIdx,
+}
+
 primitives!(
     rects: RectPrimitive => 0,
     glyphs: GlyphPrimitive => 2,

iris/core/src/render/sdf.rs

@@ -0,0 +1,54 @@
+//! The rounded-rect coverage function, on the CPU.
+//!
+//! `shader.wgsl`'s `distance_from_rect`/`rounded_rect_coverage` are a
+//! transliteration of these two, line for line, and
+//! `mask_sdf_matches_the_shader` in `iris`'s layout tests compares the two
+//! at a grid of points against values the shader itself produced. They are
+//! kept together here, in the crate both a renderer and a hit test can
+//! reach, because LAYOUT.md's "Masks with a shape" turns on the two
+//! agreeing: a masked corner that cannot be tapped and a masked corner
+//! that is not drawn have to be the same corner, and they are only the
+//! same corner while one function decides both.
+//!
+//! Window pixels throughout, matching the shader's `pos` -- not `UiRegion`
+//! units, which the shader has already resolved by the time it evaluates
+//! this.
+
+use crate::util::Vec2;
+
+/// The signed distance from `pos` to a rounded rect given by its centre,
+/// its corner offset (half its size) and its corner `radius`. Negative
+/// inside.
+pub fn distance_from_rect(pos: Vec2, center: Vec2, corner: Vec2, radius: f32) -> f32 {
+    // vec from center to pixel
+    let p = pos - center;
+    // vec from inner rect corner to pixel
+    let q = Vec2::new(
+        p.x.abs() - (corner.x - radius),
+        p.y.abs() - (corner.y - radius),
+    );
+    let clamped = Vec2::new(q.x.max(0.0), q.y.max(0.0));
+    (clamped.x * clamped.x + clamped.y * clamped.y).sqrt() - radius
+}
+
+/// How much of the pixel at `pos` a rounded rect covers, anti-aliased over
+/// the half-pixel either side of its edge: 1 well inside, 0 well outside.
+///
+/// The half-pixel feather is why a hit test asks for **more than a half**
+/// rather than "any coverage at all": half is where the geometric edge is,
+/// so the two answer the same question the drawn shape does.
+pub fn rounded_rect_coverage(pos: Vec2, top_left: Vec2, bot_right: Vec2, radius: f32) -> f32 {
+    let edge: f32 = 0.5;
+    let corner = (bot_right - top_left) / 2.0;
+    let center = top_left + corner;
+    let dist = distance_from_rect(pos, center, corner, radius);
+    1.0 - smoothstep(-edge.min(radius), edge, dist)
+}
+
+/// WGSL's `smoothstep`, which Rust has no equivalent of. Undefined in WGSL
+/// when `low == high`, which is why the caller above never passes a zero
+/// radius into the low edge without `edge` bounding it.
+fn smoothstep(low: f32, high: f32, x: f32) -> f32 {
+    let t = ((x - low) / (high - low)).clamp(0.0, 1.0);
+    t * t * (3.0 - 2.0 * t)
+}

iris/core/src/render/shader.wgsl

@@ -30,10 +30,12 @@ struct GlyphInfo {
     flags: u32,
 }
 
+/// Mirrors `Mask` in data.rs: the slot of the primitive whose coverage
+/// clips this mask's subtree, and the mask it nests inside
+/// (`4294967295u` at the top).
 struct Mask {
-    x: UiSpan,
-    y: UiSpan,
-    move_idx: u32,
+    primitive: u32,
+    parent: u32,
 }
 
 /// One widget's cumulative on-screen translation and the slot of the
@@ -53,11 +55,6 @@ struct UiScalar {
     abs: f32,
 }
 
-struct UiVec2 {
-    rel: vec2<f32>,
-    abs: vec2<f32>,
-}
-
 // The shared glyph atlas: every page is one layer. Growing it recreates this
 // texture with headroom and copies the old layers across -- see
 // GpuTextures::grow_array -- rather than the binding_array<texture_2d<f32>>
@@ -79,8 +76,15 @@ var samp: sampler;
 var<storage> masks: array<Mask>;
 @group(3) @binding(1)
 var<storage> move_offsets: array<MoveOffset>;
+// Every primitive's placement, in one arena all layers share. The vertex
+// stage reads the primitive it is drawing (its slot arrives as the only
+// vertex attribute); the fragment stage reads a *mask's* primitive, which
+// is generally a different one in a different layer. See LAYOUT.md's
+// "Masks with a shape" and `Primitives` in primitive.rs.
+@group(3) @binding(2)
+var<storage> instances: array<PrimitiveInstance>;
 
-// The bound on the parent walk, kept in step with `MOVE_CHAIN_LIMIT` in
+// The bound on the parent walk, kept in step with `PARENT_CHAIN_LIMIT` in
 // render_state.rs, which walks the identical chain on the CPU side for
 // hit-testing. Bounded so a malformed chain (a cyclic `parent`) cannot
 // hang the GPU -- not a claim about how deep a real tree gets. It was 16
@@ -90,7 +94,7 @@ var<storage> move_offsets: array<MoveOffset>;
 // Past the bound both walks simply stop summing, so the widget draws and
 // hit-tests short by whatever the outer slots held, with nothing on
 // screen to say so.
-const MOVE_CHAIN_LIMIT: u32 = 64u;
+const PARENT_CHAIN_LIMIT: u32 = 64u;
 
 /// Sums the pixel delta along the parent chain starting at `idx`, shared by
 /// the vertex stage (a primitive's own corners) and the fragment stage (its
@@ -98,7 +102,7 @@ const MOVE_CHAIN_LIMIT: u32 = 64u;
 fn resolve_move(idx: u32) -> vec2<f32> {
     var total = vec2<f32>(0.0, 0.0);
     var i = idx;
-    for (var step = 0u; step < MOVE_CHAIN_LIMIT; step++) {
+    for (var step = 0u; step < PARENT_CHAIN_LIMIT; step++) {
         let entry = move_offsets[i];
         total += entry.delta;
         if entry.parent == 4294967295u {
@@ -113,15 +117,20 @@ struct WindowUniform {
     dim: vec2<f32>,
 };
 
+/// Mirrors `PrimitiveInstance` in data.rs -- the placement and what to
+/// draw there. `x`/`y` are the `UiRegion`'s two spans.
+struct PrimitiveInstance {
+    x: UiSpan,
+    y: UiSpan,
+    binding: u32,
+    idx: u32,
+    mask_idx: u32,
+    move_idx: u32,
+}
+
+/// A layer's draw order: one slot into `instances` per instance drawn.
 struct InstanceInput {
-    @location(0) x_start: vec2<f32>,
-    @location(1) x_end: vec2<f32>,
-    @location(2) y_start: vec2<f32>,
-    @location(3) y_end: vec2<f32>,
-    @location(4) binding: u32,
-    @location(5) idx: u32,
-    @location(6) mask_idx: u32,
-    @location(7) move_idx: u32,
+    @location(0) slot: u32,
 }
 
 struct VertexOutput {
@@ -141,21 +150,38 @@ struct Region {
     bot_right: vec2<f32>,
 }
 
+/// One primitive's on-screen corners in window pixels. Written once and
+/// used by both stages: the vertex stage for the primitive it is drawing,
+/// the fragment stage for a mask's -- so the shape a mask clips to and the
+/// shape that was drawn cannot be computed two different ways.
+struct Corners {
+    top_left: vec2<f32>,
+    bot_right: vec2<f32>,
+}
+
+fn corners_of(inst: PrimitiveInstance) -> Corners {
+    let top_left_rel = vec2(inst.x.start.rel, inst.y.start.rel);
+    let top_left_abs = vec2(inst.x.start.abs, inst.y.start.abs);
+    let bot_right_rel = vec2(inst.x.end.rel, inst.y.end.rel);
+    let bot_right_abs = vec2(inst.x.end.abs, inst.y.end.abs);
+    let move_delta = resolve_move(inst.move_idx);
+    return Corners(
+        floor(top_left_rel * window.dim) + floor(top_left_abs) + move_delta,
+        floor(bot_right_rel * window.dim) + floor(bot_right_abs) + move_delta,
+    );
+}
+
 @vertex
 fn vs_main(
     @builtin(vertex_index) vi: u32,
     in: InstanceInput,
 ) -> VertexOutput {
     var out: VertexOutput;
+    let inst = instances[in.slot];
 
-    let top_left_rel = vec2(in.x_start.x, in.y_start.x);
-    let top_left_abs = vec2(in.x_start.y, in.y_start.y);
-    let bot_right_rel = vec2(in.x_end.x, in.y_end.x);
-    let bot_right_abs = vec2(in.x_end.y, in.y_end.y);
-
-    let move_delta = resolve_move(in.move_idx);
-    let top_left = floor(top_left_rel * window.dim) + floor(top_left_abs) + move_delta;
-    let bot_right = floor(bot_right_rel * window.dim) + floor(bot_right_abs) + move_delta;
+    let c = corners_of(inst);
+    let top_left = c.top_left;
+    let bot_right = c.bot_right;
     let size = bot_right - top_left;
 
     let uv = vec2<f32>(
@@ -165,11 +191,11 @@ fn vs_main(
     let pos = (top_left + uv * size) / window.dim * 2.0 - 1.0;
     out.clip_position = vec4<f32>(pos.x, -pos.y, 0.0, 1.0);
     out.uv = uv;
-    out.binding = in.binding;
-    out.idx = in.idx;
+    out.binding = inst.binding;
+    out.idx = inst.idx;
     out.top_left = top_left;
     out.bot_right = bot_right;
-    out.mask_idx = in.mask_idx;
+    out.mask_idx = inst.mask_idx;
 
     return out;
 }
@@ -196,21 +222,39 @@ fn fs_main(
             color = vec4(1.0, 0.0, 1.0, 1.0);
         }
     }
-    if in.mask_idx != 4294967295u {
-        let mask = masks[in.mask_idx];
-        let mask_delta = resolve_move(mask.move_idx);
-        let tl = UiVec2(vec2(mask.x.start.rel, mask.y.start.rel), vec2(mask.x.start.abs, mask.y.start.abs));
-        let br = UiVec2(vec2(mask.x.end.rel, mask.y.end.rel), vec2(mask.x.end.abs, mask.y.end.abs));
-
-        let top_left = floor(tl.rel * window.dim) + floor(tl.abs) + mask_delta;
-        let bot_right = floor(br.rel * window.dim) + floor(br.abs) + mask_delta;
-        if pos.x < top_left.x || pos.x > bot_right.x || pos.y < top_left.y || pos.y > bot_right.y {
-            color *= 0.0;
+    // Every mask on the chain, not just the innermost: a widget that set
+    // its own mask inside another is clipped by both, and the coverages
+    // multiply -- so a pixel inside two feathered corners is dimmed by
+    // both, which is what a compositor does (`Mask::parent` in data.rs).
+    var mask_idx = in.mask_idx;
+    for (var step = 0u; step < PARENT_CHAIN_LIMIT; step++) {
+        if mask_idx == 4294967295u {
+            break;
         }
+        let mask = masks[mask_idx];
+        color.a *= mask_coverage(pos, mask);
+        mask_idx = mask.parent;
     }
     return color;
 }
 
+/// How much of `pos` one mask lets through: the referenced primitive's
+/// own coverage at that pixel, from the same SDF the primitive is drawn
+/// with. Nothing about the shape is copied into the mask, so a rounded
+/// container's corner and its children's clipped corner are the same
+/// arithmetic.
+fn mask_coverage(pos: vec2<f32>, mask: Mask) -> f32 {
+    let inst = instances[mask.primitive];
+    if inst.binding != RECT {
+        // Unreachable: `Painter::set_mask` rejects a glyph or an image
+        // shape by name (see `Mask::primitive`). Letting the pixel
+        // through rather than reading a `rects` entry that is not there.
+        return 1.0;
+    }
+    let c = corners_of(inst);
+    return rounded_rect_coverage(pos, c.top_left, c.bot_right, rects[inst.idx].radius);
+}
+
 fn draw_texture(region: Region) -> vec4<f32> {
     return textureSample(image_texture, samp, region.uv);
 }
@@ -226,19 +270,35 @@ fn draw_glyph(region: Region, g: GlyphInfo) -> vec4<f32> {
     return color;
 }
 
+/// The anti-aliased coverage of a rounded rect at one pixel -- the one
+/// function both a drawn rect and a mask go through, and the
+/// transliteration of `iris_core::rounded_rect_coverage` on the CPU,
+/// which the hit test uses so a corner that cannot be tapped and a corner
+/// that is not drawn are the same corner.
+fn rounded_rect_coverage(
+    pos: vec2<f32>,
+    top_left: vec2<f32>,
+    bot_right: vec2<f32>,
+    radius: f32,
+) -> f32 {
+    let edge = 0.5;
+    let corner = (bot_right - top_left) / 2.0;
+    let center = top_left + corner;
+    let dist = distance_from_rect(pos, center, corner, radius);
+    return 1.0 - smoothstep(-min(edge, radius), edge, dist);
+}
+
 fn draw_rounded_rect(region: Region, rect: Rect) -> vec4<f32> {
     var color = unpack4x8unorm(rect.color);
 
     let edge = 0.5;
 
-    let size = region.bot_right - region.top_left;
-    let corner = size / 2.0;
-    let center = region.top_left + corner;
-
-    let dist = distance_from_rect(region.pos, center, corner, rect.radius);
-    color.a *= 1.0 - smoothstep(-min(edge, rect.radius), edge, dist);
+    color.a *= rounded_rect_coverage(region.pos, region.top_left, region.bot_right, rect.radius);
 
     if rect.thickness > 0.0 {
+        let size = region.bot_right - region.top_left;
+        let corner = size / 2.0;
+        let center = region.top_left + corner;
         let dist2 = distance_from_rect(region.pos, center, corner - rect.thickness, rect.inner_radius);
         color.a *= smoothstep(-min(edge, rect.inner_radius), edge, dist2);
     }

iris/core/src/ui/painter.rs

@@ -1,7 +1,10 @@
 use crate::{
-    RenderedText, Size, StrongWidget, TextAttrs, TextBuffer, TextData, TextureHandle, UiRegion,
-    UiRenderState, UiRsc, UiScalar, UiVec2, WidgetId,
-    render::{GlyphPrimitive, Mask, MaskIdx, MoveIdx, Primitive, PrimitiveHandle, PrimitiveInst},
+    Color, RenderedText, Size, StrongWidget, TextAttrs, TextBuffer, TextData, TextureHandle,
+    UiRegion, UiRenderState, UiRsc, UiScalar, UiVec2, WidgetId,
+    render::{
+        Drawn, GlyphPrimitive, Mask, MaskIdx, MoveIdx, Primitive, PrimitiveHandle, PrimitiveInst,
+        RectPrimitive,
+    },
     util::Vec2,
 };
 
@@ -26,8 +29,20 @@ pub struct Painter<'a> {
 
 impl<'a> Painter<'a> {
     fn primitive_at<P: Primitive>(&mut self, primitive: P, region: UiRegion) {
-        let h = self.state.layers.write(
+        self.write_primitive(primitive, region, Drawn::Yes);
+    }
+
+    /// The one path every primitive this widget owns goes through --
+    /// drawn or, for a mask's shape, only referenced.
+    fn write_primitive<P: Primitive>(
+        &mut self,
+        primitive: P,
+        region: UiRegion,
+        drawn: Drawn,
+    ) -> u32 {
+        let h = self.state.write_primitive(
             self.layer,
+            drawn,
             PrimitiveInst {
                 id: self.id,
                 primitive,
@@ -40,7 +55,9 @@ impl<'a> Painter<'a> {
             // TODO: I have no clue if this works at all :joy:
             self.rsc.ui_mut().masks.push_ref(self.mask);
         }
+        let slot = h.slot;
         self.primitives.push(h);
+        slot
     }
 
     /// Writes a primitive to be rendered
@@ -53,8 +70,16 @@ impl<'a> Painter<'a> {
     }
 
     /// Clip everything this widget draws, itself and its descendants, to
-    /// `region`. One per widget: a second call would need the two to be
-    /// intersected, which nothing here does.
+    /// `region`. One call per widget; a widget drawn inside another
+    /// widget's mask nests instead -- the new mask chains to the inherited
+    /// one (`Mask::parent`) and the fragment stage multiplies both
+    /// coverages, which is what lets a transcript row's code fence clip
+    /// to itself *and* to the list it scrolls inside.
+    ///
+    /// The clip is a **primitive**, not a rectangle copied into the mask:
+    /// this writes an undrawn `RectPrimitive` at `region` and points the
+    /// mask at it, so the fragment stage evaluates the same rounded-rect
+    /// coverage a drawn rect gets. See LAYOUT.md's "Masks with a shape".
     ///
     /// The slot is allocated once and **rewritten in place** on every
     /// later draw rather than pushed again, because a descendant whose own
@@ -62,24 +87,109 @@ impl<'a> Painter<'a> {
     /// so keeps pointing at whichever slot it was drawn under. See
     /// `ActiveData::own_mask` for what pushing a fresh one cost.
     pub fn set_mask(&mut self, region: UiRegion) {
-        assert!(self.mask == MaskIdx::NONE);
+        let shape = self.write_primitive(RectPrimitive::color(Color::NONE), region, Drawn::No);
+        self.set_mask_to(shape);
+    }
+
+    /// Clip everything this widget draws after this call to `shape`'s
+    /// own shape -- the first primitive `shape`'s subtree drew, which
+    /// must already have been drawn this frame
+    /// (`UiRenderState::first_primitive`). What `.masked_by()` uses to
+    /// clip a container's content to the rounded background it draws,
+    /// with no radius argument anywhere that could fall out of step with
+    /// the one being drawn.
+    pub fn set_mask_to_widget<W: ?Sized>(&mut self, shape: &StrongWidget<W>) {
+        let slot = self.state.first_primitive(shape.id()).unwrap_or_else(|| {
+            panic!(
+                "'{}' was given as a mask's shape but drew no primitive, so there is nothing to \
+                 clip to",
+                self.rsc.widgets().label(shape.id()),
+            )
+        });
+        self.set_mask_to(slot);
+    }
+
+    /// Points this widget's mask at a primitive that has already been
+    /// written -- the shared half of [`Self::set_mask`].
+    fn set_mask_to(&mut self, shape: u32) {
+        // `assert!`, not `debug_assert!`: one comparison per widget draw,
+        // and the second call silently *replacing* the first is a widget
+        // drawn unclipped -- which reaches the screen and nothing says so.
+        // Every build anybody runs here is release
+        // (docs/REVIEW-2026-09-07.md's R1).
+        assert!(
+            self.own_mask == MaskIdx::NONE || self.mask != self.own_mask,
+            "set_mask called twice while drawing one widget: the second would replace the first \
+             rather than nest inside it",
+        );
+        // A glyph would need a CPU-side alpha plane for the hit test to
+        // agree with the shader, and a standalone image a bind-group
+        // switch the fragment stage cannot make -- see `Mask::primitive`.
+        // Named here rather than left to the shader, which would read a
+        // rect that is not there and clip to nothing.
+        let binding = self.state.primitives.instance(shape).binding;
+        assert_eq!(
+            binding,
+            RectPrimitive::BINDING,
+            "a mask's shape must be a rect primitive; primitive {shape} is binding {binding}",
+        );
+        let parent = self.mask;
         let mask = Mask {
-            region,
-            move_idx: self.move_slot,
+            primitive: shape,
+            parent,
         };
-        if self.own_mask == MaskIdx::NONE {
+        let old_parent = if self.own_mask == MaskIdx::NONE {
             let slot = self.rsc.ui_mut().masks.push(mask);
             // The one ref this widget holds on its own slot, so the slot
             // outlives any single frame's primitives; released in
             // `UiRenderState::remove`'s `undraw` branch.
             self.rsc.ui_mut().masks.push_ref(slot);
             self.own_mask = slot;
+            MaskIdx::NONE
         } else {
+            let old = self.rsc.ui().masks[self.own_mask.idx()].parent;
             *self.rsc.ui_mut().masks.get_mut(self.own_mask) = mask;
+            old
+        };
+        // The chain link's own ref, taken before the old one is dropped so
+        // that re-chaining to the same slot cannot free it in between.
+        // Released here when the link changes, and in
+        // `UiRenderState::remove` when this widget's slot goes.
+        if old_parent != parent {
+            if parent != MaskIdx::NONE {
+                self.rsc.ui_mut().masks.push_ref(parent);
+            }
+            if old_parent != MaskIdx::NONE {
+                self.rsc.ui_mut().masks.remove(old_parent);
+            }
         }
         self.mask = self.own_mask;
     }
 
+    /// Ask for this widget to be drawn again on the next frame, from
+    /// inside its own `draw` -- for a layout that can only discover a
+    /// correction to itself by laying out once (`List::clamp_to_content`,
+    /// which learns how far past its content the list is from the walk it
+    /// has just done). The mark is the same one `Widgets::get_dyn_mut`
+    /// sets, so `UiRenderState::update` picks it up exactly as it does any
+    /// other dirty widget; it does **not** by itself ask the platform for
+    /// a frame, which is the caller's own `RequestRedraw` handle.
+    ///
+    /// The correction it asks for must converge, or this is a widget that
+    /// redraws forever.
+    pub fn draw_again(&mut self) {
+        self.rsc.widgets_mut().needs_redraw.insert(self.id);
+    }
+
+    /// Whether anything is clipping what this widget draws -- its own
+    /// [`Self::set_mask`], or one an ancestor set that it inherited. What
+    /// a widget whose contents may legitimately extend past its own box
+    /// (`iris::widget::List`, which draws a row straddling an edge in
+    /// full) asserts before relying on being cut off there.
+    pub fn is_masked(&self) -> bool {
+        self.mask != MaskIdx::NONE
+    }
+
     /// Draws a widget within this widget's region, returning the size it
     /// reported using.
     pub fn widget<W: ?Sized>(&mut self, id: &StrongWidget<W>) -> Size {
@@ -170,7 +280,7 @@ impl<'a> Painter<'a> {
     /// the layer's one instanced draw, so it goes through
     /// `Primitives::write_image` instead of `primitive_at`/`Primitive::vec`.
     fn write_image(&mut self, texture_idx: u32, region: UiRegion) {
-        let h = self.state.layers.write_image(
+        let h = self.state.write_image(
             self.layer,
             self.id,
             texture_idx,
@@ -216,8 +326,10 @@ impl<'a> Painter<'a> {
         // A caller re-emitting quads placed against an atlas that has since
         // been cleared draws every glyph from coordinates now holding
         // something else. Caught at the submission rather than on screen,
-        // where it reads as fragments of unrelated letters.
-        debug_assert_eq!(
+        // where it reads as fragments of unrelated letters. `assert_eq!`
+        // for R1's reason: two integers per laid-out string, not per
+        // glyph, and the failure is unreadable text on a release build.
+        assert_eq!(
             text.generation,
             self.atlas_generation(),
             "glyphs placed against atlas generation {} submitted against {}: the holder did not \

iris/core/src/ui/render_state.rs

@@ -1,12 +1,37 @@
+use std::sync::Mutex;
+use std::time::{Duration, Instant};
+
 use crate::{
     ActiveData, IdLike, MaskIdx, MoveIdx, Painter, PixelRegion, PrimitiveLayers, RegionAlign,
     StrongWidget, UiRegion, UiRsc, UiVec2, WidgetId, Widgets,
-    render::{IMAGE_BINDING, MoveOffset},
+    render::{
+        Drawn, MoveOffset, NOT_DRAWN, Primitive, PrimitiveHandle, PrimitiveInst, Primitives,
+        RectPrimitive, rounded_rect_coverage,
+    },
     util::{HashMap, HashSet, Id, Vec2},
 };
 
+/// What [`UiRenderState::update`] did on its last call -- read back by the
+/// `iris::frame` diagnostic (`iris::diagnostics::log_frame` in the `iris`
+/// crate) so a report can tell a full relayout from a frame that only
+/// redrew a handful of dirty widgets from one that drew nothing at all.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum RedrawKind {
+    /// Neither the root nor any widget changed -- `update` did nothing.
+    None,
+    /// [`UiRenderState::redraw_all`]: a new root, or a resize.
+    All,
+    /// [`UiRenderState::redraw_updates`]: only the widgets `needs_redraw`
+    /// named.
+    Updates,
+}
+
 pub struct UiRenderState {
     pub active: HashMap<WidgetId, ActiveData>,
+    /// Every primitive in the tree, in one arena -- see [`Primitives`] for
+    /// why it is not per layer.
+    pub primitives: Primitives,
+    /// What each layer draws, in order: slots into `primitives`.
     pub layers: PrimitiveLayers,
     pub(super) output_size: Vec2,
     /// Physical pixels per `dp` -- see `Len::dp`'s field doc. `1.0` (an
@@ -32,21 +57,6 @@ pub struct UiRenderState {
     /// ever drawn and was never emptied.
     draw_started: HashSet<WidgetId>,
 
-    /// The widget currently holding exclusive pointer input, if any --
-    /// `iris::sense::SensorUi::run_sensors` reads and clears this every
-    /// call. Interior mutability (a `Mutex`, not a bare `Cell`, since a
-    /// `CursorData` reaching this through an async `task_on` handler needs
-    /// `Send`/`Sync`) because `run_sensors` takes `&self` (widgets are
-    /// dispatched to, not owned, at that layer) and this render state is
-    /// the one structure both backends (winit, android-view) already hold
-    /// across frames, the same way `old_root`/`resized` are -- see
-    /// `iris::sense`'s pointer-capture doc for why a drag needs this: once
-    /// a gesture has committed to panning or selecting, every later sample
-    /// of it must reach the same widget even if the finger has moved off
-    /// whatever hit region first noticed the press. Never held across an
-    /// await or another lock -- every access here is a single get/set.
-    captured: std::sync::Mutex<Option<WidgetId>>,
-
     /// `Widget::draw` calls and `Primitives::region_mut` rewrites since the
     /// last `take_counters`. LAYOUT.md section 8's pass conditions are
     /// stated in terms of these two: an unchanged frame must cost 0 of
@@ -58,6 +68,30 @@ pub struct UiRenderState {
     /// Text layouts actually computed -- bumped by `Painter::render_text`,
     /// which `TextView::render` only reaches on a cache miss.
     pub(super) shape_count: u64,
+
+    /// `Instant::now()` at construction -- the zero every `iris::frame` line
+    /// dates itself from, so a report's `now=` is comparable to a harness's
+    /// own `t_ms` (`Harness::new` builds its `base` the same way, in the
+    /// same constructor call) without either side needing the wall clock.
+    epoch: Instant,
+    /// How many times [`Self::update`] has run -- the `iris::frame` line's
+    /// frame number. Counts every call, including one that found nothing to
+    /// redraw, so a gap in the sequence in a report is a frame this state
+    /// was never asked to run at all (a stalled event loop), not one that
+    /// ran and did nothing.
+    frame_no: u64,
+    /// How long the redraw phase of the last [`Self::update`] took --
+    /// [`Self::redraw_all`] or [`Self::redraw_updates`], whichever ran, or
+    /// zero if neither did. Read back by `iris::diagnostics::log_frame`.
+    last_layout: Duration,
+    last_redraw_kind: RedrawKind,
+    /// When the sensor dispatch (`SensorUi::run_sensors`, in the `iris`
+    /// crate) last saw an input sample, dated by the sample's own clock
+    /// (`CursorState::time`) rather than when the dispatch ran -- same
+    /// reasoning as that field's own doc. A `Mutex` because `run_sensors`
+    /// takes `&self` and this is the one render state both backends
+    /// already share across frames.
+    last_input_at: Mutex<Option<Instant>>,
 }
 
 /// The bound on the parent walk -- see `resolve_move` in shader.wgsl,
@@ -70,23 +104,33 @@ pub struct UiRenderState {
 /// prints the chain). A chain past the bound is not reported anywhere at
 /// run time; both walks just stop summing, so the widget is drawn and hit
 /// tested short by whatever the outer slots held.
-pub const MOVE_CHAIN_LIMIT: usize = 64;
+///
+/// Named for the walk rather than for one of its two subjects: it bounds
+/// the move-offset chain *and* the mask chain (`Mask::parent`, walked in
+/// the fragment stage), and `MOVE_CHAIN_LIMIT` said only the first
+/// (docs/REVIEW-2026-09-07.md).
+pub const PARENT_CHAIN_LIMIT: usize = 64;
 
 impl UiRenderState {
     pub fn new() -> Self {
         Self {
             active: Default::default(),
+            primitives: Default::default(),
             layers: Default::default(),
             output_size: Vec2::ZERO,
             density: 1.0,
             old_root: None,
             resized: false,
             draw_started: Default::default(),
-            captured: Default::default(),
             draw_count: 0,
             region_mut_count: 0,
             mov_count: 0,
             shape_count: 0,
+            epoch: Instant::now(),
+            frame_no: 0,
+            last_layout: Duration::ZERO,
+            last_redraw_kind: RedrawKind::None,
+            last_input_at: Mutex::new(None),
         }
     }
 
@@ -108,6 +152,75 @@ impl UiRenderState {
         )
     }
 
+    /// Writes a primitive into the arena and, unless it is
+    /// [`Drawn::No`], into `layer`'s draw order.
+    pub(super) fn write_primitive<P: Primitive>(
+        &mut self,
+        layer: usize,
+        drawn: Drawn,
+        inst: PrimitiveInst<P>,
+    ) -> PrimitiveHandle {
+        let (slot, data_idx) = self.primitives.alloc(inst);
+        let pos = match drawn {
+            Drawn::Yes => self.layers[layer].push(slot, false),
+            Drawn::No => NOT_DRAWN,
+        };
+        PrimitiveHandle {
+            layer,
+            pos,
+            slot,
+            data_idx,
+            binding: P::BINDING,
+        }
+    }
+
+    /// A standalone image, which draws with its own bind group rather
+    /// than sharing the layer's one instanced draw.
+    pub(super) fn write_image(
+        &mut self,
+        layer: usize,
+        id: WidgetId,
+        texture_idx: u32,
+        region: UiRegion,
+        mask_idx: MaskIdx,
+        move_idx: MoveIdx,
+    ) -> PrimitiveHandle {
+        let slot = self
+            .primitives
+            .alloc_image(id, texture_idx, region, mask_idx, move_idx);
+        let pos = self.layers[layer].push(slot, true);
+        PrimitiveHandle {
+            layer,
+            pos,
+            slot,
+            data_idx: 0,
+            binding: crate::render::IMAGE_BINDING,
+        }
+    }
+
+    /// Compacts every layer's draw order around the primitives freed
+    /// this frame, corrects the handles that moved, and only then hands
+    /// the arena slots back for reuse -- that order is the whole reason
+    /// `Primitives::freed` exists. Once per frame, at the end of
+    /// [`Self::update`], so the harness (which has no renderer) applies
+    /// it exactly as a real backend does.
+    fn apply_free(&mut self) {
+        for (layer, order) in self.layers.iter_mut() {
+            for change in order.apply_free() {
+                let owner = self.primitives.owner(change.slot);
+                if let Some(active) = self.active.get_mut(&owner) {
+                    for h in &mut active.primitives {
+                        if h.layer == layer && h.slot == change.slot {
+                            h.pos = change.pos;
+                            break;
+                        }
+                    }
+                }
+            }
+        }
+        self.primitives.release_freed();
+    }
+
     pub fn resize(&mut self, size: impl Into<Vec2>) {
         self.output_size = size.into();
         self.resized = true;
@@ -119,7 +232,16 @@ impl UiRenderState {
     /// different triggers (a surface resize on every rotation or keyboard
     /// open; a density change only if the app follows the display to a
     /// different screen, which Android surfaces separately).
+    ///
+    /// Marks the tree for a full redraw when the value actually changes:
+    /// every `Len::dp` already resolved and every glyph already shaped
+    /// (`Text::shape` keys its cache on `(attrs, width, density)`) belongs
+    /// to the old one, and nothing else would ask for them again
+    /// (docs/REVIEW-2026-09-07.md's R5).
     pub fn set_density(&mut self, density: f32) {
+        if density != self.density {
+            self.resized = true;
+        }
         self.density = density;
     }
 
@@ -150,17 +272,94 @@ impl UiRenderState {
             "a previous frame left {} widget(s) marked as mid-draw",
             self.draw_started.len(),
         );
-        if self.needs_redraw_all(root) {
+        // Timed unconditionally -- an `Instant::now()` pair is cheap enough
+        // not to move the `--phone` bench's frame time (checked when this
+        // was added), and gating it behind the trace toggle would leave
+        // `iris::frame` with nothing to report the one frame somebody just
+        // turned tracing on to look at.
+        let layout_start = Instant::now();
+        let kind = if self.needs_redraw_all(root) {
             self.redraw_all(root, rsc);
             self.old_root = root.map(|r| r.id());
             self.resized = false;
+            RedrawKind::All
         } else if rsc.widgets().has_updates() {
             self.redraw_updates(rsc);
-        }
+            RedrawKind::Updates
+        } else {
+            RedrawKind::None
+        };
+        self.last_layout = layout_start.elapsed();
+        self.last_redraw_kind = kind;
+        self.frame_no += 1;
+        // After the redraw and before anything reads the frame: every
+        // slot freed above is still named by its layer's draw order until
+        // this runs.
+        self.apply_free();
         #[cfg(debug_assertions)]
         debug_assert!(self.primitive_counts_agree(), "{}", self.orphan_report(rsc),);
     }
 
+    /// `Instant::now()` at construction -- see the field's own doc.
+    pub fn epoch(&self) -> Instant {
+        self.epoch
+    }
+
+    /// How many times [`Self::update`] has run, counting from 1.
+    pub fn frame_number(&self) -> u64 {
+        self.frame_no
+    }
+
+    /// How long the last [`Self::update`]'s redraw phase took.
+    pub fn last_layout_duration(&self) -> Duration {
+        self.last_layout
+    }
+
+    /// What the last [`Self::update`] did -- see [`RedrawKind`].
+    pub fn last_redraw_kind(&self) -> RedrawKind {
+        self.last_redraw_kind
+    }
+
+    /// Records that a real input sample was just dispatched, dated by the
+    /// sample's own clock -- called once per sensor pass, so `iris::frame`'s
+    /// `since_input` can answer "how stale was the input
+    /// this frame drew" instead of a caller guessing from the frame
+    /// interval. `&self` because `run_sensors` only ever has that -- see
+    /// `last_input_at`'s field doc.
+    pub fn note_input(&self, at: Instant) {
+        if let Ok(mut guard) = self.last_input_at.lock() {
+            *guard = Some(at);
+        }
+    }
+
+    /// `now - ` the last input sample's own timestamp, or `None` if no
+    /// input has ever reached this render state (a cold start, or a screen
+    /// that only ever animates on its own). Saturates to zero rather than
+    /// panicking if `now` is earlier than the input sample somehow was --
+    /// a diagnostic reading wrong is not worth a crash over.
+    pub fn time_since_input(&self, now: Instant) -> Option<Duration> {
+        let at = *self.last_input_at.lock().ok()?;
+        at.map(|at| now.saturating_duration_since(at))
+    }
+
+    /// Primitive instances every currently-active widget owns, summed --
+    /// what `iris::frame`'s `primitives=` reports. Not a per-frame delta:
+    /// `redraw_updates` only rewrites what changed, so this is "how much is
+    /// on screen", which is what a report reads as "did this frame have
+    /// more to draw than the last one", not "how much work did this frame
+    /// do" (`take_counters` answers that).
+    ///
+    /// A mask's shape does not count: it is a [`Drawn::No`] primitive
+    /// that is never rasterized, so including it would put one extra on
+    /// the line for every masked widget and make a number Iris reads off
+    /// a phone report disagree with what is drawn.
+    pub fn active_primitive_count(&self) -> usize {
+        self.active
+            .values()
+            .map(|a| a.primitives.iter().filter(|h| h.pos != NOT_DRAWN).count())
+            .sum()
+    }
+
     fn redraw_all(&mut self, root: Option<&StrongWidget>, rsc: &mut dyn UiRsc) {
         self.clear(rsc);
         // free all resources & cache
@@ -252,7 +451,7 @@ impl UiRenderState {
                 // instead of redrawing. See LAYOUT.md section 3.
                 let from = active.region;
                 for h in &active.primitives {
-                    let r = self.layers[h.layer].region_mut(h);
+                    let r = self.primitives.region_mut(h);
                     *r = r.outside(&from).within(&region);
                     self.region_mut_count += 1;
                 }
@@ -324,10 +523,9 @@ impl UiRenderState {
         // own new one -- and `ActiveData::mask`'s only consumer is
         // `redraw`, which feeds it back in as the *inherited* mask. Storing
         // the set one instead handed a `Masked` its own mask on every
-        // targeted redraw, tripping `set_mask`'s nested-mask assert:
-        // `assertion failed: self.mask == MaskIdx::NONE`, an abort the
-        // first time the composer's scroll area was redrawn on the
-        // emulator.
+        // targeted redraw -- an abort the first time the composer's scroll
+        // area was redrawn on the emulator, and now (masks nest) a mask
+        // whose parent is itself, which `set_mask`'s own assert names.
         let inherited_mask = mask;
         let mut painter = Painter {
             state: self,
@@ -490,21 +688,26 @@ impl UiRenderState {
         let mut active = self.active.remove(&id);
         if let Some(active) = &mut active {
             for h in &active.primitives {
-                let mask = self.layers.free(h);
+                let mask = self.primitives.free(h);
+                if h.pos != NOT_DRAWN {
+                    self.layers[h.layer].free(h.pos, h.is_image());
+                }
                 if mask != MaskIdx::NONE {
                     rsc.ui_mut().masks.remove(mask);
                 }
             }
+            Self::remask_shape_users(&self.active, id, active.own_mask, &active.primitives, rsc);
             active.textures.clear();
             rsc.ui_mut().textures.free();
             if undraw {
                 // A captured widget that goes away mid-gesture (List's
                 // virtualisation retiring a row, a rebuild) must not leave
-                // the pointer permanently captured by an id nothing will
-                // ever draw again -- `captured`'s own path out.
-                if *self.captured.lock().unwrap() == Some(id) {
-                    *self.captured.lock().unwrap() = None;
-                }
+                // the pointer captured by an id nothing will ever draw
+                // again. That path out is the sensor pass's, not this
+                // one's: `iris::sense::SensorUi::run_sensors` releases a
+                // capture whose widget no longer resolves to a region,
+                // which covers this case and every other way an id can
+                // stop being drawn.
                 // Permanent removal: retire this widget's own move slot
                 // (the self-ownership ref taken when it was allocated) and
                 // the up-link ref it held on its parent's slot -- read from
@@ -514,8 +717,15 @@ impl UiRenderState {
                 // section 2's lifecycle note).
                 if active.own_mask != MaskIdx::NONE {
                     // The self-ownership ref `Painter::set_mask` took when
-                    // it allocated this widget's own mask slot.
+                    // it allocated this widget's own mask slot, and the
+                    // chain link's ref on the mask this one nests inside
+                    // -- read from the arena entry, for the same reason
+                    // the move slot's parent is.
+                    let outer = rsc.ui().masks[active.own_mask.idx()].parent;
                     rsc.ui_mut().masks.remove(active.own_mask);
+                    if outer != MaskIdx::NONE {
+                        rsc.ui_mut().masks.remove(outer);
+                    }
                 }
                 let parent_slot = rsc.ui_mut().move_offsets[active.move_slot.idx()].parent;
                 rsc.ui_mut().move_offsets.remove(active.move_slot);
@@ -528,6 +738,54 @@ impl UiRenderState {
         active
     }
 
+    /// A mask whose shape primitive was just freed clips to a slot that
+    /// now holds something else, so the widget that owns it is marked for
+    /// redraw -- its own `set_mask` is the only thing that resolves the
+    /// slot, and it is the same mechanism a dirty widget already goes
+    /// through.
+    ///
+    /// `own` is the mask belonging to the widget being removed and is
+    /// skipped: this runs in the middle of that widget's own redraw,
+    /// which sets its mask again on the way out, and a mark left on
+    /// itself would redraw it every frame from then on. Skipping it is
+    /// also what keeps the O(active) scan off the ordinary path -- a
+    /// plain `.masked()` frees exactly its own shape, so `stale` is empty
+    /// and this returns before touching `active`.
+    ///
+    /// Both `Vec`s start empty and stay unallocated in that case, and
+    /// membership is a linear scan of two lists that are a handful long
+    /// (a widget's own primitives, and the live masks): this runs once
+    /// per widget removed, which is once per dirty widget per frame, and
+    /// a set built there would be an allocation on the phone's frame
+    /// path in exchange for nothing at these sizes.
+    fn remask_shape_users(
+        active: &HashMap<WidgetId, ActiveData>,
+        id: WidgetId,
+        own: MaskIdx,
+        freed: &[PrimitiveHandle],
+        rsc: &mut dyn UiRsc,
+    ) {
+        let mut stale: Vec<MaskIdx> = Vec::new();
+        for (i, mask) in rsc.ui().masks.iter().enumerate() {
+            let idx = Id::preset(i as u32);
+            if idx != own && freed.iter().any(|h| h.slot == mask.primitive) {
+                stale.push(idx);
+            }
+        }
+        if stale.is_empty() {
+            return;
+        }
+        let mut owners: Vec<WidgetId> = Vec::new();
+        for (widget, data) in active {
+            if *widget != id && stale.contains(&data.own_mask) {
+                owners.push(*widget);
+            }
+        }
+        for owner in owners {
+            rsc.widgets_mut().needs_redraw.insert(owner);
+        }
+    }
+
     fn remove_rec(&mut self, id: WidgetId, rsc: &mut dyn UiRsc) -> Option<ActiveData> {
         let inst = self.remove(id, true, rsc);
         if let Some(inst) = &inst {
@@ -543,6 +801,7 @@ impl UiRenderState {
             rsc.on_undraw(&active);
         }
         self.layers.clear();
+        self.primitives.clear();
         rsc.widgets_mut().needs_redraw.clear();
         rsc.free();
     }
@@ -585,8 +844,9 @@ impl UiRenderState {
     /// Primitive instances still bound for the GPU whose owner is no
     /// longer in `active`, or whose owner's `ActiveData` no longer names
     /// them: a copy nothing can move, clip, resize or free, redrawn every
-    /// frame at whatever position it last had. `(layer, inst_idx, owner)`
-    /// each.
+    /// frame at whatever position it last had. `(slot, owner)` each --
+    /// the arena knows which primitive, not which layer's draw order still
+    /// names it.
     ///
     /// Asserted empty at the end of every [`Self::update`], because this
     /// is exactly the shape of the duplicated transcript row on Iris's
@@ -595,20 +855,15 @@ impl UiRenderState {
     /// much alive -- it is the *earlier* set of primitives that got
     /// stranded when the widget was drawn a second time without the first
     /// draw being freed. O(primitives), debug builds only.
-    pub fn orphaned_primitives(&self) -> Vec<(usize, usize, WidgetId)> {
+    pub fn orphaned_primitives(&self) -> Vec<(u32, WidgetId)> {
         let mut orphans = Vec::new();
-        for (layer, primitives) in self.layers.iter() {
-            for (inst_idx, owner, is_image) in primitives.live_instances() {
-                let owned = self.active.get(&owner).is_some_and(|a| {
-                    a.primitives.iter().any(|h| {
-                        h.layer == layer
-                            && h.inst_idx == inst_idx
-                            && (h.binding == IMAGE_BINDING) == is_image
-                    })
-                });
-                if !owned {
-                    orphans.push((layer, inst_idx, owner));
-                }
+        for (slot, owner, _) in self.primitives.live_instances() {
+            let owned = self
+                .active
+                .get(&owner)
+                .is_some_and(|a| a.primitives.iter().any(|h| h.slot == slot));
+            if !owned {
+                orphans.push((slot, owner));
             }
         }
         orphans
@@ -622,7 +877,7 @@ impl UiRenderState {
     /// transcript is tens of thousands and made a debug build on a phone
     /// too slow to finish a benchmark run.
     fn primitive_counts_agree(&self) -> bool {
-        let live: usize = self.layers.iter().map(|(_, p)| p.live_count()).sum();
+        let live: usize = self.primitives.live_count();
         let owned: usize = self.active.values().map(|a| a.primitives.len()).sum();
         live == owned
     }
@@ -636,10 +891,10 @@ impl UiRenderState {
         let mut lines: Vec<String> = orphans
             .iter()
             .take(8)
-            .map(|(layer, idx, owner)| {
+            .map(|(slot, owner)| {
                 let alive = self.active.contains_key(owner);
                 format!(
-                    "  layer {layer} instance {idx}: owner '{}' ({owner:?}),                      owner still active: {alive}",
+                    "  instance {slot}: owner '{}' ({owner:?}), owner still active: {alive}",
                     rsc.widgets().label(*owner),
                 )
             })
@@ -655,27 +910,6 @@ impl UiRenderState {
         )
     }
 
-    /// Give `id` exclusive pointer input from the next `run_sensors` call
-    /// on -- see `captured`'s field doc. Overwrites any previous capture
-    /// (a gesture that starts a new one has already decided the old one
-    /// is over).
-    pub fn capture_pointer(&self, id: WidgetId) {
-        *self.captured.lock().unwrap() = Some(id);
-    }
-
-    /// Release exclusive pointer input, if any is held -- called once
-    /// `run_sensors` has delivered the terminal `Drop` to the capturing
-    /// widget, or by that widget itself if it decides the gesture is over
-    /// some other way.
-    pub fn release_pointer(&self) {
-        *self.captured.lock().unwrap() = None;
-    }
-
-    /// The widget currently holding exclusive pointer input, if any.
-    pub fn captured_pointer(&self) -> Option<WidgetId> {
-        *self.captured.lock().unwrap()
-    }
-
     pub fn debug(&self, widgets: &Widgets, label: &str) -> impl Iterator<Item = &ActiveData> {
         self.active.iter().filter_map(move |(&id, inst)| {
             let l = widgets.label(id);
@@ -684,12 +918,12 @@ impl UiRenderState {
     }
 
     pub fn debug_layers(&self) {
-        for ((idx, depth), primitives) in self.layers.iter_depth() {
+        for ((idx, depth), order) in self.layers.iter_depth() {
             let indent = " ".repeat(depth * 2);
-            let len = primitives.instances().len();
+            let len = order.order().len();
             print!("{indent}{idx}: {len} primitives");
             if len >= 1 {
-                print!(" ({})", primitives.instances()[0].binding);
+                print!(" ({})", self.primitives.instance(order.order()[0]).binding);
             }
             println!();
         }
@@ -713,13 +947,13 @@ impl UiRenderState {
 
     /// The plain-Rust twin of `resolve_move` in shader.wgsl: sums the
     /// pixel delta along the parent chain starting at `slot`. Both walks
-    /// share `MOVE_CHAIN_LIMIT` as their bound so the two cannot disagree
+    /// share `PARENT_CHAIN_LIMIT` as their bound so the two cannot disagree
     /// about where the chain ends.
     fn resolve_move_chain(&self, slot: MoveIdx, rsc: &dyn UiRsc) -> Vec2 {
         let offsets = &rsc.ui().move_offsets;
         let mut delta = Vec2::ZERO;
         let mut at = slot;
-        for i in 0..MOVE_CHAIN_LIMIT {
+        for i in 0..PARENT_CHAIN_LIMIT {
             let entry = &offsets[at.idx()];
             delta.x += entry.delta[0];
             delta.y += entry.delta[1];
@@ -732,8 +966,9 @@ impl UiRenderState {
             // follow are different faults with different fixes, and the
             // slot numbers are the only thing that tells them apart.
             debug_assert!(
-                i + 1 < MOVE_CHAIN_LIMIT,
-                "move offset chain exceeded MOVE_CHAIN_LIMIT ({MOVE_CHAIN_LIMIT}): {chain} -- a \
+                i + 1 < PARENT_CHAIN_LIMIT,
+                "move offset chain exceeded PARENT_CHAIN_LIMIT ({PARENT_CHAIN_LIMIT}): {chain} \
+                 -- a \
                  repeated slot means a `parent` link is cyclic, all-distinct slots mean the tree \
                  nests deeper than shader.wgsl's own walk of the same bound",
                 chain = Self::move_chain_debug(slot, offsets)
@@ -743,13 +978,13 @@ impl UiRenderState {
     }
 
     /// The parent chain from `slot`, as `slot(dx, dy) -> ...`, walked twice
-    /// `MOVE_CHAIN_LIMIT` so a cycle shows up as a repeated slot number
+    /// `PARENT_CHAIN_LIMIT` so a cycle shows up as a repeated slot number
     /// rather than as a chain that merely stops. Only ever called from the
     /// failed assertion above.
     fn move_chain_debug(slot: MoveIdx, offsets: &[MoveOffset]) -> String {
         let mut parts = Vec::new();
         let mut at = slot;
-        for _ in 0..MOVE_CHAIN_LIMIT * 2 {
+        for _ in 0..PARENT_CHAIN_LIMIT * 2 {
             let entry = &offsets[at.idx()];
             parts.push(format!(
                 "{}({}, {})",
@@ -765,6 +1000,92 @@ impl UiRenderState {
         parts.join(" -> ")
     }
 
+    /// One primitive's corners in window pixels -- the transliteration of
+    /// `shader.wgsl`'s `corners_of`, `floor` for `floor`. The rounding is
+    /// the whole reason this is not `region.to_px()`: the shader floors
+    /// each half separately before adding the move delta, and a hit test
+    /// that skipped it would disagree with the pixels by up to one along
+    /// each edge -- invisible in every test written against a whole-pixel
+    /// layout and wrong on the phone, whose 2.55 density makes nothing
+    /// land on a whole pixel.
+    pub fn primitive_corners(&self, slot: u32, rsc: &dyn UiRsc) -> PixelRegion {
+        let inst = self.primitives.instance(slot);
+        let delta = self.resolve_move_chain(inst.move_idx, rsc);
+        let size = self.output_size;
+        let corner = |c: UiVec2| (c.get_rel() * size).floor() + c.get_abs().floor() + delta;
+        PixelRegion {
+            top_left: corner(inst.region.top_left()),
+            bot_right: corner(inst.region.bot_right()),
+        }
+    }
+
+    /// Where a mask's clip actually is on screen: the box of the
+    /// primitive it references. Its *shape* within that box is
+    /// [`Self::mask_coverage`]'s -- this is the bounding box, which is
+    /// what a test asking "is the clip over the right part of the screen"
+    /// wants and all a square-cornered mask has ever had.
+    pub fn mask_region(&self, mask: MaskIdx, rsc: &dyn UiRsc) -> PixelRegion {
+        self.primitive_corners(rsc.ui().masks[mask.idx()].primitive, rsc)
+    }
+
+    /// How much of the pixel at `pos` (window pixels) survives `mask` and
+    /// every mask it nests inside: the referenced primitives' own
+    /// coverage, multiplied along the chain. The CPU half of
+    /// `shader.wgsl`'s `fs_main` mask loop -- same order, same bound, same
+    /// `rounded_rect_coverage` -- so a corner that cannot be tapped and a
+    /// corner that is not drawn are the same corner (LAYOUT.md's "Masks
+    /// with a shape", point 4).
+    ///
+    /// A mask whose shape is not a rect covers everything, exactly as the
+    /// shader's own `mask_coverage` does: `Painter::set_mask_to` rejects
+    /// those by name, so this is the unreachable half of the same
+    /// agreement rather than a second policy.
+    pub fn mask_coverage(&self, mask: MaskIdx, pos: Vec2, rsc: &dyn UiRsc) -> f32 {
+        let mut coverage = 1.0;
+        let mut at = mask;
+        for i in 0..PARENT_CHAIN_LIMIT {
+            if at == MaskIdx::NONE {
+                return coverage;
+            }
+            let m = rsc.ui().masks[at.idx()];
+            if let Some(rect) = self.primitives.primitive_data::<RectPrimitive>(m.primitive) {
+                let c = self.primitive_corners(m.primitive, rsc);
+                coverage *= rounded_rect_coverage(pos, c.top_left, c.bot_right, rect.radius);
+            }
+            at = m.parent;
+            debug_assert!(
+                i + 1 < PARENT_CHAIN_LIMIT || at == MaskIdx::NONE,
+                "mask chain exceeded PARENT_CHAIN_LIMIT ({PARENT_CHAIN_LIMIT}) from {mask:?} -- a \
+                 repeated slot means a `parent` link is cyclic, all-distinct slots mean the tree \
+                 nests deeper than shader.wgsl's own walk of the same bound",
+            );
+        }
+        coverage
+    }
+
+    /// Whether `pos` is inside `mask` at all -- more than half covered,
+    /// which is where the drawn edge is (`rounded_rect_coverage`'s doc).
+    /// What a hit test asks.
+    pub fn mask_admits(&self, mask: MaskIdx, pos: Vec2, rsc: &dyn UiRsc) -> bool {
+        self.mask_coverage(mask, pos, rsc) > 0.5
+    }
+
+    /// The first primitive `id`'s subtree wrote this frame, depth first
+    /// in draw order -- what a mask pointed at a widget clips to
+    /// (`Painter::set_mask_to_widget`). A widget that draws more than one
+    /// (a bordered rect is one primitive; a card with a stripe is two)
+    /// gives its first; a widget that wants another names it.
+    pub fn first_primitive(&self, id: WidgetId) -> Option<u32> {
+        let active = self.active.get(&id)?;
+        if let Some(h) = active.primitives.first() {
+            return Some(h.slot);
+        }
+        active
+            .children
+            .iter()
+            .find_map(|child| self.first_primitive(*child))
+    }
+
     pub fn window_region(&self, id: &impl IdLike, rsc: &dyn UiRsc) -> Option<PixelRegion> {
         let region = self.resolved_region(id, rsc)?;
         Some(region.to_px(self.output_size))

iris/desktop-app/src/config.rs

@@ -1,19 +1,13 @@
-//! Where the desktop app keeps the enrollment it should not have to be
-//! told about a second time: `client_core::config::EnrolledServer`,
-//! persisted at `$XDG_CONFIG_HOME/ai-app-desktop/enrollment.json`,
-//! owner-only (0600) -- MACHINE.md's rule for anything holding a bearer
-//! token, and the reason `client_core::config`'s own doc comment leaves
-//! persistence and file mode to the caller.
+//! Where the desktop app keeps its enrollment: `client_core::config`'s
+//! [`EnrollmentStore`] pointed at `$XDG_CONFIG_HOME/ai-app-desktop`.
 //!
-//! JSON rather than the project's usual RON: `wg-app-link`'s RON house
-//! rules (`format`) are for configs a person hand-edits, and this file
-//! never is one -- only this program ever writes or reads it, and
-//! `serde_json` is already in the dependency graph through `client-core`,
-//! so nothing new is added to reach for it.
+//! Only the directory is this app's -- the file's name, its JSON, and its
+//! owner-only mode (MACHINE.md's rule for anything holding a bearer token)
+//! are the store's, shared with the Android client so the two cannot come
+//! to disagree about them.
 
-use client_core::config::EnrolledServer;
-use std::io;
-use std::path::{Path, PathBuf};
+use client_core::config::EnrollmentStore;
+use std::path::PathBuf;
 
 /// `$XDG_CONFIG_HOME/ai-app-desktop`, falling back to `~/.config` the way
 /// the XDG basedir spec says to when the variable is unset -- the same
@@ -32,103 +26,6 @@ pub fn config_dir() -> PathBuf {
     base.join("ai-app-desktop")
 }
 
-fn enrollment_file(dir: &Path) -> PathBuf {
-    dir.join("enrollment.json")
-}
-
-/// Persists `server` under `dir` (`config_dir()` for real use; a tempdir in
-/// the tests below), creating it if needed, and sets the file owner-only --
-/// it carries a bearer token, the same reason `server/`'s own token store
-/// is 0600.
-pub fn save_enrollment_in(dir: &Path, server: &EnrolledServer) -> io::Result<()> {
-    std::fs::create_dir_all(dir)?;
-    let path = enrollment_file(dir);
-    let json = serde_json::to_vec_pretty(server)
-        .expect("EnrolledServer holds nothing that fails to serialise");
-    std::fs::write(&path, json)?;
-    #[cfg(unix)]
-    {
-        use std::os::unix::fs::PermissionsExt;
-        std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600))?;
-    }
-    Ok(())
-}
-
-/// `Ok(None)` when nothing has been enrolled yet, rather than an error --
-/// "not enrolled" is an ordinary first-run state, not a failure (UI_RULES'
-/// "a deliberate choice is not a problem to report" applies just as well
-/// to a file that simply hasn't been written yet).
-pub fn load_enrollment_in(dir: &Path) -> io::Result<Option<EnrolledServer>> {
-    let path = enrollment_file(dir);
-    match std::fs::read(&path) {
-        Ok(bytes) => {
-            let server = serde_json::from_slice(&bytes).map_err(|e| {
-                io::Error::new(
-                    io::ErrorKind::InvalidData,
-                    format!("{} is not a valid enrollment ({e})", path.display()),
-                )
-            })?;
-            Ok(Some(server))
-        }
-        Err(e) if e.kind() == io::ErrorKind::NotFound => Ok(None),
-        Err(e) => Err(e),
-    }
-}
-
-pub fn save_enrollment(server: &EnrolledServer) -> io::Result<()> {
-    save_enrollment_in(&config_dir(), server)
-}
-
-pub fn load_enrollment() -> io::Result<Option<EnrolledServer>> {
-    load_enrollment_in(&config_dir())
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn a_saved_enrollment_reads_back_the_same() {
-        let dir = tempfile::tempdir().unwrap();
-        let server = EnrolledServer {
-            host: "127.0.0.1".to_string(),
-            port: 8547,
-            token: "tok".to_string(),
-        };
-        save_enrollment_in(dir.path(), &server).unwrap();
-        let read_back = load_enrollment_in(dir.path()).unwrap();
-        assert_eq!(read_back, Some(server));
-    }
-
-    #[test]
-    fn nothing_saved_yet_is_none_not_an_error() {
-        let dir = tempfile::tempdir().unwrap();
-        assert_eq!(load_enrollment_in(dir.path()).unwrap(), None);
-    }
-
-    #[test]
-    #[cfg(unix)]
-    fn the_saved_file_is_owner_only() {
-        use std::os::unix::fs::PermissionsExt;
-        let dir = tempfile::tempdir().unwrap();
-        let server = EnrolledServer {
-            host: "h".to_string(),
-            port: 1,
-            token: "t".to_string(),
-        };
-        save_enrollment_in(dir.path(), &server).unwrap();
-        let mode = std::fs::metadata(enrollment_file(dir.path()))
-            .unwrap()
-            .permissions()
-            .mode();
-        assert_eq!(mode & 0o777, 0o600);
-    }
-
-    #[test]
-    fn a_corrupt_file_is_named_in_the_error() {
-        let dir = tempfile::tempdir().unwrap();
-        std::fs::write(enrollment_file(dir.path()), b"not json").unwrap();
-        let err = load_enrollment_in(dir.path()).unwrap_err();
-        assert!(err.to_string().contains("enrollment.json"));
-    }
+pub fn store() -> EnrollmentStore {
+    EnrollmentStore::new(config_dir())
 }

iris/desktop-app/src/main.rs

@@ -5,18 +5,20 @@
 //!
 //! Usage:
 //!
-//!     desktop-app --ca /path/to/ca.pem --link 'aiapp://enroll?host=H&port=P&token=T'
-//!     desktop-app --ca /path/to/ca.pem   # after the first run above
+//!     desktop-app --link 'aiapp://enroll?host=H&port=P&token=T&ca=B'
+//!     desktop-app                        # after the first run above
+//!     desktop-app --ca /path/to/ca.pem   # a link that carries no CA
 //!
 //! `--link` is the same text `app/ui-sandbox.sh`'s banner prints and a
 //! phone would scan as a QR (DECISIONS.md, 2026-09-05) -- pasted rather
 //! than scanned, since a desktop has no camera to assume. It is parsed and
-//! saved to `config::save_enrollment` once; later runs read it back and
-//! `--link` is only needed again to enrol against a different server. The
-//! CA is never persisted -- it is a public certificate whose path a
-//! caller is expected to already know (`AGENTS.md`'s "prefer exercising
-//! the server directly": the same `certs/ca.pem` a `curl --cacert` call
-//! uses).
+//! saved once; later runs read it back and `--link` is only needed again
+//! to enrol against a different server.
+//!
+//! The CA comes with the link (`wg_app_link::enroll::ca_param`, which
+//! `ai-server` now always includes) and is saved with it. `--ca` is the
+//! override for a link that carries none, and names the same
+//! `certs/ca.pem` a `curl --cacert` call uses.
 
 mod app;
 mod config;
@@ -24,7 +26,7 @@ mod config;
 use client_core::config::EnrolledServer;
 
 struct Args {
-    ca_path: std::path::PathBuf,
+    ca_path: Option<std::path::PathBuf>,
     link: Option<String>,
 }
 
@@ -43,13 +45,7 @@ fn parse_args() -> Result<Args, String> {
             other => return Err(format!("unrecognised argument '{other}'")),
         }
     }
-    Ok(Args {
-        ca_path: ca_path.ok_or(
-            "--ca PATH is required (the pinned CA's certificate, e.g. \
-                    ~/.config/ai-app/certs/ca.pem)",
-        )?,
-        link,
-    })
+    Ok(Args { ca_path, link })
 }
 
 /// What `app.rs`'s `Client::new` needs to talk to the server: the enrolled
@@ -60,14 +56,17 @@ fn parse_args() -> Result<Args, String> {
 /// other way (`DefaultApp::run()` takes no payload).
 fn load_startup_config() -> Result<(EnrolledServer, Vec<u8>), String> {
     let args = parse_args()?;
+    let store = config::store();
     let server = match args.link {
         Some(link) => {
             let server = EnrolledServer::parse_link(&link)?;
-            config::save_enrollment(&server)
+            store
+                .save(&server)
                 .map_err(|e| format!("couldn't save the enrollment: {e}"))?;
             server
         }
-        None => config::load_enrollment()
+        None => store
+            .load()
             .map_err(|e| format!("couldn't read the saved enrollment: {e}"))?
             .ok_or_else(|| {
                 format!(
@@ -77,8 +76,20 @@ fn load_startup_config() -> Result<(EnrolledServer, Vec<u8>), String> {
                 )
             })?,
     };
-    let ca_pem = std::fs::read(&args.ca_path)
-        .map_err(|e| format!("couldn't read the CA at {}: {e}", args.ca_path.display()))?;
+    // `--ca` wins where it was given, so a caller can point a link's
+    // server at a certificate it did not carry -- and so the flag still
+    // means what it did before the link could carry one.
+    let ca_pem = match (&args.ca_path, &server.ca_pem) {
+        (Some(path), _) => std::fs::read(path)
+            .map_err(|e| format!("couldn't read the CA at {}: {e}", path.display()))?,
+        (None, Some(pem)) => pem.clone().into_bytes(),
+        (None, None) => {
+            return Err("this enrollment carries no CA -- pass --ca PATH (e.g. \
+                 ~/.config/ai-app/certs/ca.pem), or enrol again with a link \
+                 minted by a server that includes one"
+                .to_string());
+        }
+    };
     Ok((server, ca_pem))
 }
 

iris/examples/bench_images.rs

@@ -51,7 +51,10 @@ impl DefaultAppState for State {
         }
         let span = rsc.ui.widgets.add_strong(span);
         let span_weak = span.weak();
-        let root = rsc.ui.widgets.add_strong(Scroll::new(span.any(), Axis::Y));
+        let root = rsc
+            .ui
+            .widgets
+            .add_strong(Scroll::new(span.any(), Axis::Y, true));
         ui_state.set_root(root.any());
         Self {
             ui_state,

iris/macro/src/lib.rs

@@ -18,6 +18,12 @@ struct Input {
 }
 
 struct InputFn {
+    /// Everything written above the `fn` -- in practice a `///` doc
+    /// comment, which is why this exists: `masked_by` and its siblings
+    /// are public API and rustdoc is where their contract is read, so a
+    /// macro that silently rejected `///` sent the explanation into an
+    /// ordinary `//` comment nobody generating docs ever sees.
+    attrs: Vec<Attribute>,
     sig: Signature,
     body: Block,
 }
@@ -32,9 +38,10 @@ impl Parse for Input {
         input.parse::<Token![;]>()?;
         let mut fns = Vec::new();
         while !input.is_empty() {
+            let attrs = input.call(Attribute::parse_outer)?;
             let sig = input.parse()?;
             let body = input.parse()?;
-            fns.push(InputFn { sig, body })
+            fns.push(InputFn { attrs, sig, body })
         }
         if !input.is_empty() {
             input.error("function expected");
@@ -59,10 +66,15 @@ pub fn widget_trait(input: TokenStream) -> TokenStream {
         fns,
     } = parse_macro_input!(input as Input);
 
-    let sigs: Vec<_> = fns.iter().map(|f| f.sig.clone()).collect();
+    // The attributes go on the trait's own signature, which is the one
+    // rustdoc renders; the impl gets the bare `fn`.
+    let sigs: Vec<_> = fns
+        .iter()
+        .map(|InputFn { attrs, sig, .. }| quote! { #(#attrs)* #sig })
+        .collect();
     let impls: Vec<_> = fns
         .iter()
-        .map(|InputFn { sig, body }| quote! { #sig #body })
+        .map(|InputFn { sig, body, .. }| quote! { #sig #body })
         .collect();
 
     let Some(GenericParam::Type(state)) = generics.params.first() else {

iris/run-headless.sh

@@ -33,10 +33,10 @@
 # through positionally without disturbing the existing `-- cargo args`
 # convention above.
 #
-# The VM has a virtio-gpu render node (Vulkan 1.4 through Venus, GL 4.6
-# through virgl), so wgpu runs on the host's real GPU -- what is missing is
-# only a compositor to give winit a surface. So: a headless sway, the same
-# trick `emu` uses for the Android emulator, and `grim` to see the result.
+# The VM has a real GPU and no display (the `this-machine-graphics` skill
+# says what it is and how it fails), so what is missing here is only a
+# compositor to give winit a surface. So: a headless sway, the same trick
+# `emu` uses for the Android emulator, and `grim` to see the result.
 #
 # It is deliberately *not* `emu`'s compositor. sway tiles, so adding a window
 # to the one an emulator is sitting in resizes that emulator's window, and a

iris/src/android/platform.rs

@@ -1,10 +1,7 @@
 use crate::platform::OpenUrl;
 use android_view::{
     View,
-    jni::{
-        JNIEnv,
-        objects::{JObject, JValue},
-    },
+    jni::{JNIEnv, objects::JValue},
 };
 
 use super::view::HasAndroidUiState;
@@ -78,7 +75,7 @@ fn try_open_url<'local>(
         &context,
         "startActivity",
         "(Landroid/content/Intent;)V",
-        &[JValue::Object(&JObject::from(intent))],
+        &[JValue::Object(&intent)],
     )?;
     Ok(())
 }

iris/src/android/render.rs

@@ -88,9 +88,10 @@ pub struct FrameDiagnostics {
 }
 
 impl AndroidRenderer {
-    /// `Err` holds a full, human-readable report -- wgpu's own error text
-    /// (`UiRenderNode::new`'s doc comment) plus the adapter identity and
-    /// the limits/downlevel flags bind-group-layout validation checks
+    /// `Err` holds a full, human-readable report for **every** way this
+    /// can fail -- no surface, no adapter, no device, or wgpu's own error
+    /// text (`UiRenderNode::new`'s doc comment) plus the adapter identity
+    /// and the limits/downlevel flags bind-group-layout validation checks
     /// against -- rather than the panic wgpu's default error handler would
     /// otherwise raise with no caller able to see it. This is what aborted
     /// the P0 bench APK on Iris's phone with only "wgpu error: Validation
@@ -107,29 +108,67 @@ impl AndroidRenderer {
         height: u32,
         content_scale: f32,
     ) -> Result<Self, String> {
-        // `force-gles` (RUST.md's I5 "Where iris's frame time goes") swaps
-        // the software-Vulkan (SwiftShader) path for GLES/virgl on the same
-        // build, to isolate whether the backend itself explains the frame
-        // time gap against Compose. `cfg!` rather than a runtime switch:
-        // there is no way to hand an env var to an already-launched Android
-        // process on this machine (see the feature's doc in Cargo.toml).
-        let backends = if cfg!(feature = "force-gles") {
+        // `force-gles` (RUST.md's I5 "Where iris's frame time goes") pins
+        // the build to GLES, to isolate whether the backend itself explains
+        // the frame time gap against Compose. `cfg!` rather than a runtime
+        // switch: there is no way to hand an env var to an already-launched
+        // Android process on this machine (see the feature's doc in
+        // Cargo.toml).
+        //
+        // Otherwise: **Vulkan where it has an adapter at all, GLES where it
+        // has none.** `Backends::PRIMARY` leaves `GL` out, so a device
+        // offering only a GLES adapter had no adapter at all and this
+        // function aborted the process -- this checkout's emulator, whose
+        // Vulkan ICD carries no adapter behind it (`NotFound {
+        // active_backends: VULKAN, no_adapter_backends: VULKAN,
+        // supported_backends: VULKAN | GL }`), and the crash loop in
+        // RUST.md's queue.
+        //
+        // The choice is made *before any surface exists*, with an instance
+        // that never touches the window, because **an Android window can be
+        // connected to one graphics API only**. One instance carrying both
+        // backends does not work: `create_surface` builds a raw surface per
+        // backend, Vulkan's `vkCreateAndroidSurfaceKHR` claims the window
+        // first, and the GLES surface made from the same window then fails
+        // `configure` as lost -- measured here as "In Surface::configure /
+        // Invalid surface" followed by an abort in
+        // `Surface::get_current_texture_view`, "Surface is not configured
+        // for presentation".
+        let mut backends = if cfg!(feature = "force-gles") {
             Backends::GL
         } else {
             Backends::PRIMARY
         };
-        let instance = Instance::new(&InstanceDescriptor {
+        let mut instance = Instance::new(&InstanceDescriptor {
             backends,
             ..Default::default()
         });
+        // A build already pinned to GLES has nowhere to fall back to.
+        if backends != Backends::GL && instance.enumerate_adapters(backends).block_on().is_empty() {
+            log::warn!(
+                "iris renderer: no {backends:?} adapter on this device, falling back to GLES"
+            );
+            backends = Backends::GL;
+            instance = Instance::new(&InstanceDescriptor {
+                backends,
+                ..Default::default()
+            });
+        }
 
         // SAFETY: the `NativeWindow` outlives the surface built from it --
         // android-view drops the old renderer (and this surface with it)
         // before handing over a new window, in `surface_changed` below.
         let surface = instance
             .create_surface(SurfaceTarget::from(AndroidWindowHandle { window }))
-            .expect("Could not create android surface!");
+            .map_err(|error| format!("Could not create the android surface: {error}"))?;
 
+        // Every step from here to a live device reports rather than
+        // panics, for the one reason: on the phone these builds run on
+        // there is no `adb`, so an abort's message reaches a tombstone
+        // nobody can read and the launcher simply restarts the app --
+        // which is what a crash loop with no explanation is. The caller
+        // (`android::view::IrisViewPeer::surface_changed`) puts this
+        // string on screen and in the app's own log ring instead.
         let adapter = instance
             .request_adapter(&RequestAdapterOptions {
                 power_preference: PowerPreference::default(),
@@ -137,19 +176,7 @@ impl AndroidRenderer {
                 force_fallback_adapter: false,
             })
             .block_on()
-            .expect("Could not get adapter!");
-
-        // Requesting the device itself still panics on failure: that is a
-        // `RequestDeviceError` (a limit or feature the adapter cannot grant
-        // at all), a different and already-diagnosable failure from the one
-        // this function now recovers from -- `RUST.md`'s "Software mode ...
-        // crashes for a third, different reason" is exactly that class, and
-        // its message already names the limit and the requested/allowed
-        // values with no truncation risk (it never reaches wgpu's
-        // uncaptured-error path). What this function's `Result` return
-        // covers is the *next* class of failure: the adapter grants the
-        // device, and validation only fails once a specific bind group
-        // layout is checked against it.
+            .map_err(|error| format!("No usable GPU adapter for backends {backends:?}: {error}"))?;
 
         // Same request as the winit backend's `UiRenderer::new` -- no
         // binding-array features, see TEXTURES.md's "Recommended shape".
@@ -161,7 +188,13 @@ impl AndroidRenderer {
                 ..Default::default()
             })
             .block_on()
-            .expect("Could not get device!");
+            .map_err(|error| {
+                format!(
+                    "The adapter {} ({:?}) refused a device: {error}",
+                    adapter.get_info().name,
+                    adapter.get_info().backend,
+                )
+            })?;
 
         // wgpu's default handler for an error raised outside `UiRenderNode::
         // new`'s own error scopes (i.e. everything past device creation --
@@ -181,11 +214,27 @@ impl AndroidRenderer {
         let info = adapter.get_info();
         let adapter_name = info.name.clone();
         let adapter_backend = info.backend;
-        let adapter_driver = if info.driver_info.is_empty() {
-            info.driver.clone()
-        } else {
-            format!("{} {}", info.driver, info.driver_info)
-        };
+        // Either half can be empty -- the emulator's GLES adapter reports
+        // no `driver` and a long `driver_info`, so joining unconditionally
+        // left a leading space in every log line it appears in.
+        let adapter_driver = [info.driver.as_str(), info.driver_info.as_str()]
+            .into_iter()
+            .filter(|part| !part.is_empty())
+            .collect::<Vec<_>>()
+            .join(" ");
+        // Say which adapter won, in the same words `default::render` uses,
+        // and at startup rather than only on the Diagnostics page: the
+        // backend alone (logged by `view.rs` when a renderer is built) does
+        // not separate the cases that matter. In this checkout's emulator
+        // `Gl` is the host's real GPU through virgl, and `Gl` under
+        // `EMU_GPU=software` is SwiftShader on the CPU; on a phone `Vulkan`
+        // is the device's own driver. A frame time or a screenshot with no
+        // record of which of those produced it cannot be read, and the
+        // fallback above is silent by design.
+        log::info!(
+            "iris renderer: {adapter_name} ({adapter_backend:?}, {adapter_driver}) on \
+             {backends:?}"
+        );
 
         let surface_caps = surface.get_capabilities(&adapter);
         let surface_format = surface_caps
@@ -308,6 +357,7 @@ impl AndroidRenderer {
              mono={default_mono_family:?}\n\
              fonts resolved: regular={regular:?} bold={bold:?} italic={italic:?} \
              mono={mono:?}\n\
+             icon font: {icons:?}\n\
              wgpu errors since surface creation:\n  {errors_text}\n\n\
              {frame_report}",
             name = self.adapter_name,
@@ -322,6 +372,7 @@ impl AndroidRenderer {
             bold = font.bold_resolved,
             italic = font.italic_resolved,
             mono = font.mono_resolved,
+            icons = font.icon_family,
         )
     }
 

iris/src/android/view.rs

@@ -7,7 +7,7 @@ use android_view::{
     jni::{
         JNIEnv, JavaVM,
         objects::{GlobalRef, JValue},
-        sys::{jint, jlong},
+        sys::jint,
     },
     ndk::event::{Axis, Keycode, MotionAction},
 };
@@ -20,7 +20,7 @@ use std::{
     marker::{PhantomData, Sized},
     rc::Rc,
     sync::Arc,
-    time::{Duration, Instant},
+    time::Instant,
 };
 
 use super::{
@@ -312,12 +312,11 @@ pub struct IrisViewPeer<State: AndroidAppState> {
     pub(super) render: UiRenderState,
     pub(super) state: State,
     task_recv: TaskMsgReceiver<AndroidRsc<State>>,
-    /// `(an Instant, the input-event nanosecond stamp it was taken at)`,
-    /// captured from the first `MotionEvent` this view receives and never
-    /// changed after -- how `on_touch_event` dates every touch sample. Its
-    /// path out is the peer's own drop: it holds nothing but two numbers
-    /// and is meaningless to any other view.
-    input_clock: Option<(Instant, jlong)>,
+    /// Anchored on the first `MotionEvent` this view receives and never
+    /// re-anchored after -- how `on_touch_event` dates every touch sample.
+    /// Its path out is the peer's own drop: it holds nothing but three
+    /// numbers and is meaningless to any other view.
+    input_clock: Option<PointerClock>,
 }
 
 impl<State: 'static, I: RscIdx<AndroidRsc<State>>> std::ops::Index<I> for AndroidRsc<State> {
@@ -406,7 +405,13 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
     /// magenta and screenshotting), but no primitive ever appears on top of
     /// it -- on both the Vulkan/SwiftShader and GLES/virgl backends. Leave
     /// these in until that is root-caused; removing them loses the exact
-    /// evidence a `logcat` capture needs to reproduce the state.
+    /// evidence a `logcat` capture needs to reproduce the state. Gated on
+    /// `iris::diagnostics::trace_enabled` since 2026-09-07 (docs/RUST.md's
+    /// review, D1): unconditional, they were two `debug!` lines every
+    /// rendered frame, and `client_core::log_ring`'s `RingLogger` records
+    /// every level the app's already-`Debug` install lets through
+    /// regardless of target, so they filled the whole ring in under ten
+    /// seconds at 120Hz and left `Copy report` nothing else to show.
     fn render(&mut self, ctx: &mut CallbackCtx) {
         if self.state.android_state().renderer.is_none() {
             return;
@@ -441,18 +446,29 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
             self.state.on_insets_changed(&mut self.rsc, physical);
         }
 
-        let ui_state = self.state.android_state();
-        log::debug!(
-            "render(): root={:?} widgets={} active={} root_px={:?} out_size={:?}",
-            ui_state.root.is_some(),
-            self.rsc.widgets().len(),
-            self.render.active_widgets(),
-            ui_state
-                .root
-                .as_ref()
-                .and_then(|r| self.render.window_region(r, &self.rsc)),
-            self.window_size(),
-        );
+        // Gated the same way `iris::frame`'s own line is (docs/RUST.md's
+        // "Phone logging" review, D1): a bare `log::debug!` reaches
+        // `client_core::log_ring`'s ring regardless of level, since
+        // `RingLogger::enabled` is unconditionally `true` and the app
+        // installs at `LevelFilter::Debug` -- two of these a rendered
+        // frame filled the 2000-line ring in under ten seconds at 120Hz,
+        // leaving `Copy report` nothing but frame spam. See
+        // `iris::diagnostics`'s module doc.
+        if crate::diagnostics::trace_enabled() {
+            let ui_state = self.state.android_state();
+            log::debug!(
+                target: "iris::frame",
+                "render(): root={:?} widgets={} active={} root_px={:?} out_size={:?}",
+                ui_state.root.is_some(),
+                self.rsc.widgets().len(),
+                self.render.active_widgets(),
+                ui_state
+                    .root
+                    .as_ref()
+                    .and_then(|r| self.render.window_region(r, &self.rsc)),
+                self.window_size(),
+            );
+        }
         // iris's own frame-time report (RUST.md's I5 box, "Measurements
         // taken" (b)): started here, at the same point a redraw request
         // fires, and stopped after `renderer.draw()`'s `queue.submit` +
@@ -498,21 +514,25 @@ impl<State: AndroidAppState> IrisViewPeer<State> {
             .android_state_mut()
             .frame_report
             .record_split(frame_start.elapsed(), submit_to_present);
+        crate::diagnostics::log_frame(&self.render, frame_start, submit_to_present, animating);
         // A frame callback is one-shot, so an animation that wants
         // another frame has to say so every frame -- unlike `after_input`,
         // which only has to ask when input dirtied something.
         if animating {
             ctx.view.post_frame_callback(&mut ctx.env);
         }
-        let ui_state = self.state.android_state();
-        log::debug!(
-            "render(): after update active={} root_px={:?}",
-            self.render.active_widgets(),
-            ui_state
-                .root
-                .as_ref()
-                .and_then(|r| self.render.window_region(r, &self.rsc)),
-        );
+        if crate::diagnostics::trace_enabled() {
+            let ui_state = self.state.android_state();
+            log::debug!(
+                target: "iris::frame",
+                "render(): after update active={} root_px={:?}",
+                self.render.active_widgets(),
+                ui_state
+                    .root
+                    .as_ref()
+                    .and_then(|r| self.render.window_region(r, &self.rsc)),
+            );
+        }
 
         // I4 (RUST.md): only produces a `TreeUpdate` -- and so only queues
         // anything to raise -- when the named set actually changed this
@@ -618,17 +638,30 @@ impl<State: AndroidAppState> ViewPeer for IrisViewPeer<State> {
         // `CLOCK_MONOTONIC` an `Instant` reads, so a single
         // `(Instant, nanos)` pair converts every later sample exactly.
         // Anchoring **once** rather than per event is what keeps the times
-        // ordered: a fresh `Instant::now()` per event, minus each sample's
-        // age inside it, can date a later event's first historical sample
-        // before the previous event's last one whenever delivery jitters by
-        // more than the batch spans -- and `VelocityTracker::add_sample`'s
-        // debug assert would rightly fire on that. See `CursorState::time`.
+        // ordered, and anchoring on the first event's *oldest* sample
+        // rather than on its own time is what keeps that event's batch
+        // from collapsing onto one instant -- `sense::PointerClock`'s doc
+        // has both, and owns the arithmetic so it can be unit-tested off a
+        // device (`sense_tests.rs`). See `CursorState::time`.
         let event_time = event.event_time_nanos(&mut ctx.env);
-        let (anchor_at, anchor_nanos) =
-            *self.input_clock.get_or_insert((Instant::now(), event_time));
-        let at = |sample_time: jlong| {
-            anchor_at + Duration::from_nanos(sample_time.saturating_sub(anchor_nanos).max(0) as u64)
-        };
+        if self.input_clock.is_none() {
+            let history = event.history_size(&mut ctx.env);
+            let oldest = if history > 0 {
+                event.historical_event_time_nanos(&mut ctx.env, 0)
+            } else {
+                event_time
+            };
+            self.input_clock = Some(PointerClock::anchored(Instant::now(), event_time, oldest));
+        }
+        let mut clock = self.input_clock.expect("anchored just above");
+        // `iris::input`'s own doc (`sense::log_input_event`): collected
+        // only when tracing is on, since this is otherwise a `Vec` per
+        // `MotionEvent` for a line nobody is reading -- the JNI reads
+        // themselves (`historical_axis`/`historical_event_time_nanos`
+        // below) already happen unconditionally, for the replay this
+        // function does regardless of tracing.
+        let trace_input = crate::diagnostics::trace_enabled();
+        let mut historical_ms: Vec<(u64, f32, f32)> = Vec::new();
 
         // **Historical samples first.** A flick on a 120Hz screen is
         // delivered as one or two `MotionEvent`s with the intermediate
@@ -648,31 +681,30 @@ impl<State: AndroidAppState> ViewPeer for IrisViewPeer<State> {
             // the event's own sample as the newest of the batch; everything
             // downstream (`VelocityTracker`, `DragArbiter`'s long-press
             // clock) assumes it, so say so here rather than at each reader.
-            let mut previous = anchor_nanos;
+            // `PointerClock::sample` is what asserts it, and it carries the
+            // last sample seen *across* events, so the first sample of
+            // every event is checked against the previous event's last one
+            // rather than against the anchor.
             for pos in 0..history {
                 let hx = event.historical_axis(&mut ctx.env, Axis::X, 0, pos);
                 let hy = event.historical_axis(&mut ctx.env, Axis::Y, 0, pos);
                 let ht = event.historical_event_time_nanos(&mut ctx.env, pos);
-                debug_assert!(
-                    ht >= previous,
-                    "historical sample {pos} of {history} is dated {ht}ns, before the {previous}ns \
-                     sample ahead of it -- the input clock is not what this assumes"
-                );
-                previous = ht;
+                let sample_at = clock.sample(ht);
+                if trace_input {
+                    historical_ms.push((clock.ms_since_anchor(ht), hx, hy));
+                }
                 let ui_state = self.state.android_state_mut();
                 ui_state.cursor.pos = vec2(hx, hy);
-                ui_state.cursor.time = at(ht);
+                ui_state.cursor.time = sample_at;
                 self.run_input_frame(ctx);
             }
-            debug_assert!(
-                event_time >= previous,
-                "the event's own sample is dated {event_time}ns, before its last historical \
-                 sample at {previous}ns"
-            );
         }
 
+        let event_at = clock.sample(event_time);
+        let event_ms = clock.ms_since_anchor(event_time);
+        self.input_clock = Some(clock);
         let ui_state = self.state.android_state_mut();
-        ui_state.cursor.time = at(event_time);
+        ui_state.cursor.time = event_at;
         match action {
             MotionAction::Down => {
                 ui_state.cursor.pos = vec2(x, y);
@@ -695,6 +727,16 @@ impl<State: AndroidAppState> ViewPeer for IrisViewPeer<State> {
             }
             _ => return false,
         }
+        if trace_input {
+            let action_word = match action {
+                MotionAction::Down => "down",
+                MotionAction::Move => "move",
+                MotionAction::Up => "up",
+                MotionAction::Cancel => "cancel",
+                _ => "other",
+            };
+            crate::sense::log_input_event(action_word, x, y, event_ms, &historical_ms);
+        }
         self.after_input(ctx);
         true
     }
@@ -811,29 +853,33 @@ 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.
+                // A genuinely new renderer means a genuinely new GPU
+                // device, holding none of the textures the old one did --
+                // while the CPU side of them (`UiData::textures`, and the
+                // glyph atlas built on it) lives on `self.rsc` and
+                // survives. So every slot has to be uploaded again, and
+                // `Textures::reupload` queues exactly that, in slot order.
+                //
+                // It replaces clearing them, which threw away the *slot
+                // numbering* as well as the pixels: every `TextureHandle`
+                // a live widget still held -- one per icon or image on
+                // screen, and one per folded card at the time -- then
+                // named a slot nothing recognised, and the next frame
+                // panicked in `image_bind_group` ("texture slot 89 is not
+                // a live standalone image: None"). Re-uploading also keeps
+                // the glyph atlas, so an app switch no longer re-rasterises
+                // every glyph on screen. This only runs on the branch that
+                // actually builds a new renderer, never on the reuse
+                // branch above, where the textures are still on the device
+                // that holds them.
                 log::info!(
-                    "iris surface: new renderer built ({:?}), clearing glyph atlas: \
+                    "iris surface: new renderer built ({:?}), re-uploading textures: \
                      glyphs={} pages={}",
                     renderer.adapter_backend,
                     self.rsc.ui.text.atlas.glyph_count(),
                     self.rsc.ui.text.atlas.page_count(),
                 );
-                self.rsc.ui.text.atlas.clear();
-                self.rsc.ui.textures.reset();
+                self.rsc.ui.textures.reupload();
                 self.state.android_state_mut().renderer = Some(renderer);
                 self.render(ctx);
             }

iris/src/attr.rs

@@ -52,8 +52,16 @@ pub fn recent_click(last_click: &mut Instant) -> bool {
 /// 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").
+/// `CursorSense::Cancel` is in the set for the same reason `DragGesture`
+/// registers it: if a scroll area or a list takes the pointer mid-gesture,
+/// this field sees no `PressEnd`, and a `press_origin` left set is then
+/// compared against the *next* press -- a stray selection, or a keyboard
+/// summoned by a tap somewhere else entirely.
 fn press_track() -> CursorSenses {
-    CursorSense::click() | CursorSense::Pressing(CursorButton::Left) | CursorSense::unclick()
+    CursorSense::click()
+        | CursorSense::Pressing(CursorButton::Left)
+        | CursorSense::unclick()
+        | CursorSense::Cancel
 }
 
 pub struct Selector;
@@ -185,6 +193,7 @@ fn on_press(
                     state.focus_gained(render.window_region(&id, &*rsc));
                 }
             }
+            CursorSense::Cancel => id.edit(rsc).text.press_origin = None,
             _ => {}
         }
         return;
@@ -205,6 +214,9 @@ fn on_press(
                 ctx.text.press_origin = None;
             }
         }
+        // The gesture was taken by somebody else, so it is not a tap and
+        // must not grant focus when it ends out of this widget's sight.
+        CursorSense::Cancel => id.edit(rsc).text.press_origin = None,
         CursorSense::PressEnd(_) => {
             let was_tap = id.edit(rsc).text.press_origin.take().is_some();
             if was_tap {

iris/src/default/app.rs

@@ -27,6 +27,10 @@ pub struct App<State: AppState> {
 
 impl<State: AppState> App<State> {
     pub fn run() {
+        // The desktop's `main` in everything but name -- see
+        // `super::logging`'s doc for why the logger goes here and what
+        // its absence hid.
+        super::logging::install(log::LevelFilter::Info);
         let event_loop = EventLoop::with_user_event().build().unwrap();
         let proxy = event_loop.create_proxy();
         event_loop

iris/src/default/logging.rs

@@ -0,0 +1,86 @@
+//! A stderr logger for the desktop entry point.
+//!
+//! Without one, `log::` calls on this side go nowhere: `log`'s default is
+//! a no-op logger, and nothing in `desktop-app` or the examples ever
+//! installed a real one. That is how iris came to have a renderer that
+//! silently fell back to GLES (and, on this VM, on to llvmpipe when the
+//! host took its GPU away) with **no record anywhere of what
+//! drew the frame** -- a layer-2 screenshot off llvmpipe and one off the
+//! host GPU are the same PNG, and the difference is exactly what a
+//! screenshot is being taken to judge.
+//!
+//! Installed by [`DefaultApp::run`](super::app::DefaultApp::run) rather
+//! than by a library call somewhere, because that function already takes
+//! over the process -- it owns the event loop and does not return -- so
+//! it is the desktop's `main` in everything but name, and one install
+//! there covers `desktop-app` and every example at once. `try_init`
+//! rather than `init`: a binary that installed its own logger first keeps
+//! it, and a second `DefaultApp::run` in one process is not an error.
+//!
+//! Deliberately not `env_logger`. All this owes the reader is a level and
+//! a line, which is a page of code against a dependency plus its own
+//! filter dialect; the Android side is `android_logger` for the same
+//! reason -- one line per platform's own convention.
+
+use std::io::Write;
+
+use log::{Level, LevelFilter, Log, Metadata, Record};
+
+/// Reads one level name from `RUST_LOG` -- `off`, `error`, `warn`,
+/// `info`, `debug`, `trace`, case-insensitively. **Not env_logger's
+/// per-module filter syntax**: anything else is ignored and the default
+/// stands, rather than being silently read as "off", since a typo that
+/// turned logging off would be indistinguishable from a quiet program.
+fn level_from_env(default: LevelFilter) -> LevelFilter {
+    match std::env::var("RUST_LOG") {
+        Ok(text) => text.trim().parse().unwrap_or(default),
+        Err(_) => default,
+    }
+}
+
+struct StderrLogger {
+    level: LevelFilter,
+}
+
+impl Log for StderrLogger {
+    fn enabled(&self, metadata: &Metadata) -> bool {
+        metadata.level() <= self.level
+    }
+
+    fn log(&self, record: &Record) {
+        if !self.enabled(record.metadata()) {
+            return;
+        }
+        // One write, not a `writeln!` per part: two threads logging at
+        // once interleave otherwise, and the frame and input traces are
+        // both written from whichever thread produced them.
+        let line = format!(
+            "{level:<5} {target}: {args}\n",
+            level = match record.level() {
+                Level::Error => "ERROR",
+                Level::Warn => "WARN",
+                Level::Info => "INFO",
+                Level::Debug => "DEBUG",
+                Level::Trace => "TRACE",
+            },
+            target = record.target(),
+            args = record.args(),
+        );
+        let _ = std::io::stderr().write_all(line.as_bytes());
+    }
+
+    fn flush(&self) {
+        let _ = std::io::stderr().flush();
+    }
+}
+
+/// Installs the stderr logger unless this process already has one.
+/// Defaults to `info`, which is where the renderer says which adapter it
+/// got; `RUST_LOG=debug` adds iris's own per-frame lines.
+pub fn install(default: LevelFilter) {
+    let level = level_from_env(default);
+    let logger = Box::leak(Box::new(StderrLogger { level }));
+    if log::set_logger(logger).is_ok() {
+        log::set_max_level(level);
+    }
+}

iris/src/default/mod.rs

@@ -15,6 +15,7 @@ mod access;
 mod app;
 mod attr;
 mod input;
+mod logging;
 mod platform;
 mod render;
 
@@ -294,6 +295,31 @@ impl<State: DefaultAppState> AppState for DefaultApp<State> {
             ui_state.focus = None;
         }
         if input_changed {
+            // The winit half of `iris::input` (`sense::log_input_event`'s
+            // own doc): no batching here, so `historical` is always empty
+            // -- winit hands one `WindowEvent` per pointer sample, unlike
+            // Android's `MotionEvent`. The action is read back off the
+            // buttons `Input::event` just updated, the same test
+            // `GestureOutcome`'s callers already use to tell a press from a
+            // release. Computed only when tracing is on, same reasoning as
+            // `log_input_event` itself gating on it.
+            if crate::diagnostics::trace_enabled() {
+                let action = if cursor_state.buttons.left.is_start() {
+                    "down"
+                } else if cursor_state.buttons.left.is_end() {
+                    "up"
+                } else {
+                    "move"
+                };
+                let t_ms = cursor_state.time.duration_since(render.epoch()).as_millis() as u64;
+                crate::sense::log_input_event(
+                    action,
+                    cursor_state.pos.x,
+                    cursor_state.pos.y,
+                    t_ms,
+                    &[],
+                );
+            }
             let window_size = ui_state.window_size();
             render.run_sensors(rsc, state, cursor_state, window_size);
         }
@@ -313,11 +339,14 @@ impl<State: DefaultAppState> AppState for DefaultApp<State> {
                 // `IrisViewPeer::render`'s `post_frame_callback` does on
                 // Android. Nothing else in iris moves without an input
                 // event.
-                let animating = rsc.ui_mut().tick_animations(std::time::Instant::now());
+                let frame_start = std::time::Instant::now();
+                let animating = rsc.ui_mut().tick_animations(frame_start);
                 let ui_state = state.default_state_mut();
                 render.update(&ui_state.root, rsc);
                 ui_state.renderer.update(&mut rsc.ui, render);
+                let draw_start = std::time::Instant::now();
                 ui_state.renderer.draw();
+                crate::diagnostics::log_frame(render, frame_start, draw_start.elapsed(), animating);
                 if animating {
                     ui_state.window.request_redraw();
                 }
@@ -334,6 +363,22 @@ impl<State: DefaultAppState> AppState for DefaultApp<State> {
                 render.resize((size.width, size.height));
                 ui_state.renderer.resize(size)
             }
+            // Dragging the window to a display with a different scale.
+            // Both copies again, the pair `new` sets at startup -- read
+            // through `content_scale` rather than from the event, so
+            // `IRIS_SCALE` still pins the density it was given (the
+            // `--phone` window must not follow the monitor). winit sends
+            // the matching `Resized` separately. Before 2026-09-07 this
+            // event was unhandled, so every `dp` and every rasterised
+            // glyph stayed at the density the window opened on
+            // (docs/REVIEW-2026-09-07.md's R5) -- invisible on this
+            // machine, where every display is 1.0.
+            WindowEvent::ScaleFactorChanged { .. } => {
+                let scale = content_scale(ui_state.window.as_ref());
+                rsc.ui.text.density = scale;
+                render.set_density(scale);
+                ui_state.window.request_redraw();
+            }
             WindowEvent::KeyboardInput { event, .. } => {
                 if let Some(sel) = ui_state.focus
                     && event.state.is_pressed()

iris/src/default/render.rs

@@ -82,19 +82,40 @@ impl UiRenderer {
     pub fn new(window: Arc<Window>) -> Self {
         let size = window.inner_size();
 
-        let instance = Instance::new(&InstanceDescriptor {
-            // `force-gles` on the desktop too, not just on Android: the
-            // GLES backend has behaviour of its own (a one-layer array
-            // texture is a `GL_TEXTURE_2D` -- see
-            // `GpuTextures::create_array_texture`), and a machine with a
-            // real GPU is where that is cheap to reproduce and screenshot.
-            backends: if cfg!(feature = "force-gles") {
-                Backends::GL
-            } else {
-                Backends::PRIMARY
-            },
+        // `force-gles` on the desktop too, not just on Android: the
+        // GLES backend has behaviour of its own (a one-layer array
+        // texture is a `GL_TEXTURE_2D` -- see
+        // `GpuTextures::create_array_texture`), and a machine with a
+        // real GPU is where that is cheap to reproduce and screenshot.
+        let mut backends = if cfg!(feature = "force-gles") {
+            Backends::GL
+        } else {
+            Backends::PRIMARY
+        };
+        let mut instance = Instance::new(&InstanceDescriptor {
+            backends,
             ..Default::default()
         });
+        // The same fallback the Android backend grew in 85869d0, and for
+        // the same reason: a machine can advertise a Vulkan ICD with no
+        // device behind it, and refusing to draw at all because the only
+        // usable adapter is a GLES one is iris's bug rather than the
+        // machine's. On this VM the Vulkan device disappears whenever
+        // the host refuses a virtio-gpu context, so `run-headless.sh` --
+        // layer 2 of the test rig -- aborted with `Could not get
+        // adapter!` while GL was sitting there working. Probed before the
+        // surface exists, matching Android, where an instance carrying
+        // both backends fails worse than one carrying the wrong one.
+        if backends != Backends::GL && instance.enumerate_adapters(backends).block_on().is_empty() {
+            log::warn!(
+                "iris renderer: no {backends:?} adapter on this machine, falling back to GLES"
+            );
+            backends = Backends::GL;
+            instance = Instance::new(&InstanceDescriptor {
+                backends,
+                ..Default::default()
+            });
+        }
 
         let surface = instance
             .create_surface(window.clone())
@@ -107,7 +128,32 @@ impl UiRenderer {
                 force_fallback_adapter: false,
             })
             .block_on()
-            .expect("Could not get adapter!");
+            .unwrap_or_else(|error| {
+                panic!("No usable GPU adapter for backends {backends:?}: {error}")
+            });
+
+        // Say which adapter won, in the same words the Android backend
+        // uses. Without it a layer-2 screenshot or frame time from this
+        // window carries no record of what drew it, and the two cases that
+        // matter look identical in the PNG: the host's real GPU, and
+        // llvmpipe after this VM lost its virtio-gpu contexts. That
+        // happened on 2026-09-08, and the only reason anyone noticed is
+        // that the fallback above did not exist yet and the app aborted
+        // instead. A silent fallback needs this line to stay honest.
+        {
+            let info = adapter.get_info();
+            log::info!(
+                "iris renderer: {name} ({backend:?}, {driver}{driver_info}) on {backends:?}",
+                name = info.name,
+                backend = info.backend,
+                driver = info.driver,
+                driver_info = if info.driver_info.is_empty() {
+                    String::new()
+                } else {
+                    format!(" {}", info.driver_info)
+                },
+            );
+        }
 
         // No features beyond what wgpu asks for by default, and no
         // binding-array limits: the atlas is one texture_2d_array and a

iris/src/diagnostics.rs

@@ -0,0 +1,83 @@
+//! The trace toggle for the `iris::input`/`iris::frame` diagnostics (Iris's
+//! 2026-09-07 request: "add another button to copy input event info ...
+//! instrument a lot of the code with timings"), and the one place both
+//! call sites' `iris::frame` line is written from.
+//!
+//! **Why a crate-level flag instead of `log::log_enabled!`/
+//! `log::set_max_level`**: the app already installs its logger at
+//! `LevelFilter::Debug` (`iris/android-app/src/lib.rs`'s `JNI_OnLoad`), so
+//! a `log::Level::Debug` line reaches `client_core::log_ring`'s ring
+//! regardless of what this instrument would prefer -- `RingLogger::enabled`
+//! is unconditionally `true` by design (its own doc: "the ring wants
+//! everything"). So the level alone cannot give these two targets a
+//! default-off switch; the gate has to live on this side, checked before
+//! `log::debug!` is even reached.
+//!
+//! **Why default off matters**: the ring is 2000 lines / 256 KiB
+//! (`client_core::log_ring::DEFAULT_MAX_LINES`/`DEFAULT_MAX_BYTES`), and a
+//! 120Hz session logging both a line per touch sample and a line per frame
+//! fills that in seconds -- so a caller turns this on only for the length
+//! of whatever is being investigated, and the report says so at its top
+//! (a caller's job; see `iris::diagnostics::trace_enabled` used at the top
+//! of whatever builds the report).
+//!
+//! **Not yet wired to a control**: the Diagnostics pane that would hold the
+//! switch is in `iris/android-app/src/bench_client.rs`, which another agent
+//! has open at the same time this was written. `set_trace` is the whole
+//! surface a button needs; wiring one is a follow-up.
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::{Duration, Instant};
+
+use iris_core::UiRenderState;
+
+static TRACE: AtomicBool = AtomicBool::new(false);
+
+/// Turns the `iris::input`/`iris::frame` `debug!` lines on or off. Off by
+/// default -- see the module doc for why turning the level on alone would
+/// not do it.
+pub fn set_trace(on: bool) {
+    TRACE.store(on, Ordering::Relaxed);
+}
+
+/// Whether the `iris::input`/`iris::frame` lines are enabled right now --
+/// what a report's header reads before deciding what to say about the
+/// lines it does or doesn't hold (UI_RULES.md: "design the unknown state
+/// first").
+pub fn trace_enabled() -> bool {
+    TRACE.load(Ordering::Relaxed)
+}
+
+/// One `iris::frame` line, called once per frame from each backend's own
+/// frame function -- `android::view::IrisViewPeer::render`,
+/// `default::DefaultApp::window_event`'s `RedrawRequested` arm, and
+/// `harness::Harness::frame` -- after the draw (or, on the harness, where a
+/// draw would be; `draw` is `Duration::ZERO` there since nothing is
+/// actually submitted to a GPU).
+///
+/// `render.update(...)` must already have run this frame: this reads back
+/// what it recorded (`UiRenderState::last_layout_duration`/
+/// `last_redraw_kind`/`frame_number`) rather than timing anything itself,
+/// so a caller's own measurement of the phase around `update()` and around
+/// its own draw call are the only two `Instant` pairs in the whole path --
+/// see each call site's own comment for why it is not restructured to fit
+/// this instead.
+pub fn log_frame(render: &UiRenderState, now: Instant, draw: Duration, animating: bool) {
+    if !trace_enabled() {
+        return;
+    }
+    let since_input = render
+        .time_since_input(now)
+        .map(|d| format!("{}ms", d.as_millis()))
+        .unwrap_or_else(|| "none".to_string());
+    log::debug!(
+        target: "iris::frame",
+        "iris frame: n={} now={}ms since_input={since_input} layout={:?} draw={:?} \
+         redraw={:?} primitives={} animating={animating}",
+        render.frame_number(),
+        now.duration_since(render.epoch()).as_millis(),
+        render.last_layout_duration(),
+        draw,
+        render.last_redraw_kind(),
+        render.active_primitive_count(),
+    );
+}

iris/src/event.rs

@@ -44,13 +44,19 @@ impl<WL: WidgetLike<Rsc, Tag>, Rsc: HasEvents, Tag> Eventable<Rsc, Tag> for WL {
 
 widget_trait! {
     pub trait TaskEventable<Rsc: HasEvents + HasTasks>;
-    fn task_on<'a, E: EventLike, F: AsyncWidgetEventFn<Rsc, WL::Widget>>(
+    /// No `Data: Send` bound, deliberately: the registered handler below
+    /// takes `|_, rsc|` and the event's data never crosses into the
+    /// spawned future -- `AsyncEventIdCtx` carries the widget id and the
+    /// task handle and nothing else. The bound used to be here anyway, and
+    /// it was the whole reason `CursorData`'s pointer state was behind a
+    /// `Mutex` rather than owned by the input handler (Iris, 2026-09-08:
+    /// never reach for a lock first).
+    fn task_on<E: EventLike, F: AsyncWidgetEventFn<Rsc, WL::Widget>>(
         self,
         event: E,
         f: F,
     ) -> impl WidgetIdFn<Rsc, WL::Widget>
-        where <E::Event as Event>::Data<'a>: Send,
-        for<'b> F::CallRefFuture<'b>: Send,
+        where for<'b> F::CallRefFuture<'b>: Send,
     {
         let f = Arc::new(f);
         move |rsc| {

iris/src/harness.rs

@@ -58,6 +58,18 @@ impl TouchAction {
             _ => None,
         }
     }
+
+    /// The inverse of [`Self::parse`] -- what [`Harness::touch`] hands
+    /// [`crate::sense::log_input_event`], so an `iris::input` line and a
+    /// `.touch` file agree on one spelling of each action.
+    pub fn word(self) -> &'static str {
+        match self {
+            Self::Down => "down",
+            Self::Move => "move",
+            Self::Up => "up",
+            Self::Cancel => "cancel",
+        }
+    }
 }
 
 #[derive(Clone, Copy, Debug)]
@@ -345,8 +357,14 @@ impl Harness {
             update(&mut self.state, &mut self.rsc);
         }
         let now = self.at(t_ms);
-        self.rsc.ui.tick_animations(now);
+        let animating = self.rsc.ui.tick_animations(now);
         self.render.update(&self.state.root, &mut self.rsc);
+        // No GPU here, so there is no draw phase to time -- `draw` is
+        // always zero. `layout`/`redraw`/`primitives` are still real,
+        // because `render.update` just ran; see
+        // `iris::diagnostics::log_frame`'s own doc for why this reads
+        // those back rather than timing anything itself.
+        crate::diagnostics::log_frame(&self.render, now, Duration::ZERO, animating);
     }
 
     /// Frames every `step_ms` up to and including `end_ms` -- what a
@@ -377,6 +395,12 @@ impl Harness {
             TouchAction::Move => {}
             TouchAction::Up | TouchAction::Cancel => self.cursor.buttons.left.update(false),
         }
+        // Layer 1's half of `iris::input` (`sense::log_input_event`'s own
+        // doc): no batching happens here, so `historical` is always empty
+        // and `t_ms` is the script's own column, which is what makes this
+        // round-trip through `report_to_touch.py` back into an identical
+        // `TouchScript`.
+        crate::sense::log_input_event(action.word(), pos.x, pos.y, t_ms, &[]);
         let cursor = self.cursor.clone();
         self.render
             .run_sensors(&mut self.rsc, &mut self.state, cursor, self.size);

iris/src/layout_tests.rs

@@ -50,7 +50,10 @@ fn scrolled_rects(
         span.push(row.any());
     }
     let span = rsc.ui.widgets.add_strong(span);
-    let scroll = rsc.ui.widgets.add_strong(Scroll::new(span.any(), Axis::Y));
+    let scroll = rsc
+        .ui
+        .widgets
+        .add_strong(Scroll::new(span.any(), Axis::Y, true));
     let weak = scroll.weak();
     (weak, scroll.any(), rects)
 }
@@ -150,8 +153,8 @@ fn hit_testing_follows_a_scrolled_widget() {
 /// `ActiveData::mask` is the mask a widget was drawn **under**, not the one
 /// it set for itself -- `redraw` feeds it straight back in as the inherited
 /// mask, so storing the set one hands a `Masked` its own mask the second
-/// time round and trips `Painter::set_mask`'s nested-mask assert. That was
-/// an abort (`assertion failed: self.mask == MaskIdx::NONE`) the first time
+/// time round -- which `Painter::set_mask` asserts against, since a mask
+/// that chains to itself is a clip loop. That was an abort the first time
 /// the composer's new scroll area was redrawn on the emulator; a targeted
 /// redraw of a `Masked` is what any real screen does whenever anything
 /// inside it changes.
@@ -161,7 +164,10 @@ fn redrawing_a_masked_widget_does_not_nest_its_own_mask() {
         ui: UiData::default(),
     };
     let (_scroll, inner_root, _rects) = scrolled_rects(&mut rsc, 8);
-    let masked = rsc.ui.widgets.add_strong(Masked { inner: inner_root });
+    let masked = rsc.ui.widgets.add_strong(Masked {
+        shape: None,
+        inner: inner_root,
+    });
     let masked_id = masked.id();
     let root = masked.any();
     let mut render = UiRenderState::new();
@@ -184,7 +190,10 @@ fn a_mask_stays_put_while_its_scrolled_content_moves() {
         ui: UiData::default(),
     };
     let (scroll, inner_root, _rects) = scrolled_rects(&mut rsc, 500);
-    let masked = rsc.ui.widgets.add_strong(Masked { inner: inner_root });
+    let masked = rsc.ui.widgets.add_strong(Masked {
+        shape: None,
+        inner: inner_root,
+    });
     let masked_id = masked.id();
     let root = masked.any();
     let mut render = UiRenderState::new();
@@ -318,7 +327,10 @@ fn a_scroll_measures_the_box_it_was_offered_not_the_window() {
         x: None,
         y: Some(Len::abs(1000.0)),
     });
-    let scroll = rsc.ui.widgets.add_strong(Scroll::new(tall.any(), Axis::Y));
+    let scroll = rsc
+        .ui
+        .widgets
+        .add_strong(Scroll::new(tall.any(), Axis::Y, true));
     let scroll_w = scroll.weak();
     let scroll_id = scroll.id();
     let capped = rsc.ui.widgets.add_strong(MaxSize {
@@ -381,7 +393,10 @@ fn a_panned_widgets_own_hit_box_moves_exactly_once() {
         y: Some(Len::abs(1000.0)),
     });
     let tall_w = tall.weak();
-    let scroll = rsc.ui.widgets.add_strong(Scroll::new(tall.any(), Axis::Y));
+    let scroll = rsc
+        .ui
+        .widgets
+        .add_strong(Scroll::new(tall.any(), Axis::Y, true));
     let scroll_w = scroll.weak();
     let root = scroll.any();
 
@@ -418,7 +433,10 @@ fn a_masked_widget_keeps_one_mask_slot_that_is_always_its_own_region() {
         ui: UiData::default(),
     };
     let (_scroll, inner_root, _rects) = scrolled_rects(&mut rsc, 8);
-    let masked = rsc.ui.widgets.add_strong(Masked { inner: inner_root });
+    let masked = rsc.ui.widgets.add_strong(Masked {
+        shape: None,
+        inner: inner_root,
+    });
     let masked_id = masked.id();
     // Placed at the bottom of a `Span::DOWN` behind a `rest(1)` sibling,
     // which is what moves the bar away from the provisional slot it is
@@ -453,7 +471,7 @@ fn a_masked_widget_keeps_one_mask_slot_that_is_always_its_own_region() {
     );
     let mask = *rsc.ui.masks.iter().next().unwrap();
     assert_eq!(
-        mask.region,
+        render.primitives.instance(mask.primitive).region,
         render.active.get(&masked_id).unwrap().region,
         "the mask a descendant clips against must be this widget's current box"
     );
@@ -655,3 +673,293 @@ fn a_widget_moved_by_its_parent_and_then_placed_inside_it_lands_at_the_placement
          after={after:?}"
     );
 }
+
+// ---------------------------------------------------------------------
+// LAYOUT.md's "Masks with a shape" -- its pass conditions, at layer 1.
+//
+// The shape a mask clips to is a *primitive already drawn*, never a copy
+// of one, so "the child's clipped corner" and "the container's own corner"
+// are the same arithmetic. These say so by evaluating both and demanding
+// exact equality: an approximate assertion would also pass a second copy
+// of the radius that merely happened to agree.
+// ---------------------------------------------------------------------
+
+const RADIUS: f32 = 20.0;
+
+/// A rounded container with `.masked_by` it, holding a `Rect::REST` child
+/// that fills it -- so the child's own corners are exactly the corners
+/// being clipped away. Returns the drawn state, the mask, the child, and
+/// the shape primitive the mask points at.
+fn rounded_container(rsc: &mut TestRsc) -> (UiRenderState, MaskIdx, WidgetId, u32) {
+    let child = rsc.ui.widgets.add_strong(Rect::new(UiColor::WHITE));
+    let child_id = child.id();
+    let shape = rsc
+        .ui
+        .widgets
+        .add_strong(Rect::new(UiColor::BLACK).radius(Len::abs(RADIUS)));
+    let shape_id = shape.id();
+    let root = rsc
+        .ui
+        .widgets
+        .add_strong(Masked {
+            shape: Some(shape.any()),
+            inner: child.any(),
+        })
+        .any();
+
+    let mut render = UiRenderState::new();
+    render.resize((200.0, 100.0));
+    render.update(&root, rsc);
+
+    let mask = render
+        .active
+        .get(&child_id)
+        .expect("the child is drawn")
+        .mask;
+    assert_ne!(
+        mask,
+        MaskIdx::NONE,
+        "the child was drawn with no clip at all"
+    );
+    let slot = render
+        .first_primitive(shape_id)
+        .expect("the shape widget drew a rect");
+    (render, mask, child_id, slot)
+}
+
+/// The pass condition: the child's coverage at a corner pixel *equals*
+/// the container's own coverage there. Exactly equal, because it is the
+/// same primitive evaluated once -- LAYOUT.md's point 1.
+#[test]
+fn a_masked_child_is_clipped_by_its_container_s_own_corner() {
+    let mut rsc = TestRsc {
+        ui: UiData::default(),
+    };
+    let (render, mask, _child, slot) = rounded_container(&mut rsc);
+    let corners = render.primitive_corners(slot, &rsc);
+    let radius = render
+        .primitives
+        .primitive_data::<RectPrimitive>(slot)
+        .expect("a mask's shape is a rect")
+        .radius;
+
+    // Across the whole corner arc, not one point on it: a single sample
+    // is satisfied by a mask that clips to the box and happens to agree
+    // where the two coincide. Swept from the arc's own centre -- the
+    // straight chord between the two ends of the arc lies *inside* the
+    // circle everywhere, so a walk along it never leaves the shape and
+    // the `outside` count below is what caught that.
+    let arc_center = corners.top_left + Vec2::new(radius, radius);
+    let (mut outside, mut inside) = (0, 0);
+    for i in 0..=20 {
+        let angle = std::f32::consts::FRAC_PI_2 * i as f32 / 20.0;
+        let dir = Vec2::new(-angle.cos(), -angle.sin());
+        for out in [-1.5f32, 0.0, 1.5] {
+            let pos = arc_center + dir * (radius + out);
+            let container = rounded_rect_coverage(pos, corners.top_left, corners.bot_right, radius);
+            assert_eq!(
+                render.mask_coverage(mask, pos, &rsc),
+                container,
+                "at {pos:?} the child's clip and the container's own edge disagree",
+            );
+            if container < 0.5 {
+                outside += 1;
+            } else {
+                inside += 1;
+            }
+        }
+    }
+    assert!(
+        outside > 0 && inside > 0,
+        "the sweep stayed on one side of the curve ({outside} out, {inside} in), so it proved \
+         nothing about the corner"
+    );
+}
+
+/// A hit test asks the same question the pixels do: the corner the
+/// container rounded away is not there to be pressed, and a point just
+/// inside the curve is. LAYOUT.md's point 4.
+#[test]
+fn a_mask_s_shape_decides_what_can_be_pressed() {
+    let mut rsc = TestRsc {
+        ui: UiData::default(),
+    };
+    let (render, mask, _child, slot) = rounded_container(&mut rsc);
+    let corners = render.primitive_corners(slot, &rsc);
+
+    // The very corner of the box, which the radius cut off.
+    let cut = corners.top_left + Vec2::new(1.0, 1.0);
+    assert!(
+        !render.mask_admits(mask, cut, &rsc),
+        "the corner the container rounded away is still pressable",
+    );
+    // The same distance in along the diagonal, past the curve.
+    let inside = corners.top_left + Vec2::new(RADIUS, RADIUS);
+    assert!(
+        render.mask_admits(mask, inside, &rsc),
+        "a point well inside the curve is not pressable",
+    );
+    // And the middle of an edge, which no radius touches -- the half the
+    // rounding had no reason to change.
+    let edge = Vec2::new(
+        (corners.top_left.x + corners.bot_right.x) / 2.0,
+        corners.top_left.y + 1.0,
+    );
+    assert!(
+        render.mask_admits(mask, edge, &rsc),
+        "a straight edge between two corners is not pressable",
+    );
+}
+
+/// Nested masks multiply, so a pixel inside two feathered corners is
+/// dimmed by both -- LAYOUT.md's point 2, and the "alpha should be
+/// decreased / multiplied" Iris asked for. Written as a product of the
+/// two the shader would compute separately, which is what "multiply"
+/// means and what an intersection test would get wrong.
+#[test]
+fn nested_masks_multiply_their_coverage() {
+    let mut rsc = TestRsc {
+        ui: UiData::default(),
+    };
+    let child = rsc.ui.widgets.add_strong(Rect::new(UiColor::WHITE));
+    let child_id = child.id();
+
+    let inner_shape = rsc
+        .ui
+        .widgets
+        .add_strong(Rect::new(UiColor::BLACK).radius(Len::abs(RADIUS)));
+    let inner_shape_id = inner_shape.id();
+    let inner = rsc.ui.widgets.add_strong(Masked {
+        shape: Some(inner_shape.any()),
+        inner: child.any(),
+    });
+    let outer_shape = rsc
+        .ui
+        .widgets
+        .add_strong(Rect::new(UiColor::BLACK).radius(Len::abs(RADIUS)));
+    let outer_shape_id = outer_shape.id();
+    let root = rsc
+        .ui
+        .widgets
+        .add_strong(Masked {
+            shape: Some(outer_shape.any()),
+            inner: inner.any(),
+        })
+        .any();
+
+    let mut render = UiRenderState::new();
+    render.resize((200.0, 100.0));
+    render.update(&root, &mut rsc);
+
+    let mask = render.active.get(&child_id).expect("drawn").mask;
+    let one = |render: &UiRenderState, rsc: &TestRsc, id, pos| {
+        let slot = render.first_primitive(id).expect("a shape rect");
+        let c = render.primitive_corners(slot, rsc);
+        let radius = render
+            .primitives
+            .primitive_data::<RectPrimitive>(slot)
+            .unwrap()
+            .radius;
+        rounded_rect_coverage(pos, c.top_left, c.bot_right, radius)
+    };
+
+    // A point on the corner arc, where both feathers are partial -- the
+    // only place a product and a minimum differ.
+    let slot = render.first_primitive(inner_shape_id).unwrap();
+    let corners = render.primitive_corners(slot, &rsc);
+    let pos = corners.top_left + Vec2::new(RADIUS * 0.3, RADIUS * 0.3);
+    let inner_cov = one(&render, &rsc, inner_shape_id, pos);
+    let outer_cov = one(&render, &rsc, outer_shape_id, pos);
+    assert!(
+        inner_cov > 0.0 && inner_cov < 1.0,
+        "the sample point is not inside a feather ({inner_cov}), so this proves nothing"
+    );
+    assert_eq!(
+        render.mask_coverage(mask, pos, &rsc),
+        inner_cov * outer_cov,
+        "two nested masks must multiply, not intersect",
+    );
+}
+
+/// A plain `.masked()` -- no shape given -- still clips to the widget's
+/// own box with square corners, which is what every list and scroll area
+/// relies on. The half the shape work had no reason to touch, and the one
+/// that would silently round every existing clip if `set_mask` ever wrote
+/// a radius of its own.
+#[test]
+fn a_plain_mask_still_clips_to_a_square_box() {
+    let mut rsc = TestRsc {
+        ui: UiData::default(),
+    };
+    let (_scroll, inner_root, _rects) = scrolled_rects(&mut rsc, 8);
+    let root = rsc
+        .ui
+        .widgets
+        .add_strong(Masked {
+            shape: None,
+            inner: inner_root,
+        })
+        .any();
+    let mut render = UiRenderState::new();
+    render.resize((200.0, 100.0));
+    render.update(&root, &mut rsc);
+
+    let mask = *rsc.ui.masks.iter().next().expect("one mask");
+    let corners = render.primitive_corners(mask.primitive, &rsc);
+    let mask_idx = MaskIdx::preset(0);
+    assert!(
+        render.mask_admits(mask_idx, corners.top_left + Vec2::new(0.5, 0.5), &rsc),
+        "a square clip must admit its own corner pixel",
+    );
+    assert!(
+        !render.mask_admits(mask_idx, corners.top_left - Vec2::new(2.0, 2.0), &rsc),
+        "a square clip must reject a point outside it",
+    );
+}
+
+/// A scroll area created to be *read* opens at the beginning of its
+/// content, however many frames it takes to learn how long that content
+/// is.
+///
+/// The bug this pins: `content_len` was `0.0` both for "nothing here" and
+/// for "not drawn yet", so the first frame's clamp found a range of zero,
+/// read `amt == len` as "sitting at the end", and set `snap_end` -- and
+/// the frame after, now knowing the real length, jumped to it. On screen
+/// that was a code fence opening at the end of its longest line, in the
+/// middle of a word (`iris/run-headless.sh phone`, 2026-09-08).
+#[test]
+fn a_scroll_area_opens_at_the_start_of_content_it_has_not_measured_yet() {
+    for (name, at_end, want) in [("read", false, 0.0), ("written", true, 4900.0)] {
+        let mut rsc = TestRsc {
+            ui: UiData::default(),
+        };
+        let fill = rsc.ui.widgets.add_strong(Rect::new(UiColor::WHITE)).any();
+        let tall = rsc.ui.widgets.add_strong(Sized {
+            inner: fill,
+            x: None,
+            y: Some(Len::abs(5000.0)),
+        });
+        let scroll = rsc
+            .ui
+            .widgets
+            .add_strong(Scroll::new(tall.any(), Axis::Y, at_end));
+        let weak = scroll.weak();
+        let root = scroll.any();
+
+        let mut render = UiRenderState::new();
+        render.resize((800.0, 100.0));
+        // Twice: the first draw is the one that measures the content, and
+        // the defect only showed on the second. The touch in between is
+        // what asks for that second draw -- an unchanged frame draws
+        // nothing at all, which is the point of the frame before it.
+        render.update(&root, &mut rsc);
+        let _ = rsc.ui.widgets.get_mut(&weak);
+        render.update(&root, &mut rsc);
+
+        let amt = rsc.ui.widgets.get(&weak).unwrap().amt();
+        assert!(
+            (amt - want).abs() < 0.01,
+            "an area to be {name} should have opened at {want}, got {amt}"
+        );
+    }
+}

iris/src/lib.rs

@@ -20,6 +20,7 @@ pub mod android;
 pub mod default;
 
 pub mod attr;
+pub mod diagnostics;
 pub mod event;
 pub mod harness;
 pub mod platform;

iris/src/sense_tests.rs

@@ -7,7 +7,7 @@
 //! impl need no GPU or window.
 
 use crate::prelude::*;
-use std::{cell::Cell, rc::Rc};
+use std::{cell::Cell, rc::Rc, time::Instant};
 
 struct SenseRsc {
     ui: UiData,
@@ -130,7 +130,7 @@ fn a_button_over_a_list_scrolls_the_list_and_still_clicks() {
 /// the press, in a gap, or off the loaded content entirely. Before pointer
 /// capture, `run_sensors`' hit test simply delivered nothing that frame,
 /// so a widget mid-drag never saw its release and never got a chance to
-/// start a fling. `UiRenderState::capture_pointer`/`DragGesture` fix this
+/// start a fling. `PointerRequests::capture`/`DragGesture` fix this
 /// by giving the drag's widget every frame regardless of where the
 /// pointer is, including the terminal `Drop` in place of `PressEnd`.
 #[test]
@@ -157,7 +157,7 @@ fn a_release_outside_every_hit_region_still_reaches_the_captured_widget() {
                     // would gate this on a `DragArbiter`/`DragGesture`
                     // decision, but this test only needs to exercise the
                     // capture-and-release mechanics themselves.
-                    ctx.data.render.capture_pointer(draggable_weak.id());
+                    ctx.data.pointer.capture(draggable_weak.id());
                     let _ = rsc;
                 }
                 CursorSense::Drop => dropped.set(true),
@@ -175,7 +175,7 @@ fn a_release_outside_every_hit_region_still_reaches_the_captured_widget() {
     press.buttons.left = ActivationState::Start;
     render.run_sensors(&mut rsc, &mut state, press, (100.0, 100.0).into());
     assert_eq!(
-        render.captured_pointer(),
+        pointer_input(&mut rsc).holder(),
         Some(draggable.id()),
         "the press should have taken capture"
     );
@@ -192,7 +192,7 @@ fn a_release_outside_every_hit_region_still_reaches_the_captured_widget() {
          the widget holding pointer capture"
     );
     assert_eq!(
-        render.captured_pointer(),
+        pointer_input(&mut rsc).holder(),
         None,
         "Drop must release the capture"
     );
@@ -236,7 +236,7 @@ fn capturing_one_widget_starves_every_other_widget_of_events() {
     let mut render = UiRenderState::new();
     render.resize((100.0, 100.0));
     render.update(&root, &mut rsc);
-    render.capture_pointer(a_weak.id());
+    pointer_input(&mut rsc).set_holder(Some(a_weak.id()));
 
     let mut state = ();
     let cursor = cursor_at((50.0, 50.0).into());
@@ -315,5 +315,304 @@ fn a_finger_drag_over_a_scroll_area_pans_it() {
 
     // And the gesture holds the pointer, so the rest of it reaches this
     // widget even once the finger leaves its box.
-    assert_eq!(render.captured_pointer(), Some(scroll.id()));
+    assert_eq!(pointer_input(&mut rsc).holder(), Some(scroll.id()));
+}
+
+/// docs/REVIEW-2026-09-07.md's D4. The first `MotionEvent` a view sees can
+/// be a `Move` -- the `Down` went to another view, or the view was attached
+/// mid-gesture -- and its batched samples are older than its own
+/// timestamp. Anchoring on that timestamp clamped every one of them onto
+/// the anchor, so the tracker saw three samples at one instant, the Lsq2
+/// fit went degenerate, and the flick read 0 px/s.
+#[test]
+fn the_first_events_batched_samples_are_dated_apart() {
+    const MS: i64 = 1_000_000;
+    let now = Instant::now();
+    // A 120Hz batch: three historical samples at 0/4/8ms and the event's
+    // own at 12ms.
+    let clock = PointerClock::anchored(now, 12 * MS, 0);
+
+    assert_eq!(
+        clock.at(12 * MS),
+        now,
+        "the event's own sample is the one that arrived now"
+    );
+    let batch = [clock.at(0), clock.at(4 * MS), clock.at(8 * MS)];
+    assert!(
+        batch[0] < batch[1] && batch[1] < batch[2] && batch[2] < now,
+        "the batch must keep the 4ms between its samples, got {:?}",
+        batch
+            .iter()
+            .map(|t| now.duration_since(*t))
+            .collect::<Vec<_>>()
+    );
+    assert_eq!(clock.ms_since_anchor(8 * MS), 8);
+}
+
+/// The same clock has to keep ordering *across* events: the sample it
+/// compares a new event's first sample against is the previous event's
+/// last one, never the anchor.
+#[test]
+fn the_clock_orders_samples_across_events() {
+    const MS: i64 = 1_000_000;
+    let mut clock = PointerClock::anchored(Instant::now(), 12 * MS, 0);
+    let first = clock.sample(12 * MS);
+    let second = clock.sample(28 * MS);
+    assert!(second > first);
+    assert_eq!(
+        second.duration_since(first),
+        std::time::Duration::from_millis(16)
+    );
+}
+
+/// Iris's 2026-09-08 phone report, first half: "it keeps snapping back to
+/// some position when horizontally scrolling."
+///
+/// A `Scroll` that has committed to a pan holds the pointer, so the
+/// gesture's end arrives as `CursorSense::Drop` -- and `scrollable_on`
+/// used to register `click_or_drag | unclick` only, which `should_run`
+/// never matches a `Drop` against. So the widget never learned its own
+/// gesture had ended: its `DragArbiter` stayed `Panning` at the position
+/// the finger left, and the *next* drag's first frame was measured from
+/// there and applied in one step. The registration is
+/// `CursorSense::drag_senses()` now, which is the rule for every widget
+/// driving a `DragGesture` rather than a fact about this one.
+#[test]
+fn a_scroll_area_that_captured_the_pointer_learns_its_gesture_ended() {
+    let mut rsc = SenseRsc {
+        ui: UiData::default(),
+        events: EventManager::default(),
+    };
+    let scroll_strong = rect(UiColor::WHITE)
+        .height(Len::abs(1000.0))
+        .scrollable()
+        .add_strong(&mut rsc);
+    let scroll = scroll_strong.weak();
+    let root = scroll_strong.any();
+
+    let mut render = UiRenderState::new();
+    render.resize((100.0, 100.0));
+    render.update(&root, &mut rsc);
+    rsc.ui.widgets.get_mut(&scroll).unwrap().scroll(0.0);
+    render.update(&root, &mut rsc);
+
+    let mut state = ();
+    let win = Vec2::new(100.0, 100.0);
+    let mut send = |render: &UiRenderState, rsc: &mut SenseRsc, y: f32, button| {
+        let mut c = cursor_at((50.0, y).into());
+        c.buttons.left = button;
+        render.run_sensors(rsc, &mut state, c, win);
+    };
+
+    // One pan of 40px past the slop, then a release well outside the
+    // widget -- the ordinary shape of a flick.
+    send(&render, &mut rsc, 80.0, ActivationState::Start);
+    send(
+        &render,
+        &mut rsc,
+        80.0 - (DRAG_SLOP + 40.0),
+        ActivationState::On,
+    );
+    let after_first = rsc.ui.widgets.get(&scroll).unwrap().amt();
+    assert!((after_first - 40.0).abs() < 0.01, "amt={after_first}");
+    send(&render, &mut rsc, 400.0, ActivationState::End);
+    assert_eq!(
+        pointer_input(&mut rsc).holder(),
+        None,
+        "the release must give the pointer back"
+    );
+
+    // A second gesture, starting where the first one did. If the arbiter
+    // were still panning from the release position, this first frame
+    // would apply the whole distance between the two at once.
+    send(&render, &mut rsc, 80.0, ActivationState::Start);
+    let after_second = rsc.ui.widgets.get(&scroll).unwrap().amt();
+    assert!(
+        (after_second - after_first).abs() < 0.01,
+        "a fresh touch-down moved the content by {} -- the previous \
+         gesture was never closed",
+        after_second - after_first,
+    );
+}
+
+/// The second half of the same report: "tapping sometimes seems to make
+/// the scrolling jump, particularly when tapping on things that have
+/// events like horizontal scrolling."
+///
+/// Two widgets see the same press -- a scroll area and, under it,
+/// something tracking the gesture for a list. When the scroll area
+/// captures, the other one is cut off completely: no `PressEnd`, no
+/// `Drop`. It has to be told, or its gesture stays open at an origin
+/// belonging to a finger that has long gone, and the next unrelated touch
+/// is measured from it.
+#[test]
+fn taking_the_pointer_cancels_everyone_else_tracking_the_press() {
+    let mut rsc = SenseRsc {
+        ui: UiData::default(),
+        events: EventManager::default(),
+    };
+
+    // The bystander *contains* the capturer, which is the real shape: a
+    // transcript's `List` and one row's own text both track the same
+    // press, and a `Stack`'s siblings would be on separate layers where
+    // only the topmost is dispatched to at all.
+    let capturer = rsc.ui.widgets.add_strong(Rect::new(UiColor::WHITE));
+    let capturer_weak = capturer.weak();
+    let bystander = rsc.ui.widgets.add_strong(Stack {
+        children: vec![capturer.any()],
+        size: StackSize::default(),
+    });
+    let bystander_weak = bystander.weak();
+
+    let capturer_saw = Rc::new(Cell::new(0u32));
+    {
+        let capturer_saw = capturer_saw.clone();
+        rsc.register_event(
+            capturer_weak,
+            CursorSense::drag_senses(),
+            move |ctx, _rsc| {
+                capturer_saw.set(capturer_saw.get() + 1);
+                if matches!(ctx.data.sense, CursorSense::Pressing(_)) {
+                    ctx.data.pointer.capture(capturer_weak.id());
+                }
+            },
+        );
+    }
+    let cancelled = Rc::new(Cell::new(0u32));
+    let ended = Rc::new(Cell::new(0u32));
+    {
+        let (cancelled, ended) = (cancelled.clone(), ended.clone());
+        rsc.register_event(
+            bystander_weak,
+            CursorSense::drag_senses(),
+            move |ctx, _rsc| match ctx.data.sense {
+                CursorSense::Cancel => cancelled.set(cancelled.get() + 1),
+                CursorSense::PressEnd(_) | CursorSense::Drop => ended.set(ended.get() + 1),
+                _ => {}
+            },
+        );
+    }
+
+    let root = bystander.any();
+    let mut render = UiRenderState::new();
+    render.resize((100.0, 100.0));
+    render.update(&root, &mut rsc);
+
+    let mut state = ();
+    let win = Vec2::new(100.0, 100.0);
+    let mut down = cursor_at((50.0, 50.0).into());
+    down.buttons.left = ActivationState::Start;
+    render.run_sensors(&mut rsc, &mut state, down, win);
+    assert_eq!(cancelled.get(), 0, "nothing has captured yet");
+
+    let mut moved = cursor_at((50.0, 20.0).into());
+    moved.buttons.left = ActivationState::On;
+    render.run_sensors(&mut rsc, &mut state, moved, win);
+    assert!(capturer_saw.get() > 0, "the capturer never saw the press");
+    assert_eq!(
+        pointer_input(&mut rsc).holder(),
+        Some(capturer_weak.id()),
+        "the capture should have been taken on this frame"
+    );
+    assert_eq!(
+        cancelled.get(),
+        1,
+        "the widget that lost the gesture must be told exactly once"
+    );
+
+    // And exactly once: the frames after the capture reach the capturer
+    // alone, so there is nothing left to cancel.
+    let mut more = cursor_at((50.0, 10.0).into());
+    more.buttons.left = ActivationState::On;
+    render.run_sensors(&mut rsc, &mut state, more, win);
+    let mut up = cursor_at((50.0, 10.0).into());
+    up.buttons.left = ActivationState::End;
+    render.run_sensors(&mut rsc, &mut state, up, win);
+    assert_eq!(cancelled.get(), 1, "cancelled more than once");
+    assert_eq!(
+        ended.get(),
+        0,
+        "a cancelled widget must not also be told the gesture ended \
+         normally -- acting on that is the tap it never made"
+    );
+}
+
+/// Iris's rule for nested scrolling, 2026-09-08: "it should only trigger
+/// horizontal if you drag left or right, and vertical should fall through
+/// if you drag up or down."
+///
+/// One mechanism does both, and it is `DragArbiter`'s existing axis test:
+/// each scroll area's gesture commits only on its own axis, so a drag
+/// along the other one is never claimed and the enclosing area's gesture
+/// -- which sees the same press, being an ancestor rather than a sibling
+/// layer -- is the one that commits and captures. This pins the pair,
+/// including the direction the change had no reason to touch.
+#[test]
+fn a_drag_pans_whichever_nested_scroll_area_owns_its_axis() {
+    for (name, to, pans, still) in [
+        ("vertical", Vec2::new(50.0, 80.0 - (DRAG_SLOP + 40.0)), 0, 1),
+        (
+            "horizontal",
+            Vec2::new(50.0 - (DRAG_SLOP + 40.0), 80.0),
+            1,
+            0,
+        ),
+    ] {
+        let mut rsc = SenseRsc {
+            ui: UiData::default(),
+            events: EventManager::default(),
+        };
+        // 1000px square of content in a 100px window: room to pan either
+        // way, in an X area inside a Y one.
+        let seen = Rc::new(Cell::new(None));
+        let record = seen.clone();
+        let outer_strong = rect(UiColor::WHITE)
+            .width(Len::abs(1000.0))
+            .height(Len::abs(1000.0))
+            .scrollable_on(Axis::X)
+            // The inner area's own handle, taken as the chain is built --
+            // the whole point is to exercise `scrollable_on`'s real
+            // registration on both, so neither is assembled by hand.
+            .with_id(move |_rsc, id| {
+                record.set(Some(id));
+                id
+            })
+            .scrollable()
+            .add_strong(&mut rsc);
+        let inner = seen.get().unwrap();
+        let outer = outer_strong.weak();
+        let root = outer_strong.any();
+        let areas = [outer, inner];
+
+        let mut render = UiRenderState::new();
+        render.resize((100.0, 100.0));
+        render.update(&root, &mut rsc);
+        // The second frame, where each area knows its content length --
+        // LAYOUT.md section 4's one-frame lag, and what drops `snap_end`.
+        for a in areas {
+            rsc.ui.widgets.get_mut(&a).unwrap().scroll(0.0);
+        }
+        render.update(&root, &mut rsc);
+
+        let mut state = ();
+        let win = Vec2::new(100.0, 100.0);
+        let mut down = cursor_at((50.0, 80.0).into());
+        down.buttons.left = ActivationState::Start;
+        render.run_sensors(&mut rsc, &mut state, down, win);
+        let mut drag = cursor_at(to);
+        drag.buttons.left = ActivationState::On;
+        render.run_sensors(&mut rsc, &mut state, drag, win);
+
+        let moved = rsc.ui.widgets.get(&areas[pans]).unwrap().amt();
+        let unmoved = rsc.ui.widgets.get(&areas[still]).unwrap().amt();
+        assert!(
+            (moved - 40.0).abs() < 0.01,
+            "a {name} drag should have panned the {name} area by the 40px \
+             past the slop, got {moved}"
+        );
+        assert_eq!(
+            unmoved, 0.0,
+            "a {name} drag must not move the area that owns the other axis"
+        );
+    }
 }

iris/src/widget/list.rs

@@ -94,11 +94,22 @@
 //! row that has never been measured, so this stays independent of how many
 //! rows exist outside the loaded window.
 //!
-//! **What is deliberately not solved here.** No overscroll clamping: a
-//! `scroll()` past the first or last row leaves a gap rather than rubber-
-//! banding back (mirrors `Scroll`'s own documented one-frame-lag
-//! tolerance in LAYOUT.md, just not even auto-corrected -- there is
-//! nothing to measure "how much content is left" without walking it).
+//! **Only what overlaps the viewport is drawn, and it is drawn whole.**
+//! One rule, `intersects_viewport`, used by both halves of that sentence:
+//! a row straddling either edge is drawn in full and clipped by the
+//! `.masked()` its caller must place it in (`List::draw` asserts that),
+//! and a row that has left the viewport is not drawn at all. The walk
+//! still traverses whatever lies between the anchor and the viewport, and
+//! `rehome_anchor` moves the anchor back onto a visible row every frame so
+//! that "whatever lies between" stays empty however far the list is
+//! panned.
+//!
+//! **Overscroll is taken back on the next frame, never rubber-banded.** A
+//! `scroll()` or a fling past the first or last row leaves a gap for one
+//! frame; `clamp_to_content` measures it from the ends the walk already
+//! placed and gives it back (the same one-frame-lag `Scroll`'s content
+//! length has, per LAYOUT.md). A list shorter than its viewport is not
+//! overscrolled and is left alone, still bottom-anchored.
 
 use crate::prelude::*;
 use iris_core::util::HashMap;
@@ -415,8 +426,9 @@ impl List {
 
     /// Move the anchor's edge by `amt` pixels. Positive moves later
     /// content into view (mirrors `Scroll::scroll`'s sign convention).
-    /// Deliberately unclamped -- see the module doc's "what is not
-    /// solved here."
+    /// Unclamped here, on purpose: it is one write, and there is nothing
+    /// at this point that knows where the content ends. The next `draw`
+    /// gives back whatever this moved past ([`Self::clamp_to_content`]).
     pub fn scroll(&mut self, amt: f32) {
         if let Some(a) = &mut self.anchor {
             a.offset -= amt;
@@ -475,9 +487,32 @@ impl List {
         // through) would propagate silently into `deceleration_for`'s
         // `.ln()` -- the fling either never settles or jumps to NaN
         // positions with nothing on screen saying why (docs/
-        // REVIEW-2026-09-06.md finding 3).
-        debug_assert!(velocity_px_per_s.is_finite());
-        if velocity_px_per_s == 0.0 || self.anchor.is_none() {
+        // REVIEW-2026-09-06.md finding 3). A plain `assert!` rather than a
+        // `debug_assert!`: it is one comparison per *gesture*, and every
+        // build anybody runs -- the emulator's and Iris's phone's -- is
+        // release, where a debug-only guard against silently wrong output
+        // is no guard at all (docs/REVIEW-2026-09-07.md's R1).
+        assert!(velocity_px_per_s.is_finite());
+        // Compose's two thresholds at a release, and **only** those two.
+        //
+        // The maximum is `ViewConfiguration.getScaledMaximumFlingVelocity()`
+        // (8000dp/s), which `DragGestureNode.sendDragStopped` passes into
+        // `VelocityTracker.calculateVelocity(maximumVelocity)`. It is
+        // applied here rather than in the tracker because the tracker
+        // works in pixels and has no density; this widget takes one from
+        // the painter in `draw`.
+        //
+        // The minimum is 1px/s, from `DefaultFlingBehavior.performFling`'s
+        // `abs(initialVelocity) > 1f` and its own stated reason ("we need
+        // it since spline curve gives us NaNs") -- not
+        // `ViewConfiguration.getScaledMinimumFlingVelocity()`'s 50dp/s,
+        // which Compose's scrolling never consults: its single use in
+        // either artifact is `NestedScrollInteropConnection`, for View
+        // interop. A 50dp/s floor would swallow slow, deliberate releases
+        // that Compose flings, so it is deliberately not here.
+        let max = MAX_FLING_VELOCITY_DP_S * self.density;
+        let velocity_px_per_s = velocity_px_per_s.clamp(-max, max);
+        if velocity_px_per_s.abs() <= 1.0 || self.anchor.is_none() {
             self.fling = None;
             return;
         }
@@ -541,14 +576,22 @@ impl List {
         // observable nor observed while `distance_fraction` returned `t`
         // (`android_fling_spline`'s doc), which is why this is here rather
         // than the total-travel line the release log already carries.
-        log::debug!(
-            "iris fling tick: t={:.3}s dy={:+.1}px speed={:.0}px/s of {:.0} left={:.1}px",
-            elapsed.as_secs_f32(),
-            delta,
-            f.calc.velocity_at(velocity, elapsed),
-            velocity,
-            f.calc.distance(velocity) - target,
-        );
+        // Gated on `iris::diagnostics::trace_enabled` since 2026-09-07
+        // (docs/RUST.md's review, D1): one line per fling *tick*,
+        // unconditional, was enough on its own to help fill the log
+        // ring -- see `android::view::IrisViewPeer::render`'s own doc for
+        // the same finding on its two per-frame lines.
+        if crate::diagnostics::trace_enabled() {
+            log::debug!(
+                target: "iris::frame",
+                "iris fling tick: t={:.3}s dy={:+.1}px speed={:.0}px/s of {:.0} left={:.1}px",
+                elapsed.as_secs_f32(),
+                delta,
+                f.calc.velocity_at(velocity, elapsed),
+                velocity,
+                f.calc.distance(velocity) - target,
+            );
+        }
         self.scroll(delta);
 
         // Clamp: a fling moving toward the start that has already reached
@@ -790,6 +833,115 @@ impl List {
         }
     }
 
+    /// Move the anchor onto a row that is actually on screen, without
+    /// moving anything that is drawn: the row it re-homes to keeps the
+    /// exact top edge this frame's layout gave it.
+    ///
+    /// [`Self::scroll`] moves the anchor's *offset* and nothing else, so
+    /// panning away from the anchor's own row leaves that row further and
+    /// further outside the viewport, and every row between it and the
+    /// viewport has to be walked on every frame from then on -- before
+    /// `place`'s intersection test, drawn too. Measured on the bench
+    /// fixture before this: 8 scrolls of 3000px left **64 rows** placed in
+    /// a 2012px viewport, ~59 of them off-screen, and the ones above it
+    /// drawn straight over the header (docs/IRIS_TODO.md, 2026-09-07).
+    /// Re-homing each frame makes the walk O(visible) again whatever
+    /// distance was travelled, which is what the module doc claims.
+    ///
+    /// Only when the anchor's own row has left the viewport, so
+    /// `update_snap_end`'s pinned-to-newest anchor -- last slot, bottom
+    /// edge at the viewport's own bottom, which intersects it -- is left
+    /// exactly as it is rather than rewritten into a top-edge anchor that
+    /// no longer reads as flush with the end.
+    fn rehome_anchor(&mut self) {
+        let Some(anchor) = self.anchor else {
+            return;
+        };
+        if self.extents.values().any(|e| e.slot == anchor.slot) {
+            return;
+        }
+        // The topmost row on screen, so the anchor's offset stays a small
+        // number near the viewport's own leading edge rather than
+        // whatever the last row's bottom happens to be.
+        let Some(first) = self
+            .extents
+            .values()
+            .min_by(|a, b| a.top.total_cmp(&b.top))
+            .copied()
+        else {
+            // Nothing on screen at all -- a list scrolled past its own
+            // content (`scroll` is deliberately unclamped). There is no
+            // on-screen row to re-home to, and inventing one would move
+            // the list; leave the anchor where it is and let the next
+            // scroll or `repair_anchor` bring content back.
+            return;
+        };
+        self.anchor = Some(Anchor {
+            slot: first.slot,
+            edge: Edge::Top,
+            offset: first.top,
+        });
+    }
+
+    /// Take back an empty band at one edge that content on the other side
+    /// of the viewport could fill -- the correction that makes a `scroll`
+    /// or a fling past the end of the content settle *on* the end rather
+    /// than beyond it.
+    ///
+    /// `top`/`bottom` are the extreme edges this frame's walk actually
+    /// placed, so the gap is already measured: `at_start` means nothing is
+    /// above `top`, and if `top` is nevertheless below the viewport's own
+    /// leading edge then those pixels are empty and always will be. This
+    /// is the whole of what the module doc used to list as deliberately
+    /// unsolved ("no overscroll clamping ... nothing to measure how much
+    /// content is left without walking it") -- true of *total* content
+    /// height, but the walk hands back both ends of the loaded run for
+    /// free, which is all a clamp needs. `tick_fling` stops a fling that
+    /// has reached an end, but stops it wherever the spline's last step
+    /// had already put it: a hard fling to the top of the bench fixture
+    /// left the first row **1398px below** a 600px viewport, i.e. the
+    /// whole screen blank, and it stayed there (docs/IRIS_TODO.md,
+    /// 2026-09-07: "black from the header down").
+    ///
+    /// **Only when the opposite end is not also inside the viewport.**
+    /// Both at once means the content is shorter than the viewport, where
+    /// the space is not overscroll at all -- it is a bottom-anchored list
+    /// with three rows in it, and pulling those to the top would be this
+    /// widget rejecting its own default (`repair_anchor`).
+    ///
+    /// Applied to the anchor, so it lands on the *next* frame rather than
+    /// re-running this one: the same one-frame-lag `Scroll` accepts for
+    /// its content length, and one frame is 8ms on the phone.
+    fn clamp_to_content(&mut self, painter: &mut Painter, top: f32, bottom: f32) {
+        if self.at_start == self.at_end {
+            return;
+        }
+        // `at_start`/`at_end` already carry the sign of their own gap
+        // (`top >= 0.0`, `bottom <= viewport_len`), so this is the gap
+        // itself, positive to move content toward the leading edge.
+        let gap = if self.at_start {
+            top
+        } else {
+            bottom - self.viewport_len
+        };
+        // Sub-pixel gaps are what floating-point row heights leave behind
+        // every frame; correcting one would ask for another frame, which
+        // would leave another, and the list would never stop redrawing.
+        if gap.abs() < 0.5 {
+            return;
+        }
+        self.scroll(gap);
+        // Nothing else will ask: the frame this correction was discovered
+        // in has already been laid out, and a fling that ran out at an end
+        // (`tick_fling`'s `hit_bound`) has stopped requesting frames --
+        // which is exactly the case that left the list parked past its own
+        // first row.
+        painter.draw_again();
+        if let Some(redraw) = &self.redraw {
+            redraw.request_redraw();
+        }
+    }
+
     fn update_snap_end(&mut self) {
         self.snap_end = match self.anchor {
             Some(a) => {
@@ -801,6 +953,22 @@ impl List {
         };
     }
 
+    /// **The one rule for what this list draws**: a row is on screen if
+    /// any part of it is, so a row straddling either edge is drawn *in
+    /// full* and one that has left the viewport entirely is not drawn at
+    /// all. Both halves matter and they failed in opposite directions on
+    /// Iris's phone (docs/IRIS_TODO.md, 2026-09-07): rows already scrolled
+    /// past were still being drawn, over the header above the list, and
+    /// the part of a straddling row above the viewport had nothing
+    /// clipping it. The viewport here is the list's own box -- `0 ..
+    /// viewport_len`, `painter.region()` in window terms -- which is the
+    /// same box `List::draw` requires a mask on, so that what this test
+    /// admits and what the clip keeps are one region rather than two that
+    /// can disagree.
+    fn intersects_viewport(&self, top: f32, bottom: f32) -> bool {
+        bottom > 0.0 && top < self.viewport_len
+    }
+
     fn abs_region(axis: Axis, start: f32, end: f32) -> UiRegion {
         let span = UiSpan::new(UiScalar::abs(start), UiScalar::abs(end));
         UiRegion::from_axis(axis, span, UiSpan::FULL)
@@ -847,7 +1015,10 @@ impl List {
         // sites, not by this function, which would otherwise fail with a
         // bare "index out of bounds" and no context (docs/
         // REVIEW-2026-09-06.md finding 2). `slot_widget`, called from
-        // here, is what actually indexes/`.expect`s on it.
+        // here, is what actually indexes/`.expect`s on it. Stays a
+        // `debug_assert!` under R1's rule: this runs once per row placed
+        // per frame, and its release failure is the `.expect` below rather
+        // than something silently wrong on screen.
         debug_assert!(
             self.slot_exists(slot),
             "place() called with a slot that doesn't exist: {slot:?}"
@@ -865,6 +1036,26 @@ impl List {
         let key = self.slot_key(slot);
         let cached = key.and_then(|k| self.heights.get(&k).copied());
 
+        // A row entirely outside the viewport is traversed but not drawn
+        // -- see `intersects_viewport`. The walk still has to *pass
+        // through* it, because its height is what says where the rows
+        // behind it land, but nothing about it reaches the screen, so
+        // drawing it costs a redraw (and, unclipped, paints over whatever
+        // is above the list) for content nobody can see. Only possible
+        // for a row whose height is already known: a first-time row has
+        // to be drawn to be measured at all, which is why the extent
+        // below is recorded from the intersection test rather than from
+        // "was this drawn".
+        if let Some(h) = cached {
+            let (top, bottom) = match placement {
+                Placement::Top(top) => (top, top + h),
+                Placement::Bottom(bottom) => (bottom - h, bottom),
+            };
+            if !self.intersects_viewport(top, bottom) {
+                return (top, bottom);
+            }
+        }
+
         let (top, bottom, height) = match (placement, cached) {
             (Placement::Top(top), Some(h)) => {
                 // Offered a box sized to the *cached* height (cheap to
@@ -928,7 +1119,13 @@ impl List {
         };
         if let Some(k) = key {
             self.heights.insert(k, height);
-            self.extents.insert(k, RowExtent { slot, top, bottom });
+            // `extents` is what is *on screen* (`key_at`'s doc, and
+            // `rehome_anchor` below reads it as exactly that), so a
+            // first-time row that had to be drawn to be measured and
+            // turned out to be off-screen does not go in it.
+            if self.intersects_viewport(top, bottom) {
+                self.extents.insert(k, RowExtent { slot, top, bottom });
+            }
         }
         (top, bottom)
     }
@@ -961,6 +1158,26 @@ impl Widget for List {
         // density, and `draw` is where this widget meets the only thing
         // that knows it. See `fling`.
         self.density = painter.density();
+        // A row that straddles either edge is drawn in full
+        // (`intersects_viewport`), so the part of it outside this list's
+        // box is on screen unless something clips it -- and with nothing
+        // clipping it, a transcript panned to its top edge drew code and
+        // paragraphs straight through the header bar above it on Iris's
+        // phone (docs/IRIS_TODO.md, 2026-09-07). Clipping is `.masked()`,
+        // one mechanism, applied by whoever places the list -- a `List`
+        // cannot set the mask itself, since `Painter::set_mask` allows one
+        // mask per widget and rows of this list already use their own
+        // (`transcript-ui`'s `row.rs`, `tool.rs`). So it checks instead.
+        //
+        // `assert!`, not `debug_assert!`: one bool per draw, and what it
+        // catches is a `List` painting over its surroundings with nothing
+        // on screen saying so -- the fault e922b73 was written to fix.
+        // Every build that runs is release (docs/REVIEW-2026-09-07.md's R1).
+        assert!(
+            painter.is_masked(),
+            "a `List` must be drawn inside something `.masked()`: it draws rows straddling both \
+             edges in full, so the parts outside its own box reach the screen otherwise",
+        );
         let output_len = painter.output_size().axis(axis);
         self.viewport_len = painter.region().axis(axis).len().to_abs(output_len);
 
@@ -1013,6 +1230,26 @@ impl Widget for List {
         self.at_start = self.prev_slot(idx_top).is_none() && top >= 0.0;
         self.at_end = self.next_slot(idx_bottom).is_none() && bottom <= self.viewport_len;
 
+        // Both halves of `intersects_viewport`'s rule, checked where they
+        // are cheap to check: what this frame put on screen is exactly
+        // what overlaps the viewport, and nothing above or below it can
+        // be seen. The first failed silently for a whole build -- an
+        // off-screen row draws correctly, it is just in the wrong place.
+        // `assert!` for R1's reason: it walks the rows *on screen*, a
+        // handful, once per draw, and a release build is the only build
+        // this fault has ever been seen in.
+        assert!(
+            self.extents
+                .values()
+                .all(|e| self.intersects_viewport(e.top, e.bottom)),
+            "a row outside the viewport (0..{}) is recorded as on screen: {:?}",
+            self.viewport_len,
+            self.extents
+                .values()
+                .find(|e| !self.intersects_viewport(e.top, e.bottom)),
+        );
+        self.rehome_anchor();
+        self.clamp_to_content(painter, top, bottom);
         self.update_snap_end();
         Size::REST
     }
@@ -1068,9 +1305,60 @@ mod tests {
     /// calling `List`'s own methods through `Widgets::get`/`get_mut`, which
     /// need a `Sized` widget type) and the erased root `UiRenderState::update`
     /// draws.
+    ///
+    /// The root is a `Masked` around the list rather than the list
+    /// itself, because that is what every real caller has to do -- a
+    /// `List` draws the row straddling each edge in full and asserts
+    /// something is clipping it (`List::draw`). The mask is the full
+    /// window here, which is also the list's own box.
     fn add_list(rsc: &mut TestRsc, list: List) -> (WeakWidget<List>, StrongWidget) {
         let strong = rsc.ui.widgets.add_strong(list);
-        (strong.weak(), strong.any())
+        let weak = strong.weak();
+        let root = rsc.ui.widgets.add_strong(Masked {
+            shape: None,
+            inner: strong.any(),
+        });
+        (weak, root.any())
+    }
+
+    /// The case the top-edge cull and the overscroll clamp both had no
+    /// reason to touch: fewer rows than fit. Every one of them is drawn
+    /// (nothing here is outside the viewport), and `clamp_to_content`
+    /// leaves the list bottom-anchored -- the gap above the first row is
+    /// not overscroll, it is where this widget puts a short list, and
+    /// pulling it to the top would be the clamp overriding
+    /// `repair_anchor`'s own default.
+    #[test]
+    fn a_list_shorter_than_the_viewport_is_drawn_whole_and_stays_at_the_bottom() {
+        let mut rsc = TestRsc {
+            ui: UiData::default(),
+        };
+        let mut list = List::new(Axis::Y);
+        push_rows(&mut rsc, &mut list, &[0, 1, 2], 20.0);
+        let (list_weak, root) = add_list(&mut rsc, list);
+
+        let mut render = UiRenderState::new();
+        render.resize((100.0, 100.0));
+        // Several frames, since the clamp acts on the frame *after* the
+        // one that measured a gap: a wrong one would walk the rows up the
+        // screen 40px at a time rather than settle.
+        for _ in 0..4 {
+            render.update(&root, &mut rsc);
+            let list_ref = rsc.ui.widgets.get(&list_weak).unwrap();
+            assert_eq!(
+                list_ref.extents.len(),
+                3,
+                "every row of a short list is on screen"
+            );
+            let first = list_ref.extents[&0];
+            let last = list_ref.extents[&2];
+            assert!(
+                (first.top - 40.0).abs() < 0.01 && (last.bottom - 100.0).abs() < 0.01,
+                "a 60px list in a 100px viewport moved off the bottom: rows {}..{}",
+                first.top,
+                last.bottom,
+            );
+        }
     }
 
     #[test]
@@ -1142,7 +1430,7 @@ mod tests {
             bg_ids.push(bg_id);
             list.push_back(ListRow::new(key, row));
         }
-        let root = rsc.ui.widgets.add_strong(list).any();
+        let (_, root) = add_list(&mut rsc, list);
 
         let mut render = UiRenderState::new();
         render.resize((100.0, 100.0));
@@ -1323,7 +1611,12 @@ mod tests {
             render.update(&root, &mut rsc);
             render.take_counters();
 
-            rsc.ui.widgets.get_mut(&list_weak).unwrap().scroll(5.0);
+            // Backwards, into content that exists: a list opens flush with
+            // its newest end, so scrolling *forward* from there is
+            // overscroll, and `clamp_to_content` lays out a second time to
+            // give it back -- a correct extra pass, but not the ordinary
+            // scroll tick whose cost this test is about.
+            rsc.ui.widgets.get_mut(&list_weak).unwrap().scroll(-5.0);
             render.update(&root, &mut rsc);
             let (draws, _rewrites, moves, _shapes) = render.take_counters();
 
@@ -1616,9 +1909,28 @@ mod tests {
     /// Enough rows, tall enough, that a fling toward the start has real
     /// room to travel before `at_start` clamps it -- shared by the fling
     /// tests below.
+    /// How far a `build_flingable_list` list has scrolled from its very
+    /// first row, in pixels: read off the topmost row on screen, whose
+    /// content position is exactly `slot * ROW_H` because every row there
+    /// is that tall. Measures the list's own accumulated movement (the
+    /// thing `scroll`/`tick_fling` write) rather than the spline's
+    /// bookkeeping, and unlike a single row's extent it stays defined
+    /// however far the list travels -- `extents` holds only what is
+    /// on screen (`List::intersects_viewport`).
+    fn scroll_position(list: &List) -> f32 {
+        let top = list
+            .extents
+            .values()
+            .min_by(|a, b| a.top.total_cmp(&b.top))
+            .expect("something is on screen");
+        top.slot as f32 * FLING_ROW_H - top.top
+    }
+
+    const FLING_ROW_H: f32 = 20.0;
+
     fn build_flingable_list(rsc: &mut TestRsc) -> (WeakWidget<List>, StrongWidget, UiRenderState) {
         let mut list = List::new(Axis::Y);
-        push_rows(rsc, &mut list, &(0..200).collect::<Vec<_>>(), 20.0);
+        push_rows(rsc, &mut list, &(0..200).collect::<Vec<_>>(), FLING_ROW_H);
         let (list_weak, root) = add_list(rsc, list);
         let mut render = UiRenderState::new();
         render.resize((100.0, 600.0));
@@ -1764,32 +2076,22 @@ mod tests {
 
         rsc.ui.widgets.get_mut(&list_weak).unwrap().fling(8000.0);
         let start = Instant::now();
-        let mut prev_top = rsc.ui.widgets.get(&list_weak).unwrap().extents[&0].top;
+        let mut prev = scroll_position(rsc.ui.widgets.get(&list_weak).unwrap());
         let mut deltas = Vec::new();
         for step in 1..600 {
             let now = start + std::time::Duration::from_millis(step * 16);
             let still = rsc.ui.widgets.get_mut(&list_weak).unwrap().tick_fling(now);
             render.update(&root, &mut rsc);
-            let Some(top) = rsc
-                .ui
-                .widgets
-                .get(&list_weak)
-                .unwrap()
-                .extents
-                .get(&0)
-                .map(|e| e.top)
-            else {
-                break; // row 0 scrolled out of the loaded extents
-            };
-            deltas.push((prev_top - top).abs());
-            prev_top = top;
+            let at = scroll_position(rsc.ui.widgets.get(&list_weak).unwrap());
+            deltas.push((at - prev).abs());
+            prev = at;
             if !still {
                 break;
             }
         }
         assert!(
             deltas.len() >= 3,
-            "fling settled or left row 0's extent before collecting enough samples"
+            "fling settled before collecting enough samples"
         );
         // Skip the first tick (the slop-transition jump the arbiter
         // applies is a `List::fling`-adjacent concern, not this curve,
@@ -1871,15 +2173,26 @@ mod tests {
                 break;
             }
         }
+        // The frame that gives back whatever the fling's last step spent
+        // past the first row -- `clamp_to_content` writes the anchor at
+        // the end of a draw, so it lands on the next one.
+        render.update(&root, &mut rsc);
         let list_ref = rsc.ui.widgets.get(&list_weak).unwrap();
         assert!(
             list_ref.at_start,
             "fling should have clamped at the first row"
         );
+        // Not `>= -0.5`: `extents` used to hold every row the walk placed,
+        // on screen or not, so that read was satisfied by a first row
+        // sitting *1398px below* a 600px viewport with the whole screen
+        // blank -- the assertion could not fail in the direction the bug
+        // actually went. Both edges, so neither an overshoot past the top
+        // nor one left uncorrected can pass.
         let first = list_ref.extents[&0];
         assert!(
-            first.top >= -0.5,
-            "clamped fling overshot the first row's top: {}",
+            first.top.abs() < 0.5,
+            "a fling stopped at the start must leave the first row flush with the top, not {}px \
+             from it",
             first.top
         );
     }

iris/src/widget/mask.rs

@@ -1,12 +1,36 @@
 use crate::prelude::*;
 
+/// Clips `inner` -- and everything below it -- to a shape.
+///
+/// The shape is a **primitive**, never a rectangle or a radius stored
+/// here: with `shape`, the widget named there is drawn behind `inner`
+/// filling the same box and the clip is its first primitive, so a rounded
+/// container's corner and the corner its content is cut to are the same
+/// arithmetic and cannot fall out of step. Without one, this writes an
+/// undrawn rect at its own region, which is the plain "clip to my box"
+/// every list and scroll area wants. See docs/LAYOUT.md's "Masks with a
+/// shape".
 pub struct Masked {
+    /// The widget whose first primitive is the clip, drawn behind
+    /// `inner`, or `None` for this widget's own box.
+    pub shape: Option<StrongWidget>,
     pub inner: StrongWidget,
 }
 
 impl Widget for Masked {
     fn draw(&mut self, painter: &mut Painter) -> Size {
-        painter.set_mask(painter.region());
+        match &self.shape {
+            // Layered the way `Stack` layers a background under its
+            // content, and for the same reason: within one layer the draw
+            // order is undefined once anything has been freed.
+            Some(shape) => {
+                painter.child_layer();
+                painter.widget(shape);
+                painter.set_mask_to_widget(shape);
+                painter.next_layer();
+            }
+            None => painter.set_mask(painter.region()),
+        }
         painter.widget(&self.inner)
     }
 }

iris/src/widget/position/scroll.rs

@@ -1,5 +1,5 @@
 use crate::prelude::*;
-use crate::sense::{DragGesture, GestureOutcome};
+use crate::sense::{DragGesture, GestureOutcome, PointerRequests, PressState};
 use std::time::Instant;
 
 pub struct Scroll {
@@ -8,7 +8,18 @@ pub struct Scroll {
     amt: f32,
     snap_end: bool,
     container_len: f32,
-    content_len: f32,
+    /// How long the content is along `axis`, as of the last draw --
+    /// `None` until this widget has drawn once.
+    ///
+    /// An `Option` rather than a `0.0` that stands in for both, because
+    /// the two answers led somewhere different and the code could not
+    /// tell them apart: on the first frame the clamp in `update_amt`
+    /// computed a scroll range of zero, concluded from `amt == len` that
+    /// the area was sitting at its end, and set `snap_end` -- so the next
+    /// frame, now knowing the real length, jumped to it. A code fence
+    /// therefore opened at the end of its longest line, mid-word
+    /// (`iris/run-headless.sh phone`, 2026-09-08).
+    content_len: Option<f32>,
     /// Touch panning, from the same `DragGesture` `List` is driven by
     /// (`transcript-ui::Selection::drag`) rather than a second copy of its
     /// wiring: arbitration, `DRAG_SLOP` and pointer capture all live in
@@ -48,13 +59,19 @@ impl Widget for Scroll {
         let container_len = painter.px_size().axis(axis);
         self.container_len = container_len;
 
-        if self.snap_end {
-            self.amt = self.content_len - self.container_len;
+        if self.snap_end
+            && let Some(content_len) = self.content_len
+        {
+            self.amt = content_len - self.container_len;
         }
         self.update_amt();
 
         let mut region = UiRegion::FULL;
-        region.axis_mut(axis).end = region.axis(axis).start.offset(self.content_len);
+        // The container's own length until the content has been measured:
+        // a zero-length region on the first frame would place the child's
+        // primitives against a box of no size.
+        let offered = self.content_len.unwrap_or(container_len);
+        region.axis_mut(axis).end = region.axis(axis).start.offset(offered);
         let region = region.offset(Vec2::from_axis(axis, -self.amt, 0.0));
 
         let used = painter.widget_within(&self.inner, region);
@@ -62,10 +79,11 @@ impl Widget for Scroll {
         // A child reporting `rel` means "this fraction of what I was
         // offered", and what it was offered is this scroll area -- so the
         // container, again, is what that resolves against.
-        self.content_len = used
-            .axis(axis)
-            .apply_rest(painter.density())
-            .to_abs(container_len);
+        self.content_len = Some(
+            used.axis(axis)
+                .apply_rest(painter.density())
+                .to_abs(container_len),
+        );
 
         // The **content's** size, not the container's. A parent that can
         // grow (the composer's bar) should hug the text until its own cap
@@ -80,14 +98,18 @@ impl Widget for Scroll {
 }
 
 impl Scroll {
-    pub fn new(inner: StrongWidget, axis: Axis) -> Self {
+    /// `at_end` starts the area pinned to the end of its content and
+    /// keeps it there while the content grows -- see
+    /// `WidgetLike::scrollable_to_end`. `false` starts at the beginning,
+    /// which is what anything being *read* wants.
+    pub fn new(inner: StrongWidget, axis: Axis, at_end: bool) -> Self {
         Self {
             inner,
             axis,
             amt: 0.0,
-            snap_end: true,
+            snap_end: at_end,
             container_len: 0.0,
-            content_len: 0.0,
+            content_len: None,
             gesture: DragGesture::on(axis),
         }
     }
@@ -113,20 +135,22 @@ impl Scroll {
     /// released velocity is deliberately dropped rather than approximated.
     pub fn drag(
         &mut self,
-        render: &UiRenderState,
+        pointer: &PointerRequests,
         id: WidgetId,
         sense: CursorSense,
         pos_window: Vec2,
         now: Instant,
     ) {
-        // `already_selected: false` -- a scroll area has no selection of
-        // its own to extend, so a horizontal drag stays `Undecided` and a
+        // A default `PressState`: a scroll area has no selection of its
+        // own to extend, so a horizontal drag stays `Undecided` and a
         // vertical one past the slop pans, which is the whole contract
-        // here. A caller that *does* own a selection (the transcript's
-        // `Selection`) drives `DragGesture` itself instead.
+        // here; and it never flings (see this method's doc), so there is
+        // never a moving target to catch either. A caller that *does* own
+        // a selection, or a fling (the transcript's `Selection`), drives
+        // `DragGesture` itself instead.
         match self
             .gesture
-            .handle(render, id, sense, pos_window, now, false)
+            .handle(pointer, id, sense, pos_window, now, PressState::default())
         {
             // `scroll(dy)`, not `scroll(-dy)` -- `Selection::drag` passes
             // `-dy` to `List::scroll` because a `List`'s anchor offset and
@@ -141,14 +165,24 @@ impl Scroll {
             | GestureOutcome::Tapped
             | GestureOutcome::SelectStart
             | GestureOutcome::SelectExtend
+            | GestureOutcome::Cancelled
             | GestureOutcome::Released(_) => {}
         }
     }
 
+    /// Clamp `amt` into the range the content allows, and re-read whether
+    /// this area is sitting at its end.
+    ///
+    /// Both are skipped until the content has been measured: with no
+    /// length there is no range to clamp into, and "at the end" is a
+    /// question that cannot be answered yet -- answering it anyway is
+    /// what `content_len`'s doc describes.
     pub fn update_amt(&mut self) {
-        self.amt = self.amt.max(0.0);
-        let len = (self.content_len - self.container_len).max(0.0);
-        self.amt = self.amt.min(len);
+        let Some(content_len) = self.content_len else {
+            return;
+        };
+        let len = (content_len - self.container_len).max(0.0);
+        self.amt = self.amt.clamp(0.0, len);
         self.snap_end = self.amt == len;
     }
 
@@ -179,8 +213,8 @@ mod tests {
         let mut ui = UiData::default();
         let inner = ui.widgets.add_strong(Rect::new(UiColor::WHITE)).any();
         let id = inner.id();
-        let mut s = Scroll::new(inner, Axis::Y);
-        s.content_len = 1000.0;
+        let mut s = Scroll::new(inner, Axis::Y, true);
+        s.content_len = Some(1000.0);
         s.container_len = 100.0;
         s.amt = 400.0;
         s.snap_end = false;
@@ -189,7 +223,7 @@ mod tests {
 
     fn press(
         s: &mut Scroll,
-        render: &UiRenderState,
+        render: &PointerRequests,
         id: WidgetId,
         sense: CursorSense,
         y: f32,
@@ -201,7 +235,7 @@ mod tests {
     #[test]
     fn a_vertical_finger_drag_pans_the_content_with_the_finger() {
         let (_ui, mut s, id) = area();
-        let render = UiRenderState::new();
+        let render = PointerRequests::default();
         let t = Instant::now();
         press(
             &mut s,
@@ -244,7 +278,7 @@ mod tests {
     #[test]
     fn a_press_that_stays_inside_the_slop_does_not_scroll() {
         let (_ui, mut s, id) = area();
-        let render = UiRenderState::new();
+        let render = PointerRequests::default();
         let t = Instant::now();
         press(
             &mut s,
@@ -284,7 +318,7 @@ mod tests {
     #[test]
     fn a_horizontal_drag_does_not_scroll() {
         let (_ui, mut s, id) = area();
-        let render = UiRenderState::new();
+        let render = PointerRequests::default();
         let t = Instant::now();
         s.drag(
             &render,
@@ -309,7 +343,7 @@ mod tests {
     #[test]
     fn a_pan_past_the_end_clamps_instead_of_running_off() {
         let (_ui, mut s, id) = area();
-        let render = UiRenderState::new();
+        let render = PointerRequests::default();
         let t = Instant::now();
         s.drag(
             &render,

iris/src/widget/text/mod.rs

@@ -78,12 +78,20 @@ impl TextView {
         }
         self.width = width;
         let tex = painter.render_text(&mut self.buf, &self.attrs, width);
-        log::debug!(
-            "iris text render: chars={} width={width:?} glyphs={} size={:?}",
-            self.buf.text().chars().count(),
-            tex.glyphs.len(),
-            tex.size,
-        );
+        // Gated on `iris::diagnostics::trace_enabled` since 2026-09-07
+        // (docs/RUST.md's review, D1): one line per text *shape* (a cache
+        // miss), unconditional, is many per frame while rows compose --
+        // see `android::view::IrisViewPeer::render`'s own doc for the same
+        // finding on its two per-frame lines.
+        if crate::diagnostics::trace_enabled() {
+            log::debug!(
+                target: "iris::frame",
+                "iris text render: chars={} width={width:?} glyphs={} size={:?}",
+                self.buf.text().chars().count(),
+                tex.glyphs.len(),
+                tex.size,
+            );
+        }
         self.tex = Some(tex.clone());
         self.attrs.changed = false;
         self.buf.changed = false;

iris/src/widget/trait_fns.rs

@@ -87,14 +87,34 @@ widget_trait! {
         self.scrollable_on(Axis::Y)
     }
 
-    // `scrollable` along `axis`. A code fence pans across its own long
-    // lines exactly the way a transcript pans down its rows, so the two
-    // are one function with the axis passed in rather than a second copy
-    // -- `DragArbiter::on` is the other half. (A `///` doc comment here
-    // is not accepted by `widget_trait!`, which parses its body itself.)
+    /// `scrollable_on`, but starting pinned to the **end** of its content
+    /// and staying there while the content grows -- what a composer wants,
+    /// where the newest line is the one being written.
+    ///
+    /// Explicit, because the other kind is not a variation on it: a code
+    /// fence opened at the end of its longest line, which is the middle
+    /// of a word (seen in `iris/run-headless.sh phone`, 2026-09-08). The
+    /// two behaviours are one mechanism with the starting edge passed in,
+    /// and both names say which they are rather than one of them being a
+    /// default nobody reads.
+    fn scrollable_to_end(self, axis: Axis) -> impl WidgetIdFn<Rsc, Scroll> where Rsc: HasEvents {
+        self.scroll_area(axis, true)
+    }
+
+    /// `scrollable` along `axis`. A code fence pans across its own long
+    /// lines exactly the way a transcript pans down its rows, so the two
+    /// are one function with the axis passed in rather than a second copy
+    /// -- `DragArbiter::on` is the other half.
     fn scrollable_on(self, axis: Axis) -> impl WidgetIdFn<Rsc, Scroll> where Rsc: HasEvents {
+        self.scroll_area(axis, false)
+    }
+
+    /// The one implementation behind [`Self::scrollable_on`] and
+    /// [`Self::scrollable_to_end`] -- see the latter for what `at_end`
+    /// decides.
+    fn scroll_area(self, axis: Axis, at_end: bool) -> impl WidgetIdFn<Rsc, Scroll> where Rsc: HasEvents {
         move |state| {
-            Scroll::new(self.add_strong(state), axis)
+            Scroll::new(self.add_strong(state), axis, at_end)
                 .on(CursorSense::Scroll, move |ctx, rsc| {
                     let delta = ctx.data.scroll_delta.axis(axis) * 50.0;
                     ctx.widget(rsc).scroll(delta);
@@ -104,21 +124,32 @@ widget_trait! {
                 // has the arbitration and why there is no fling. The wheel
                 // above and this are the two inputs of one scroll, so they
                 // are registered together rather than left to each caller.
-                .on(
-                    CursorSense::click_or_drag() | CursorSense::unclick(),
-                    |ctx, rsc| {
-                        let id = ctx.widget.id();
-                        let (sense, pos) = (ctx.data.sense, ctx.data.cursor.pos);
-                        ctx.widget(rsc)
-                            .drag(ctx.data.render, id, sense, pos, ctx.data.cursor.time);
-                    },
-                )
+                .on(CursorSense::drag_senses(), |ctx, rsc| {
+                    let id = ctx.widget.id();
+                    let (sense, pos) = (ctx.data.sense, ctx.data.cursor.pos);
+                    ctx.widget(rsc)
+                        .drag(ctx.data.pointer, id, sense, pos, ctx.data.cursor.time);
+                })
                 .add(state)
         }
     }
 
     fn masked(self) -> impl WidgetFn<Rsc, Masked> {
         move |state| Masked {
+            shape: None,
+            inner: self.add_strong(state),
+        }
+    }
+
+    /// Clip to `shape` rather than to a plain box: `shape` is drawn
+    /// behind this widget, filling the same region, and what clips is the
+    /// primitive it drew -- so a rounded background and the corner its
+    /// content is cut to are one rect, with no radius passed twice.
+    /// Replaces `.masked().background(w)`, which drew the two but clipped
+    /// to the box.
+    fn masked_by<T>(self, shape: impl WidgetLike<Rsc, T>) -> impl WidgetFn<Rsc, Masked> {
+        move |state| Masked {
+            shape: Some(shape.add_strong(state)),
             inner: self.add_strong(state),
         }
     }

iris/tests/mask_sdf.rs

@@ -0,0 +1,403 @@
+//! The CPU rounded-rect SDF and the shader's own must agree.
+//!
+//! LAYOUT.md's "Masks with a shape" turns on it: the fragment stage clips
+//! a masked subtree with `shader.wgsl`'s `rounded_rect_coverage`, and the
+//! hit test (`UiRenderState::mask_admits`) clips the *same* subtree with
+//! `iris_core::rounded_rect_coverage`, so a corner that cannot be tapped
+//! and a corner that is not drawn are the same corner only while the two
+//! functions answer the same. Nothing else checks that: both sides are
+//! individually plausible and drift shows up as a control that is a pixel
+//! or two off, which is exactly what nobody notices.
+//!
+//! So this runs **the real shader text**, lifted out of
+//! `iris_core::SHAPE_SHADER` by name rather than copied here, over a grid
+//! of points, and compares what came back with the Rust function at the
+//! same points. This is the only test in the workspace that needs a GPU;
+//! everything else about masks is layer 1 (docs/RUST.md's "Three test
+//! layers"). It fails rather than skips when there is no adapter, because
+//! a check that quietly did not run reads exactly like a check that
+//! passed.
+//!
+//! **It is a render pass, and it asks for `iris_core::device_limits()`,
+//! because those are the two things iris itself does.** The first version
+//! of this test was a compute pass, which meant asking for compute limits
+//! that `device_limits()` deliberately zeroes -- docs/RUST.md, 2026-09-05:
+//! nothing in `iris`/`iris-core` creates a `ComputePipeline` or writes a
+//! `@compute` stage, so the limits stopped being requested rather than a
+//! fallback being built for a capability nothing uses. A test that needs
+//! a capability the thing under test has never needed is testing the
+//! wrong device, which is reason enough.
+//!
+//! It is **not** why that version crashed; see [`vulkan_instance`] for
+//! what that crash actually was and why nothing here has to work around
+//! it any more.
+
+// `OnceLock<wgpu::Instance>` needs `Instance: Sync`, and wgpu's type
+// graph is deep enough that proving it overflows rustc's default trait
+// recursion limit of 128. Nothing here is recursive; the limit is a
+// compile-time budget, and this is the documented way to raise it.
+#![recursion_limit = "256"]
+
+use std::sync::OnceLock;
+
+use iris_core::{SHAPE_SHADER, rounded_rect_coverage, util::Vec2};
+use pollster::FutureExt;
+use wgpu::util::DeviceExt;
+
+/// The rect the grid is sampled against, in window pixels. Deliberately
+/// off the whole-pixel grid: the shader floors a primitive's corners, but
+/// `rounded_rect_coverage` is handed pixels either side of that and has to
+/// agree at fractional positions too -- the phone's 2.55 density puts
+/// nothing on a whole pixel.
+const TOP_LEFT: Vec2 = Vec2::new(10.5, 20.25);
+const BOT_RIGHT: Vec2 = Vec2::new(170.75, 90.0);
+
+/// Radii spanning what the widgets actually ask for, plus the two edges of
+/// the function's own domain: a square corner, and one large enough that
+/// `min(edge, radius)` stops mattering.
+const RADII: [f32; 5] = [0.0, 0.75, 8.0, 20.0, 34.0];
+
+/// The grid, as an attachment: one texel per probe point. `GRID_W` is a
+/// multiple of 64 so that a row of `R32Float` is 256-byte aligned, which
+/// is what `copy_texture_to_buffer` requires; at `STEP` this spans the
+/// rect above and about four pixels of margin on every side, so the
+/// feather is sampled rather than stepped over.
+const GRID_W: u32 = 384;
+const GRID_H: u32 = 192;
+const STEP: f32 = 0.5;
+const ORIGIN: Vec2 = Vec2::new(TOP_LEFT.x - 4.0, TOP_LEFT.y - 4.0);
+
+/// f32 arithmetic in two compilers, not one: `length`/`sqrt` and
+/// `smoothstep` are each allowed a unit or two in the last place, and the
+/// GPU may contract a multiply-add the CPU does not. A coverage is in
+/// [0, 1], so this is about six decimal digits -- four orders of magnitude
+/// tighter than the half-pixel feather the hit test reads, which is what
+/// the agreement is actually for.
+const TOLERANCE: f32 = 1e-5;
+
+#[test]
+fn mask_sdf_matches_the_shader() {
+    let gpu = Gpu::open();
+    let mut worst = 0.0f32;
+    let mut worst_at = (Vec2::new(0.0, 0.0), 0.0f32, 0.0f32, 0.0f32);
+    let (mut inside, mut feather, mut outside) = (0u32, 0u32, 0u32);
+
+    for radius in RADII {
+        let coverage = run_shader(&gpu, radius);
+        for y in 0..GRID_H {
+            for x in 0..GRID_W {
+                let pos = probe_at(x, y);
+                let got = coverage[(y * GRID_W + x) as usize];
+                let want = rounded_rect_coverage(pos, TOP_LEFT, BOT_RIGHT, radius);
+                let diff = (got - want).abs();
+                if diff > worst {
+                    worst = diff;
+                    worst_at = (pos, radius, want, got);
+                }
+                if got > 0.999 {
+                    inside += 1;
+                } else if got > 0.001 {
+                    feather += 1;
+                } else {
+                    outside += 1;
+                }
+            }
+        }
+    }
+
+    let (pos, radius, want, got) = worst_at;
+    assert!(
+        worst <= TOLERANCE,
+        "shader.wgsl's rounded_rect_coverage and iris_core's disagree by {worst} at {pos:?} \
+         (radius {radius}): the CPU says {want}, the GPU {got}. One of the two was edited \
+         without the other -- they are transliterations and have to stay so, or a masked \
+         corner stops being tappable where it is drawn.",
+    );
+
+    // The half that would pass on a function returning a constant.
+    assert!(
+        inside > 0 && feather > 0 && outside > 0,
+        "the grid never crossed an edge ({inside} in, {feather} on the feather, {outside} out), \
+         so agreeing proved nothing",
+    );
+}
+
+/// The probe position of texel `(x, y)` -- the one place the mapping
+/// lives, so the CPU side and the fragment stage cannot walk different
+/// grids.
+fn probe_at(x: u32, y: u32) -> Vec2 {
+    Vec2::new(ORIGIN.x + x as f32 * STEP, ORIGIN.y + y as f32 * STEP)
+}
+
+/// `shader.wgsl`'s own `rounded_rect_coverage`, evaluated at every texel
+/// of an `R32Float` attachment: one fragment per grid point, read back
+/// whole. A fragment stage because that is the stage the function is
+/// really called from, so what this compares is the code path that draws
+/// rather than a second one built to be measurable.
+fn run_shader(gpu: &Gpu, radius: f32) -> Vec<f32> {
+    let Gpu { device, queue, .. } = gpu;
+    let module = device.create_shader_module(wgpu::ShaderModuleDescriptor {
+        label: Some("mask sdf probe"),
+        source: wgpu::ShaderSource::Wgsl(probe_source().into()),
+    });
+
+    // R32Float, not an 8-bit colour format: a coverage quantised to 1/255
+    // could not be compared against the CPU's at anything like TOLERANCE,
+    // and the comparison would then be measuring the texture rather than
+    // the two functions.
+    let texture = device.create_texture(&wgpu::TextureDescriptor {
+        label: Some("mask sdf coverage"),
+        size: wgpu::Extent3d {
+            width: GRID_W,
+            height: GRID_H,
+            depth_or_array_layers: 1,
+        },
+        mip_level_count: 1,
+        sample_count: 1,
+        dimension: wgpu::TextureDimension::D2,
+        format: wgpu::TextureFormat::R32Float,
+        usage: wgpu::TextureUsages::RENDER_ATTACHMENT | wgpu::TextureUsages::COPY_SRC,
+        view_formats: &[],
+    });
+    let view = texture.create_view(&Default::default());
+
+    let probe = Probe {
+        top_left: [TOP_LEFT.x, TOP_LEFT.y],
+        bot_right: [BOT_RIGHT.x, BOT_RIGHT.y],
+        origin: [ORIGIN.x, ORIGIN.y],
+        step: [STEP, STEP],
+        radius,
+        _pad: [0.0; 3],
+    };
+    let uniform = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("mask sdf probe"),
+        contents: bytemuck::bytes_of(&probe),
+        usage: wgpu::BufferUsages::UNIFORM,
+    });
+
+    let pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+        label: Some("mask sdf probe"),
+        layout: None,
+        vertex: wgpu::VertexState {
+            module: &module,
+            entry_point: Some("probe_vs"),
+            compilation_options: Default::default(),
+            buffers: &[],
+        },
+        fragment: Some(wgpu::FragmentState {
+            module: &module,
+            entry_point: Some("probe_fs"),
+            compilation_options: Default::default(),
+            targets: &[Some(wgpu::TextureFormat::R32Float.into())],
+        }),
+        primitive: Default::default(),
+        depth_stencil: None,
+        multisample: Default::default(),
+        multiview_mask: None,
+        cache: None,
+    });
+    let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
+        label: Some("mask sdf probe"),
+        layout: &pipeline.get_bind_group_layout(0),
+        entries: &[wgpu::BindGroupEntry {
+            binding: 0,
+            resource: uniform.as_entire_binding(),
+        }],
+    });
+
+    // `copy_texture_to_buffer` wants each row 256-byte aligned; GRID_W is
+    // chosen so that it already is, rather than padding and unpicking the
+    // padding on the way out.
+    let row_bytes = GRID_W * 4;
+    assert_eq!(row_bytes % 256, 0, "GRID_W must keep rows 256-byte aligned");
+    let out_size = u64::from(row_bytes) * u64::from(GRID_H);
+    let read_buf = device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some("mask sdf readback"),
+        size: out_size,
+        usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::COPY_DST,
+        mapped_at_creation: false,
+    });
+
+    let mut enc = device.create_command_encoder(&Default::default());
+    {
+        let mut pass = enc.begin_render_pass(&wgpu::RenderPassDescriptor {
+            label: Some("mask sdf probe"),
+            color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+                view: &view,
+                depth_slice: None,
+                resolve_target: None,
+                ops: wgpu::Operations {
+                    load: wgpu::LoadOp::Clear(wgpu::Color::BLACK),
+                    store: wgpu::StoreOp::Store,
+                },
+            })],
+            depth_stencil_attachment: None,
+            multiview_mask: None,
+            timestamp_writes: None,
+            occlusion_query_set: None,
+        });
+        pass.set_pipeline(&pipeline);
+        pass.set_bind_group(0, &bind_group, &[]);
+        pass.draw(0..3, 0..1);
+    }
+    enc.copy_texture_to_buffer(
+        texture.as_image_copy(),
+        wgpu::TexelCopyBufferInfo {
+            buffer: &read_buf,
+            layout: wgpu::TexelCopyBufferLayout {
+                offset: 0,
+                bytes_per_row: Some(row_bytes),
+                rows_per_image: Some(GRID_H),
+            },
+        },
+        wgpu::Extent3d {
+            width: GRID_W,
+            height: GRID_H,
+            depth_or_array_layers: 1,
+        },
+    );
+    queue.submit([enc.finish()]);
+
+    let slice = read_buf.slice(..);
+    slice.map_async(wgpu::MapMode::Read, |r| r.expect("mapping the readback"));
+    device
+        .poll(wgpu::PollType::wait_indefinitely())
+        .expect("waiting for the probe");
+    let coverage = bytemuck::cast_slice::<u8, f32>(&slice.get_mapped_range()).to_vec();
+    read_buf.unmap();
+    coverage
+}
+
+/// One `wgpu::Instance` for the process, created on first use and never
+/// destroyed.
+///
+/// **Why it is a static rather than a value the test owns.** Destroying
+/// the last `VkInstance` makes the Vulkan loader `dlclose` the ICD, and
+/// Mesa's ICD here registers a `pthread_key_create` destructor pointing
+/// into its own text without being linked `-z nodelete`. glibc then calls
+/// that destructor when the thread exits -- through an address that is no
+/// longer mapped. libtest runs every `#[test]` on a spawned thread, so a
+/// test that opens and closes an instance segfaults *after* printing its
+/// result, which reads exactly like the test failing. Measured
+/// 2026-09-08 with `rigs/gpu-probe`'s `teardown` bin: it
+/// needs no wgpu (raw `ash` does it too), no GPU work, and no device --
+/// an instance created and destroyed on a spawned thread is enough, and
+/// keeping any one instance alive is enough to prevent it.
+///
+/// Devices, queues and everything else drop normally; only the instance
+/// is held, which is what wgpu asks for anyway (one instance per
+/// process). So this costs one instance for the length of a test binary
+/// and buys ordinary drops everywhere else.
+fn vulkan_instance() -> &'static wgpu::Instance {
+    static INSTANCE: OnceLock<wgpu::Instance> = OnceLock::new();
+    INSTANCE.get_or_init(wgpu::Instance::default)
+}
+
+/// The device this test draws with.
+struct Gpu {
+    device: wgpu::Device,
+    queue: wgpu::Queue,
+}
+
+impl Gpu {
+    /// Opens the device this test draws with, and reports which adapter
+    /// answered, because that is not a detail here: a run on llvmpipe and
+    /// a run on the host's GPU are otherwise indistinguishable in the
+    /// log, and only one of them is a check of what the phone will do.
+    fn open() -> Self {
+        let instance = vulkan_instance();
+        let adapter = instance
+            .request_adapter(&wgpu::RequestAdapterOptions::default())
+            .block_on()
+            .expect(
+                "no wgpu adapter on this machine, so the CPU/shader SDF agreement went \
+                 unchecked. This VM has a virtio-gpu render node (the `this-machine-graphics` \
+                 skill says what it is and how it fails); if that is gone, fix it rather \
+                 than deleting this test.",
+            );
+        let info = adapter.get_info();
+        eprintln!(
+            "mask_sdf: {} ({:?}, {})",
+            info.name, info.backend, info.driver
+        );
+        let (device, queue) = adapter
+            .request_device(&wgpu::DeviceDescriptor {
+                // What iris itself asks for -- see this file's header.
+                required_limits: iris_core::device_limits(),
+                ..Default::default()
+            })
+            .block_on()
+            .expect("could not get a device from the adapter");
+        Self { device, queue }
+    }
+}
+
+/// What the fragment stage needs to turn its own texel into a probe
+/// position: the rect being sampled, and where texel (0, 0) sits.
+#[repr(C)]
+#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
+struct Probe {
+    top_left: [f32; 2],
+    bot_right: [f32; 2],
+    origin: [f32; 2],
+    step: [f32; 2],
+    radius: f32,
+    _pad: [f32; 3],
+}
+
+/// The probe module: the two functions **lifted from `shader.wgsl`
+/// itself**, plus an entry point that calls the outer one. Lifted rather
+/// than copied so there is nothing to keep in step -- an edit to the
+/// shader is what this test is for, and a copy here would be edited along
+/// with it.
+fn probe_source() -> String {
+    format!(
+        "{}\n{}\n\
+         struct Probe {{\n\
+         top_left: vec2<f32>,\n\
+         bot_right: vec2<f32>,\n\
+         origin: vec2<f32>,\n\
+         step: vec2<f32>,\n\
+         radius: f32,\n\
+         }}\n\
+         @group(0) @binding(0) var<uniform> probe: Probe;\n\
+         @vertex\n\
+         fn probe_vs(@builtin(vertex_index) vi: u32) -> @builtin(position) vec4<f32> {{\n\
+         var xy = array(vec2(-1.0, -1.0), vec2(3.0, -1.0), vec2(-1.0, 3.0));\n\
+         return vec4(xy[vi], 0.0, 1.0);\n\
+         }}\n\
+         @fragment\n\
+         fn probe_fs(@builtin(position) pos: vec4<f32>) -> @location(0) f32 {{\n\
+         let at = probe.origin + floor(pos.xy) * probe.step;\n\
+         return rounded_rect_coverage(at, probe.top_left, probe.bot_right, probe.radius);\n\
+         }}\n",
+        wgsl_fn("distance_from_rect"),
+        wgsl_fn("rounded_rect_coverage"),
+    )
+}
+
+/// One WGSL function's whole text, from its `fn` keyword to the `}` that
+/// closes its body, found by matching braces. Panics by name when the
+/// function is not there, which is what a rename looks like from here.
+fn wgsl_fn(name: &str) -> &'static str {
+    let start = SHAPE_SHADER
+        .find(&format!("fn {name}("))
+        .unwrap_or_else(|| panic!("shader.wgsl has no `fn {name}(` -- renamed, or gone"));
+    let body = SHAPE_SHADER[start..]
+        .find('{')
+        .expect("a wgsl fn signature is followed by its body");
+    let mut depth = 0usize;
+    for (i, c) in SHAPE_SHADER[start + body..].char_indices() {
+        match c {
+            '{' => depth += 1,
+            '}' => {
+                depth -= 1;
+                if depth == 0 {
+                    return &SHAPE_SHADER[start..start + body + i + 1];
+                }
+            }
+            _ => {}
+        }
+    }
+    panic!("`fn {name}`'s body in shader.wgsl is never closed");
+}

iris/transcript-fixture/Cargo.toml

@@ -22,3 +22,7 @@ serde_json = { version = "1", features = ["float_roundtrip"] }
 
 [dev-dependencies]
 winit = { workspace = true }
+# For the `iris::input`/`iris::frame` round-trip test: a capturing `log::Log`
+# to read back what `iris::diagnostics::log_frame`/`sense::log_input_event`
+# wrote, pinned to the same version `iris/Cargo.toml` already carries.
+log = "0.4.28"

iris/transcript-fixture/tests/catch_a_fling.rs

@@ -0,0 +1,211 @@
+//! Layer 1 of docs/RUST.md's "Three test layers", for catching a fling:
+//! the real transcript screen over the real bench fixture, at the phone's
+//! size and density, with no window, no compositor and no GPU.
+//!
+//! docs/IRIS_TODO.md's 2026-09-07 night report -- "sometimes when I try to
+//! catch it while it's still moving (particularly if I drag) then it fails
+//! to stop & snap to where finger is". The finger goes down on content
+//! that is still travelling and the content does not follow it until
+//! `DRAG_SLOP` has been crossed, which at a fling's speed is several
+//! frames of the content sliding *away* from a finger that is already
+//! down. Compose does not do that: a down while `isScrollInProgress`
+//! starts the drag immediately (`scrollable`'s `startDragImmediately`).
+
+use iris::harness::{Harness, TouchAction, TouchScript};
+use iris::prelude::*;
+use iris::sense::DRAG_SLOP;
+use transcript_fixture::{PHONE_FRAME_MS, PHONE_SCALE, phone_size};
+
+/// The screen open on the fixture, framed twice -- once to draw, once for
+/// `List::repair_anchor` to resolve the opening `snap_end` into a real
+/// anchor, which is what every assertion about scroll position reads.
+fn opened() -> (Harness, transcript_ui::TranscriptScreen) {
+    let mut h = Harness::new(phone_size(), PHONE_SCALE);
+    let opened = transcript_fixture::open(&mut h.rsc, &mut h.state).expect("the fixture folds");
+    h.frame(0);
+    h.frame(PHONE_FRAME_MS);
+    (h, opened.screen)
+}
+
+/// Where the content is, in window pixels: the top of whichever row is
+/// under the middle of the viewport. `List` has no travel accessor and
+/// this needs none -- a row's own extent moves exactly as far as the
+/// content does, and the row is picked once so the two readings compare.
+fn tracked_row(h: &mut Harness, screen: &transcript_ui::TranscriptScreen) -> (RowKey, f32) {
+    let middle = phone_size().y / 2.0;
+    let list = (screen.list)(&mut h.rsc);
+    let key = list.key_at(middle).expect("a row under the viewport");
+    let (top, _) = list.extent(key).expect("that row has an extent");
+    (key, top)
+}
+
+fn row_top(h: &mut Harness, screen: &transcript_ui::TranscriptScreen, key: RowKey) -> f32 {
+    (screen.list)(&mut h.rsc)
+        .extent(key)
+        .expect("the tracked row is still loaded")
+        .0
+}
+
+/// Three finger samples 8ms apart, each moving `STEP` further down the
+/// screen. `STEP * 3` is deliberately **under** `DRAG_SLOP`: a gesture
+/// this small moves nothing at all on a settled list (the control below),
+/// so anything it moves here is the catch and not the slop being crossed.
+const STEP: f32 = 2.0;
+const SAMPLES: usize = 3;
+const CATCH_X: f32 = 540.0;
+
+/// Feeds the down and its `SAMPLES` moves from `y0` at `t0`, asserting
+/// after each one that the content moved by exactly the finger's own
+/// delta. Returns the release time.
+fn drag_from(
+    h: &mut Harness,
+    screen: &transcript_ui::TranscriptScreen,
+    key: RowKey,
+    y0: f32,
+    t0: u64,
+    expect_tracking: bool,
+) -> u64 {
+    let before = row_top(h, screen, key);
+    h.touch(TouchAction::Down, Vec2::new(CATCH_X, y0), t0);
+    assert_eq!(
+        row_top(h, screen, key),
+        before,
+        "the down itself must not move the content, only stop it"
+    );
+    let mut t = t0;
+    for i in 1..=SAMPLES {
+        let moved = STEP * i as f32;
+        t = t0 + 8 * i as u64;
+        h.touch(TouchAction::Move, Vec2::new(CATCH_X, y0 + moved), t);
+        let travelled = row_top(h, screen, key) - before;
+        if expect_tracking {
+            assert!(
+                (travelled - moved).abs() < 0.5,
+                "sample {i}: the finger has moved {moved}px since the down and the content \
+                 {travelled:.1}px -- it is not pinned to the finger"
+            );
+        } else {
+            assert!(
+                travelled.abs() < 0.5,
+                "sample {i}: a {moved}px drag is inside DRAG_SLOP ({DRAG_SLOP}px) and must move \
+                 nothing, but the content moved {travelled:.1}px"
+            );
+        }
+    }
+    t += 8;
+    h.touch(
+        TouchAction::Up,
+        Vec2::new(CATCH_X, y0 + STEP * SAMPLES as f32),
+        t,
+    );
+    t
+}
+
+/// The report itself: flick, let the fling run for 150ms, then put a
+/// finger down and drag it a little. From the down onwards the content is
+/// pinned to the finger, sample for sample -- no slop, and no coasting
+/// past the place the finger stopped it.
+#[test]
+fn a_press_on_a_flinging_list_pins_the_content_to_the_finger() {
+    let (mut h, screen) = opened();
+
+    let flick = TouchScript::parse(include_str!("../touch/flick-120hz.touch"))
+        .unwrap_or_else(|e| panic!("flick-120hz.touch: {e}"));
+    h.replay(&flick);
+
+    // Frames, not a file: the second half of this gesture has to arrive
+    // *while* the fling is ticking, and a `.touch` replay inserts no
+    // frames between its samples, so a fling recorded that way would be
+    // running on paper and stationary in fact.
+    let catch_at = flick.end_ms() + 150;
+    h.frames_until(
+        flick.end_ms() + PHONE_FRAME_MS,
+        catch_at - PHONE_FRAME_MS,
+        PHONE_FRAME_MS,
+    );
+    assert!(
+        (screen.list)(&mut h.rsc).is_scrolling(),
+        "the fling must still be running 150ms in, or this test catches nothing"
+    );
+
+    let (key, _) = tracked_row(&mut h, &screen);
+    drag_from(&mut h, &screen, key, 1200.0, catch_at, true);
+}
+
+/// The other half of the same rule: a catch that never moved at all is a
+/// `Released(None)`, not a tap. Compose's scrollable consumes that DOWN,
+/// so no click detector under it ever sees the gesture -- stopping a
+/// fling with a finger must not also follow the link it landed on, and
+/// must not hand the list a velocity to start again with.
+#[test]
+fn a_catch_that_never_moved_is_not_a_tap_and_does_not_fling() {
+    let (mut h, screen) = opened();
+
+    let flick = TouchScript::parse(include_str!("../touch/flick-120hz.touch"))
+        .unwrap_or_else(|e| panic!("flick-120hz.touch: {e}"));
+    h.replay(&flick);
+    let catch_at = flick.end_ms() + 150;
+    h.frames_until(
+        flick.end_ms() + PHONE_FRAME_MS,
+        catch_at - PHONE_FRAME_MS,
+        PHONE_FRAME_MS,
+    );
+    assert!(
+        (screen.list)(&mut h.rsc).is_scrolling(),
+        "the fling must still be running 150ms in, or this test catches nothing"
+    );
+
+    let (key, _) = tracked_row(&mut h, &screen);
+    h.touch(TouchAction::Down, Vec2::new(CATCH_X, 1200.0), catch_at);
+    let stopped_at = row_top(&mut h, &screen, key);
+    h.touch(TouchAction::Up, Vec2::new(CATCH_X, 1200.0), catch_at + 8);
+
+    assert_eq!(
+        (screen.list)(&mut h.rsc).fling_velocity(),
+        None,
+        "a press that stopped a fling and moved nothing must not start another"
+    );
+    h.frames_until(catch_at + 16, catch_at + 500, PHONE_FRAME_MS);
+    assert!(
+        (row_top(&mut h, &screen, key) - stopped_at).abs() < 0.5,
+        "the content moved after a catch was released without moving"
+    );
+    assert_eq!(
+        h.state.opened_urls,
+        Vec::<String>::new(),
+        "a catch is not a tap: nothing under it may be followed"
+    );
+}
+
+/// The half this change had no reason to touch: on a list that is *not*
+/// moving, the same tiny drag is still inside `DRAG_SLOP` and still moves
+/// nothing. Without this, making every press pin the content would pass
+/// the test above and take the slop away from every ordinary press.
+#[test]
+fn the_same_small_drag_on_a_settled_list_moves_nothing() {
+    let (mut h, screen) = opened();
+
+    let flick = TouchScript::parse(include_str!("../touch/flick-120hz.touch"))
+        .unwrap_or_else(|e| panic!("flick-120hz.touch: {e}"));
+    h.replay(&flick);
+    // Long past the spline's own 2071ms for this recording.
+    let settled = h.frames_until(
+        flick.end_ms() + PHONE_FRAME_MS,
+        flick.end_ms() + 4000,
+        PHONE_FRAME_MS,
+    );
+    assert!(
+        !(screen.list)(&mut h.rsc).is_scrolling(),
+        "the fling must have stopped, or this is the same case as the test above"
+    );
+
+    let (key, _) = tracked_row(&mut h, &screen);
+    drag_from(
+        &mut h,
+        &screen,
+        key,
+        1200.0,
+        settled + PHONE_FRAME_MS,
+        false,
+    );
+}

iris/transcript-fixture/tests/input_log_roundtrip.rs

@@ -0,0 +1,189 @@
+//! Layer 1 of docs/RUST.md's "Three test layers", for the diagnostics
+//! themselves rather than a widget: `iris::diagnostics::set_trace` gates
+//! `iris::input`/`iris::frame` (Iris's 2026-09-07 request, "add another
+//! button to copy input event info ... instrument a lot of the code with
+//! timings"), and `docs/REVIEW-2026-09-07.md`'s D1 found that the switch
+//! existed but four older per-frame `debug!` lines were not wired to it,
+//! filling the app's 2000-line log ring with frame spam before `Copy
+//! report` had a chance to include anything else. This is what a fix to
+//! that has to prove, both directions:
+//!
+//! 1. **Off** (the default): replaying a real gesture through a real
+//!    screen leaves the ring holding nothing below `info` -- so the
+//!    lines D1 named, and everything this pass gated the same way, really
+//!    are silent by default rather than merely "usually quiet."
+//! 2. **On**: the same replay produces `iris::input` lines that
+//!    `report_to_touch.py` turns back into the exact `TouchScript` that
+//!    was replayed, and `iris::frame` lines with real, non-zero
+//!    durations dated on the harness's own clock.
+//!
+//! **Single capturing logger, single test function** (this file's only
+//! `#[test]`): `log::set_logger` can succeed exactly once per process, and
+//! AGENTS.md's "tracing caches callsite interest process-wide" lesson is
+//! the general form of why every exercise of a logging path has to share
+//! one subscriber -- so if a second test here ever needs the ring's
+//! contents, it must extend this one rather than install its own.
+
+use std::process::{Command, Stdio};
+use std::sync::{Mutex, OnceLock};
+
+use iris::harness::{Harness, TouchScript};
+use transcript_fixture::{PHONE_FRAME_MS, PHONE_SCALE, phone_size};
+
+/// Records every line's level and formatted message -- enough to answer
+/// both "is the ring quiet" (no line at `Debug` or below) and "what did
+/// tracing actually write" (the `iris::input` lines, read back by
+/// `report_to_touch.py`).
+struct CaptureLogger {
+    lines: Mutex<Vec<(log::Level, String)>>,
+}
+
+static LOGGER: OnceLock<CaptureLogger> = OnceLock::new();
+
+impl log::Log for CaptureLogger {
+    fn enabled(&self, _metadata: &log::Metadata) -> bool {
+        true
+    }
+    fn log(&self, record: &log::Record) {
+        self.lines
+            .lock()
+            .unwrap()
+            .push((record.level(), record.args().to_string()));
+    }
+    fn flush(&self) {}
+}
+
+/// Installs the capture logger at `Debug` -- the same level
+/// `iris/android-app/src/lib.rs`'s `JNI_OnLoad` installs at, which is
+/// exactly why `iris::diagnostics::trace_enabled` has to be the gate
+/// (its own module doc) rather than the level.
+fn logger() -> &'static CaptureLogger {
+    let logger = LOGGER.get_or_init(|| CaptureLogger {
+        lines: Mutex::new(Vec::new()),
+    });
+    // Ignore "already set": a previous call in this same test binary
+    // already won, and it is the same logger either way.
+    let _ = log::set_logger(logger);
+    log::set_max_level(log::LevelFilter::Debug);
+    logger
+}
+
+fn drain(logger: &CaptureLogger) -> Vec<(log::Level, String)> {
+    std::mem::take(&mut *logger.lines.lock().unwrap())
+}
+
+fn opened() -> (Harness, transcript_ui::TranscriptScreen) {
+    let mut h = Harness::new(phone_size(), PHONE_SCALE);
+    let opened = transcript_fixture::open(&mut h.rsc, &mut h.state).expect("the fixture folds");
+    h.frame(0);
+    h.frame(PHONE_FRAME_MS);
+    (h, opened.screen)
+}
+
+#[test]
+fn tracing_is_silent_off_and_round_trips_the_flick_on() {
+    let logger = logger();
+
+    // --- (1) off: a real flick through a real screen leaves the ring
+    // with nothing at `Debug` or below.
+    iris::diagnostics::set_trace(false);
+    drain(logger); // whatever `opened()` itself logged while building
+    let (mut h, screen) = opened();
+    drain(logger); // and whatever opening logged
+    let flick = TouchScript::parse(include_str!("../touch/flick-120hz.touch"))
+        .unwrap_or_else(|e| panic!("flick-120hz.touch: {e}"));
+    h.replay(&flick);
+    let _ = (screen.list)(&mut h.rsc); // touch the screen the same way a real caller would
+    let quiet = drain(logger);
+    let debug_lines: Vec<_> = quiet
+        .iter()
+        .filter(|(level, _)| *level == log::Level::Debug)
+        .collect();
+    assert!(
+        debug_lines.is_empty(),
+        "tracing is off, but the ring would still have held these `debug!` lines: {debug_lines:#?}"
+    );
+
+    // --- (2) on: the same replay, from a fresh screen so the anchor and
+    // sequence numbers match `flick-120hz.touch` exactly again.
+    iris::diagnostics::set_trace(true);
+    let (mut h, screen) = opened();
+    drain(logger);
+    h.replay(&flick);
+    let _ = (screen.list)(&mut h.rsc);
+    let traced = drain(logger);
+    iris::diagnostics::set_trace(false); // leave it off for any test after this one
+
+    let input_lines: Vec<&str> = traced
+        .iter()
+        .filter(|(_, msg)| msg.contains("iris input: action="))
+        .map(|(_, msg)| msg.as_str())
+        .collect();
+    assert_eq!(
+        input_lines.len(),
+        flick.samples.len(),
+        "expected one `iris::input` line per replayed sample, got:\n{input_lines:#?}"
+    );
+    let frame_lines: Vec<&str> = traced
+        .iter()
+        .filter(|(_, msg)| msg.starts_with("iris frame:"))
+        .map(|(_, msg)| msg.as_str())
+        .collect();
+    assert!(
+        !frame_lines.is_empty(),
+        "expected at least one `iris::frame` line once tracing was on"
+    );
+    for line in &frame_lines {
+        // `layout=` and `draw=` are `{:?}`-formatted `Duration`s, so a real
+        // one reads like `12.34µs`/`1.2ms`, never the bare `0ns` a
+        // no-op frame would print.
+        assert!(
+            !line.contains("layout=0ns"),
+            "a frame that redrew should not report zero layout time: {line}"
+        );
+    }
+
+    // --- the round trip: pipe every `iris::input` line through
+    // `report_to_touch.py` and parse the result back into a `TouchScript`,
+    // which must equal the one that was replayed. `report_to_touch.py`
+    // is prefix-agnostic (it `search`es for the marker), so handing it
+    // the bare message is the same as handing it a real ring line.
+    let report = input_lines.join("\n");
+    let script_path = concat!(env!("CARGO_MANIFEST_DIR"), "/../benches/report_to_touch.py");
+    let mut child = Command::new("python3")
+        .arg(script_path)
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped())
+        .spawn()
+        .expect("python3 must be on PATH to run report_to_touch.py");
+    {
+        use std::io::Write;
+        child
+            .stdin
+            .take()
+            .unwrap()
+            .write_all(report.as_bytes())
+            .unwrap();
+    }
+    let output = child.wait_with_output().expect("report_to_touch.py exited");
+    assert!(
+        output.status.success(),
+        "report_to_touch.py failed: {}",
+        String::from_utf8_lossy(&output.stderr)
+    );
+    let touch_text = String::from_utf8(output.stdout).expect("report_to_touch.py wrote UTF-8");
+    let round_tripped =
+        TouchScript::parse(&touch_text).unwrap_or_else(|e| panic!("round-tripped script: {e}"));
+
+    assert_eq!(
+        round_tripped.samples.len(),
+        flick.samples.len(),
+        "round trip produced a different number of samples:\n{touch_text}"
+    );
+    for (original, back) in flick.samples.iter().zip(round_tripped.samples.iter()) {
+        assert_eq!(original.t_ms, back.t_ms);
+        assert_eq!(original.action, back.action);
+        assert_eq!(original.pos, back.pos);
+    }
+}

iris/transcript-fixture/tests/phone_screen.rs

@@ -47,21 +47,53 @@ fn a_recorded_flick_releases_with_a_velocity_and_flings_the_list() {
     let velocity = (screen.list)(&mut h.rsc)
         .fling_velocity()
         .expect("the flick must release as a pan with a velocity, not a tap");
+    // Compose's own answer for this recording's five samples, printed by
+    // `iris/benches/velocity_reference.py` -- not a number read off this
+    // code. Negative because the flick runs *down* the screen and
+    // `Selection::drag` flings the list by `-v` (see its `Released` arm).
+    // The 2026-09-07 before/after: the old average estimator read
+    // -12250px/s here, which is the fling Iris reported as too slow.
     assert!(
-        velocity.abs() > 1_000.0,
-        "a 188px, 16ms flick is thousands of px/s; got {velocity}"
+        (velocity + 15_250.0).abs() < 20.0,
+        "expected ~-15250px/s from velocity_reference.py, got {velocity}"
     );
 
-    // Android's own spline says how long a fling at this speed runs. The
-    // list learns its density from the painter, so this is the same
-    // curve it is using.
-    let expected = FlingCalculator::new(PHONE_SCALE).duration(velocity);
-    let end = flick.end_ms() + expected.as_millis() as u64 * 2;
+    // `iris/benches/fling_spline_reference.py`'s own line for this exact
+    // case -- `density=2.55 v=15250.0: distance=11057.424px
+    // duration=2.0716s`. **Not** `FlingCalculator::new(PHONE_SCALE)`,
+    // which is the calculator under test: bounding a fling with the thing
+    // being measured is the "compared the code with itself" shape 73f956f
+    // found in the spline's own tests, and it left this one able to fail
+    // in the "ran too long" direction only -- never in the "stopped dead"
+    // direction, which is what Iris actually reported
+    // (docs/REVIEW-2026-09-07.md's T1).
+    const REFERENCE_MS: u64 = 2071;
+    const REFERENCE_PX: f32 = 11057.0;
+    let end = flick.end_ms() + REFERENCE_MS * 2;
     let mut settled_at = None;
     let mut t = flick.end_ms();
+    // Travel in pixels, measured from a row's own on-screen extent, since
+    // `List` has no travel accessor and this needs none: follow whatever
+    // row is under the viewport's middle until it leaves, then pick
+    // another. Deliberately an *under*-count -- the frame a row leaves on
+    // contributes nothing -- which is why it is only ever a lower bound.
+    let middle = phone_size().y / 2.0;
+    let mut travelled = 0.0f32;
+    let mut tracked: Option<(RowKey, f32)> = None;
     while t <= end {
         h.frame(t);
-        if settled_at.is_none() && !(screen.list)(&mut h.rsc).is_scrolling() {
+        let list = (screen.list)(&mut h.rsc);
+        tracked =
+            match tracked.and_then(|(key, was)| list.extent(key).map(|(now, _)| (key, was, now))) {
+                Some((key, was, now)) => {
+                    travelled += (now - was).abs();
+                    Some((key, now))
+                }
+                None => list
+                    .key_at(middle)
+                    .and_then(|key| list.extent(key).map(|(top, _)| (key, top))),
+            };
+        if settled_at.is_none() && !list.is_scrolling() {
             settled_at = Some(t);
         }
         t += PHONE_FRAME_MS;
@@ -74,10 +106,24 @@ fn a_recorded_flick_releases_with_a_velocity_and_flings_the_list() {
     );
     let settled_at = settled_at.expect("the fling must stop on its own, not run forever");
     let ran_for = settled_at - flick.end_ms();
+    // Both directions. The lower bound is the one that fails when a fling
+    // settles on its first tick; the upper is the one that was here.
     assert!(
-        ran_for <= expected.as_millis() as u64 + PHONE_FRAME_MS * 2,
-        "the fling ran {ran_for}ms against the spline's own {}ms",
-        expected.as_millis()
+        ran_for >= REFERENCE_MS - PHONE_FRAME_MS * 2,
+        "the fling stopped after {ran_for}ms against the spline reference's {REFERENCE_MS}ms"
+    );
+    assert!(
+        ran_for <= REFERENCE_MS + PHONE_FRAME_MS * 2,
+        "the fling ran {ran_for}ms against the spline reference's {REFERENCE_MS}ms"
+    );
+    // 80% of the reference, against 10527px measured today -- the 5%
+    // shortfall is the frames a tracked row leaves the screen on. A fling
+    // that moves one row's worth fails this; scaling `tick_fling`'s delta
+    // by 0.01 reports 111px, which is how it was confirmed to fail in the
+    // direction the bug goes.
+    assert!(
+        travelled >= REFERENCE_PX * 0.8,
+        "the fling travelled {travelled:.0}px against the spline reference's {REFERENCE_PX:.0}px"
     );
 }
 

iris/transcript-fixture/tests/top_edge.rs

@@ -0,0 +1,331 @@
+//! Layer 1 of docs/RUST.md's "Three test layers", for the transcript's
+//! own edges: the real screen over the real fixture, under a header bar
+//! like the bench app's, driven by `iris::harness`.
+//!
+//! What these are about is docs/IRIS_TODO.md's 2026-09-07 phone report --
+//! rows scrolled above the viewport still drawn, over the header, and a
+//! blank band where the row straddling the top edge should be. Both are
+//! one rule (`List::intersects_viewport`): a row is drawn if any part of
+//! it is inside the list's own box, and nothing outside that box reaches
+//! the screen.
+
+use iris::harness::Harness;
+use iris::prelude::*;
+use transcript_fixture::{PHONE_FRAME_MS, PHONE_SCALE, phone_size};
+
+/// A header band above the transcript, as `bench_client.rs` puts one --
+/// the surface the rows were drawing over on the phone. Its exact height
+/// does not matter; what matters is that the list's own box does not
+/// start at the top of the window, so "above the viewport" and "off the
+/// screen" are different places.
+const HEADER_H: f32 = 300.0;
+const HEADER: UiColor = UiColor::new(28, 28, 34, 255);
+
+fn opened() -> (Harness, transcript_ui::TranscriptScreen) {
+    let mut h = Harness::new(phone_size(), PHONE_SCALE);
+    let (opened, tree) = transcript_fixture::build_screen(&mut h.rsc).expect("the fixture folds");
+    let content = WidgetPtr::new().add(&mut h.rsc);
+    content(&mut h.rsc).set(tree);
+    let root = (rect(HEADER).height(abs(HEADER_H)), content.height(rest(1)))
+        .span(Dir::DOWN)
+        .add_strong(&mut h.rsc)
+        .any();
+    h.state.set_root(root);
+    h.frame(0);
+    h.frame(PHONE_FRAME_MS);
+    (h, opened.screen)
+}
+
+/// The list's own on-screen box, in window pixels.
+fn list_box(h: &Harness, screen: &transcript_ui::TranscriptScreen) -> PixelRegion {
+    h.render
+        .window_region(&screen.list.id(), &h.rsc)
+        .expect("the list is on screen")
+}
+
+/// Every row the list drew this frame, as `(top, bottom)` window pixels,
+/// topmost first. A `List`'s direct children are exactly its rows, and
+/// `draw_inner`'s old-children diffing means a row it did not place this
+/// frame is not among them.
+fn drawn_rows(h: &Harness, screen: &transcript_ui::TranscriptScreen) -> Vec<(f32, f32)> {
+    let mut rows: Vec<(f32, f32)> = h
+        .render
+        .active
+        .get(&screen.list.id())
+        .expect("the list is drawn")
+        .children
+        .iter()
+        .filter_map(|id| h.render.window_region(id, &h.rsc))
+        .map(|px| (px.top_left.y, px.bot_right.y))
+        .collect();
+    rows.sort_by(|a, b| a.0.total_cmp(&b.0));
+    rows
+}
+
+/// Scrolls `amount` (negative walks back through older rows) and runs the
+/// frame it asks for, returning the time of the next one.
+fn scrolled(h: &mut Harness, screen: &transcript_ui::TranscriptScreen, amount: f32, t: u64) -> u64 {
+    (screen.list)(&mut h.rsc).scroll(amount);
+    h.frame(t);
+    t + PHONE_FRAME_MS
+}
+
+/// (a) of docs/IRIS_TODO.md's reproduction: with a row across the top
+/// edge, that row is placed -- the viewport's first pixel belongs to
+/// something. A rule that culled a row once its *top* left the viewport
+/// would leave a blank band here, which is the second of Iris's two
+/// screenshots.
+#[test]
+fn the_row_across_the_top_edge_is_drawn() {
+    let (mut h, screen) = opened();
+    let top = list_box(&h, &screen).top_left.y;
+    let mut t = PHONE_FRAME_MS * 2;
+
+    // 40px a frame, the shape a finger pan arrives in, through a straddle
+    // and out the other side of it many times over.
+    for _ in 0..60 {
+        t = scrolled(&mut h, &screen, -40.0, t);
+        let rows = drawn_rows(&h, &screen);
+        let first = *rows.first().expect("something is on screen");
+        assert!(
+            first.0 <= top + 0.5,
+            "a band of {:.1}px under the header belongs to no row: rows start at {:.1}, the list \
+             at {top:.1}",
+            first.0 - top,
+            first.0,
+        );
+        assert!(
+            first.1 > top,
+            "the row across the top edge was culled: it ends at {:.1}, above the list's own \
+             {top:.1}",
+            first.1,
+        );
+    }
+}
+
+/// (b): what falls outside the list's box is clipped rather than drawn
+/// over whatever is there. The straddling row above is drawn *in full*,
+/// so the only thing between its earlier lines and the header bar is this
+/// mask -- with none, the phone drew `version = "0.1.0"` behind the "Run
+/// benchmark" button.
+#[test]
+fn the_list_is_clipped_to_its_own_box() {
+    let (h, screen) = opened();
+    let active = h.render.active.get(&screen.list.id()).expect("drawn");
+    assert!(
+        active.mask != MaskIdx::NONE,
+        "the transcript's list is drawn with nothing clipping it",
+    );
+    let clip = h.render.mask_region(active.mask, &h.rsc);
+    let list = list_box(&h, &screen);
+    assert!(
+        clip.top_left.y >= list.top_left.y - 0.5 && clip.bot_right.y <= list.bot_right.y + 0.5,
+        "the clip {clip:?} reaches outside the list's own box {list:?}, so a row straddling an \
+         edge still draws past it",
+    );
+
+    // And the mask has to *reach* what the rows draw. The two above say a
+    // mask exists and sits in the right place; neither says any primitive
+    // references it, so a broken `Mask::parent` chain -- what d507ae4
+    // introduced -- would leave them green while a code fence inside a row
+    // drew unclipped again (docs/REVIEW-2026-09-07.md's T3).
+    let rows = h
+        .render
+        .active
+        .get(&screen.list.id())
+        .expect("the list is drawn")
+        .children
+        .clone();
+    let mut checked = 0;
+    for row in rows {
+        for prim in primitives_under(&h, row) {
+            assert!(
+                mask_chain(&h, prim).contains(&active.mask),
+                "a primitive of row {row:?} clips to {:?}, a chain that never reaches the list's \
+                 own mask {:?}",
+                mask_chain(&h, prim),
+                active.mask,
+            );
+            checked += 1;
+        }
+    }
+    assert!(
+        checked > 0,
+        "no row primitive was checked, so this test asserted nothing",
+    );
+}
+
+/// Every primitive `id` and its descendants drew, as `MaskIdx`es -- images
+/// excluded, since they live in a separate instance array with their own
+/// indices (`Primitives::free`).
+fn primitives_under(h: &Harness, id: WidgetId) -> Vec<MaskIdx> {
+    let Some(active) = h.render.active.get(&id) else {
+        return Vec::new();
+    };
+    let mut out: Vec<MaskIdx> = active
+        .primitives
+        .iter()
+        .filter(|p| p.binding != IMAGE_BINDING)
+        .map(|p| h.render.primitives.instance(p.slot).mask_idx)
+        .collect();
+    for child in &active.children {
+        out.extend(primitives_under(h, *child));
+    }
+    out
+}
+
+/// The chain the fragment stage walks from `mask`, outermost last.
+fn mask_chain(h: &Harness, mask: MaskIdx) -> Vec<MaskIdx> {
+    let mut chain = Vec::new();
+    let mut at = mask;
+    while at != MaskIdx::NONE {
+        assert!(
+            !chain.contains(&at),
+            "the mask chain from {mask:?} loops back to {at:?}",
+        );
+        chain.push(at);
+        at = h.rsc.ui.masks[at.idx()].parent;
+    }
+    chain
+}
+
+/// A row that has left the viewport entirely is not drawn at all. Before
+/// the fix the walk ran from the anchor -- which `scroll` leaves wherever
+/// it was, however far outside the viewport that ends up -- and drew
+/// every row on the way: 8 scrolls of 3000px left **64 rows** placed for
+/// a 2012px viewport, ~59 of them off screen and painting over the
+/// header.
+///
+/// The box is asserted on every leg *except the first*, because a row
+/// whose height has never been measured has to be drawn to be measured
+/// (`List::place`'s doc), which on the first walk back is every row
+/// entering from the top. Every later leg crosses the same rows with
+/// every height already known -- including the second walk *back*, which
+/// is there because a regression that draws rows in the wrong place while
+/// travelling backwards would otherwise be checked only by the row count
+/// (docs/REVIEW-2026-09-07.md's T2). That is also the ordinary state of a
+/// transcript being panned around in. The bound on how many rows are
+/// placed at once holds on all three.
+#[test]
+fn rows_that_have_left_the_viewport_are_not_drawn() {
+    let (mut h, screen) = opened();
+    let list = list_box(&h, &screen);
+    let mut t = PHONE_FRAME_MS * 2;
+    let bounded = |rows: &[(f32, f32)], leg: &str, step: usize| {
+        // A handful of rows whatever distance has been travelled -- the
+        // module doc's own claim about this widget.
+        assert!(
+            rows.len() <= 24,
+            "{leg} {step}: {} rows drawn for one 2012px viewport",
+            rows.len(),
+        );
+    };
+    let inside = |rows: &[(f32, f32)], leg: &str, step: usize| {
+        for &(top, bottom) in rows {
+            assert!(
+                bottom > list.top_left.y - 0.5 && top < list.bot_right.y + 0.5,
+                "{leg} {step}: a row at ({top:.1}, {bottom:.1}) is outside the list's box \
+                 {list:?} and was drawn anyway",
+            );
+        }
+    };
+
+    for step in 0..40 {
+        t = scrolled(&mut h, &screen, -400.0, t);
+        bounded(&drawn_rows(&h, &screen), "measuring", step);
+    }
+    for step in 0..40 {
+        t = scrolled(&mut h, &screen, 400.0, t);
+        let rows = drawn_rows(&h, &screen);
+        bounded(&rows, "forward", step);
+        inside(&rows, "forward", step);
+    }
+    for step in 0..40 {
+        t = scrolled(&mut h, &screen, -400.0, t);
+        let rows = drawn_rows(&h, &screen);
+        bounded(&rows, "back", step);
+        inside(&rows, "back", step);
+    }
+}
+
+/// The end the fix had no reason to touch: the row across the *bottom*
+/// edge, where the composer starts. Same rule, other direction -- and the
+/// list opens pinned there, so this is the ordinary state of the screen
+/// rather than a scrolled-to one.
+#[test]
+fn the_row_across_the_bottom_edge_is_drawn() {
+    let (mut h, screen) = opened();
+    let list = list_box(&h, &screen);
+    let mut t = PHONE_FRAME_MS * 2;
+
+    for _ in 0..40 {
+        t = scrolled(&mut h, &screen, -37.0, t);
+        let rows = drawn_rows(&h, &screen);
+        let last = *rows.last().expect("something is on screen");
+        assert!(
+            last.1 >= list.bot_right.y - 0.5,
+            "a band of {:.1}px above the composer belongs to no row",
+            list.bot_right.y - last.1,
+        );
+        assert!(
+            last.0 < list.bot_right.y,
+            "the row across the bottom edge was culled: it starts at {:.1}, below the list's own \
+             {:.1}",
+            last.0,
+            list.bot_right.y,
+        );
+    }
+}
+
+/// Panning past the first row settles *on* it rather than beyond it. The
+/// list is scrolled far further back than the fixture is long, which is
+/// what a hard fling toward the top does; before `List::clamp_to_content`
+/// it stayed wherever that left it -- the phone's "black from the header
+/// down", and a whole blank screen in `iris`'s own
+/// `fling_toward_the_start_stops_at_the_first_row`.
+#[test]
+fn scrolling_past_the_first_row_settles_on_it() {
+    let (mut h, screen) = opened();
+    let list = list_box(&h, &screen);
+    let mut t = PHONE_FRAME_MS * 2;
+
+    for _ in 0..60 {
+        t = scrolled(&mut h, &screen, -100_000.0, t);
+    }
+    // The correction is written at the end of a draw and lands on the
+    // next one.
+    h.frame(t);
+
+    let rows = drawn_rows(&h, &screen);
+    let first = *rows.first().expect("the first row is on screen");
+    assert!(
+        (first.0 - list.top_left.y).abs() < 0.5,
+        "the transcript is parked {:.1}px past its own first row, so the top of the list is blank",
+        first.0 - list.top_left.y,
+    );
+}
+
+/// The same clamp at the other end, which is where Iris met it second
+/// ("you shouldn't be able to scroll below the bottom (or above top)").
+/// The list opens flush with its newest row, so this drags *forward* off
+/// the end of the content and back.
+#[test]
+fn scrolling_past_the_last_row_settles_on_it() {
+    let (mut h, screen) = opened();
+    let list = list_box(&h, &screen);
+    let mut t = PHONE_FRAME_MS * 2;
+
+    for _ in 0..20 {
+        t = scrolled(&mut h, &screen, 100_000.0, t);
+    }
+    h.frame(t);
+
+    let rows = drawn_rows(&h, &screen);
+    let last = *rows.last().expect("the last row is on screen");
+    assert!(
+        (last.1 - list.bot_right.y).abs() < 0.5,
+        "the transcript is parked {:.1}px past its own last row, so the bottom of the list is \
+         blank",
+        list.bot_right.y - last.1,
+    );
+}

iris/transcript-ui/src/composer.rs

@@ -90,16 +90,19 @@ where
     // where the bar actually was.
     // `.scrollable().masked()`: the finger pan (`Scroll::drag`) plus the
     // clip that keeps six lines' worth of a longer message inside the
-    // bar. The mask is the caller's job rather than `Scroll`'s own,
-    // because `Painter::set_mask` allows exactly one mask per widget and
-    // a `Scroll` nested under another masked area would abort on the
-    // second -- `.masked()` is the one mechanism for clipping and this is
-    // one more use of it (tabs-ui's message area is the other).
+    // bar. The mask is the caller's job rather than `Scroll`'s own:
+    // `.masked()` is the one mechanism for clipping and this is one more
+    // use of it (tabs-ui's message area is the other). A `Scroll` nested
+    // under another masked area used to abort here; since 2026-09-07 the
+    // inner mask chains to the outer one (`Mask::parent`) and the content
+    // is clipped by both.
     // Without it the overflow paints *above* the bar, over the
     // transcript: measured before this change at 58px of stray text for a
     // 475px message in a 417px box.
     let content = field
-        .scrollable()
+        // `scrollable_to_end`: what is being typed is at the end, so a
+        // message longer than the six lines shown holds that end.
+        .scrollable_to_end(Axis::Y)
         .masked()
         .pad(dp(FIELD_PAD_DP))
         .max_height(dp(APPROX_LINE_HEIGHT_DP * MAX_LINES + FIELD_PAD_DP * 2.0))

iris/transcript-ui/src/lib.rs

@@ -397,7 +397,7 @@ where
     {
         let selection = selection.clone();
         list.on(
-            CursorSense::Pressing(CursorButton::Left) | CursorSense::Drop,
+            CursorSense::Pressing(CursorButton::Left) | CursorSense::Drop | CursorSense::Cancel,
             move |ctx, rsc| {
                 // Which *block* the finger is over, resolved from its
                 // drawn box rather than from the row's extent -- a row is
@@ -413,7 +413,7 @@ where
                     ctx.data.cursor.pos,
                     ctx.data.sense,
                     ctx.data.cursor.time,
-                    ctx.data.render,
+                    ctx.data.pointer,
                 );
             },
         )
@@ -422,7 +422,13 @@ where
 
     let (composer, composer_bar) = composer::build_composer(rsc);
 
-    let tree = (list.width(rest(1)).height(rest(1)), composer_bar)
+    // `.masked()`: the list draws the row straddling each of its edges in
+    // full (`List::intersects_viewport`), so without a clip the top of
+    // that row is drawn above the list -- through whatever the app put
+    // there, which on the phone is the header bar (docs/IRIS_TODO.md,
+    // 2026-09-07: "code and a paragraph visible behind Run benchmark").
+    // The same clip is what `List::draw` asserts it has.
+    let tree = (list.width(rest(1)).height(rest(1)).masked(), composer_bar)
         .span(Dir::DOWN)
         .add_strong(rsc)
         .any();

iris/transcript-ui/src/row.rs

@@ -238,7 +238,7 @@ where
                     cursor,
                     ctx.data.sense,
                     ctx.data.cursor.time,
-                    ctx.data.render,
+                    ctx.data.pointer,
                 );
                 // A *tap*, decided by the same `DragArbiter` the pan and
                 // the selection are: a gesture that panned the list past
@@ -266,11 +266,14 @@ where
     // delta path knowing anything about frames.
     let framed = match frame {
         BlockFrame::Plain => field.width(rest(1)).add_strong(rsc).any(),
+        // Masked *by the panel*, not inside it: the fence's own rounded
+        // rect is the clip, so content scrolled sideways is cut on the
+        // curve instead of leaving square pixels in the corners
+        // (Iris, 2026-09-07).
         BlockFrame::Verbatim { fill } => field
             .scrollable_on(Axis::X)
-            .masked()
             .pad(dp(FRAME_PAD_DP))
-            .background(rect(fill).radius(dp(FRAME_RADIUS_DP)))
+            .masked_by(rect(fill).radius(dp(FRAME_RADIUS_DP)))
             .width(rest(1))
             .add_strong(rsc)
             .any(),

iris/transcript-ui/src/selection.rs

@@ -242,7 +242,7 @@ impl Selection {
     /// selection, where it is rare and the frame is simply dropped.
     /// `pos_window` is in window space, since a pan's delta has to stay
     /// meaningful even when this frame's event landed on a different row
-    /// than the last one. `render` is `CursorData`'s own field -- what
+    /// than the last one. `pointer` is `CursorData`'s own field -- what
     /// `DragGesture` needs to take pointer capture.
     #[allow(clippy::too_many_arguments)]
     /// Returns what the gesture decided this frame, so a caller with its
@@ -257,20 +257,35 @@ impl Selection {
         pos_window: Vec2,
         sense: CursorSense,
         now: Instant,
-        render: &UiRenderState,
+        pointer: &PointerRequests,
     ) -> GestureOutcome {
-        if matches!(sense, CursorSense::PressStart(_)) {
-            // A fresh touch-down cancels any fling still coasting from
-            // the previous gesture -- `List::fling`'s own doc, and
-            // Android's `Scroller::abortAnimation` for the same reason.
+        // A fresh touch-down cancels any fling still coasting from the
+        // previous gesture -- `List::fling`'s own doc, and Android's
+        // `Scroller::abortAnimation` for the same reason -- and, since
+        // 2026-09-07, *tells the gesture there was one*. A press that
+        // caught moving content is a catch: it pans from this very sample
+        // rather than waiting out `DRAG_SLOP`, which is what pins the
+        // content to the finger instead of leaving it coasting for the
+        // first few frames (Iris's "it fails to stop & snap to where
+        // finger is"). See `DragArbiter::press_start` for Compose's own
+        // mechanism. `starts_press` rather than a `PressStart` test of our
+        // own, so this fires on the recovered-press frames too.
+        let mut press = PressState::default();
+        if self.gesture.starts_press(sense) {
+            press.scrolling = list(ui).is_scrolling();
             list(ui).cancel_fling();
         }
-        let already_selected = self.has_selection(ui);
-        let outcome =
-            self.gesture
-                .handle(render, list.id(), sense, pos_window, now, already_selected);
+        press.already_selected = self.has_selection(ui);
+        let outcome = self
+            .gesture
+            .handle(pointer, list.id(), sense, pos_window, now, press);
         match outcome {
-            GestureOutcome::Undecided => {}
+            // Somebody else took the gesture (a code fence panning
+            // sideways under the finger). Nothing here acted on it, and
+            // `DragGesture` has already forgotten it, so there is nothing
+            // to undo either -- the point is that no tap, fling or
+            // selection follows from a gesture that was never ours.
+            GestureOutcome::Cancelled | GestureOutcome::Undecided => {}
             GestureOutcome::Pan(dy) => list(ui).scroll(-dy),
             GestureOutcome::SelectStart => {
                 if let Some((key, pos_row, size)) = row {
@@ -299,8 +314,16 @@ impl Selection {
                 // The half that actually makes it move -- see
                 // `List::fling`'s doc. Without it the velocity is
                 // computed, stored, and never advanced by anything.
-                let id = list.id();
-                ui.ui_mut().animate(id);
+                //
+                // Only when `fling` actually took it: below Compose's
+                // `|v| <= 1.0`, or with no anchor, there is nothing to
+                // tick, and registering an animation for a widget that is
+                // not animating asks the next frame to find that out
+                // (docs/REVIEW-2026-09-07.md's second nit).
+                if list(ui).is_scrolling() {
+                    let id = list.id();
+                    ui.ui_mut().animate(id);
+                }
             }
             // A tap is nobody's business here -- `row.rs` reads it from
             // the returned outcome and follows a link if one was under
@@ -410,7 +433,7 @@ mod tests {
         sel.register((1, 0), field);
         assert!(sel.gesture.is_idle());
 
-        let render = UiRenderState::new();
+        let pointer = PointerRequests::default();
         let now = Instant::now();
         let size = Vec2::new(100.0, 20.0);
         // No `PressStart` is ever sent -- only the `Pressing` frames a
@@ -422,7 +445,7 @@ mod tests {
             Vec2::new(540.0, 700.0),
             CursorSense::Pressing(CursorButton::Left),
             now,
-            &render,
+            &pointer,
         );
         assert!(
             !sel.gesture.is_idle(),

iris/transcript-ui/src/tool.rs

@@ -107,19 +107,16 @@ const OUTPUT_BYTES: usize = 4096;
 /// Checked at compile time, since both are constants.
 const _: () = assert!(OUTPUT_LINES > 0 && OUTPUT_BYTES > 0);
 
-/// The mark that says a card opens, always drawn from the **monospace**
-/// face.
+/// The size of the disclosure mark, as a font size in dp.
 ///
-/// Not a style choice: `NotoSans-Regular.ttf`, which every other string
-/// here is set in, has no glyph at U+25B8/U+25BE/U+25B4 at all, while
-/// `NotoSansMono-Regular.ttf` does -- read out of both bundled `cmap`s on
-/// 2026-09-06. A missing glyph is the failure nobody who wrote the code
-/// ever sees, so the face that has the glyph is named at the one place the
-/// character is written. IRIS_TODO's "a drawn chevron" has the real fix,
-/// which needs a line primitive iris does not have.
-const CLOSED_MARK: &str = "\u{25b8}";
-const OPEN_MARK: &str = "\u{25be}";
-const UP_MARK: &str = "\u{25b4}";
+/// The mark is a glyph in the icon font iris ships (`iris::icon`, built by
+/// `iris/core/build-icon-font.sh`) -- not a codepoint out of whatever the
+/// platform resolved, which is what U+25B8/25BE/25B4 were until
+/// 2026-09-08: Iris's phone drew an empty box for them and this VM drew a
+/// dot once iris stopped bundling faces. The font's Mono face draws each
+/// glyph inside a full em, so this reads a little larger than the same
+/// number would as text.
+const MARK_DP: f32 = 9.0;
 
 /// Which cards the reader has opened, and which have had their whole
 /// output asked for.
@@ -203,7 +200,7 @@ fn on_tap<Rsc: HasEvents>(
                 ctx.data.cursor.pos,
                 ctx.data.sense,
                 ctx.data.cursor.time,
-                ctx.data.render,
+                ctx.data.pointer,
             );
             if outcome == GestureOutcome::Tapped {
                 f(rsc);
@@ -243,6 +240,13 @@ fn text<Rsc>(content: impl Into<String>, size: f32, color: UiColor) -> TextBuild
         .text_align(Align::LEFT)
 }
 
+/// One of `iris::icon`'s disclosure marks, at [`MARK_DP`] in the muted
+/// colour -- the one place this crate names the icon family, so a second
+/// icon is a second constant rather than a second way of asking.
+fn disclosure<Rsc>(glyph: &'static str) -> TextBuilder<Rsc> {
+    text(glyph, MARK_DP, MUTED_COLOR).family(Family::Icons)
+}
+
 /// A verbatim block: monospace on the surface everything verbatim in this
 /// app sits on, not wrapped, panning sideways on a finger.
 ///
@@ -266,9 +270,9 @@ where
         .wrap(false)
         .add(rsc);
     field
-        .masked()
+        .scrollable_on(Axis::X)
         .pad(dp(RAW_PAD_DP))
-        .background(rect(VERBATIM_BACKGROUND).radius(dp(RAW_RADIUS_DP)))
+        .masked_by(rect(VERBATIM_BACKGROUND).radius(dp(RAW_RADIUS_DP)))
         .width(rest(1))
         .add_strong(rsc)
         .any()
@@ -443,14 +447,9 @@ where
 
     let mut header = Span::empty(Dir::RIGHT).gap(dp(GAP_DP));
     header.push(
-        text(
-            if open { OPEN_MARK } else { CLOSED_MARK },
-            BODY_SIZE,
-            MUTED_COLOR,
-        )
-        .family(Family::Monospace)
-        .add_strong(rsc)
-        .any(),
+        disclosure(if open { icon::OPEN } else { icon::CLOSED })
+            .add_strong(rsc)
+            .any(),
     );
     header.push(
         text(tool.clone(), NAME_SIZE, NAME_COLOR)
@@ -609,9 +608,8 @@ where
 {
     let strong = WidgetPtr::new().add_strong(rsc);
     let ptr = strong.weak();
-    let mark = text(UP_MARK, BODY_SIZE, MUTED_COLOR)
-        .family(Family::Monospace)
-        .center()
+    let mark = disclosure(icon::COLLAPSE)
+        .center_text()
         .width(rest(1))
         .pad(dp(CARD_PAD_DP))
         // Anything shown only as a mark still needs a name: this is what

rigs/gpu-probe/src/bin/teardown.rs

@@ -0,0 +1,102 @@
+//! Why a GPU test segfaults *after* it has passed, and what stops it.
+//!
+//! Measured here 2026-09-08, on this VM's Venus adapter. Destroying the
+//! last `VkInstance` makes the Vulkan loader `dlclose` the ICD; Mesa's
+//! ICD (`/usr/lib/libvulkan_virtio.so`) registers a `pthread_key_create`
+//! destructor pointing into its own text and is not linked `-z nodelete`,
+//! so glibc calls that destructor through unmapped memory when the thread
+//! that used Vulkan exits. libtest runs every `#[test]` on a spawned
+//! thread, which is why it looked like "wgpu crashes on drop": the drop
+//! itself completes, and the crash lands as the thread unwinds.
+//!
+//! The four modes are the experiment, and each is one variable:
+//!
+//! | mode | what it does | 2026-09-08 |
+//! |---|---|---|
+//! | `main` | wgpu instance + device on the main thread, dropped | exits 0 |
+//! | `thread` | the same on a spawned thread | **SIGSEGV** |
+//! | `keep` | the same, but the instance is never dropped | exits 0 |
+//! | `raw` | raw Vulkan (`ash`), instance + device, spawned thread | **SIGSEGV** |
+//!
+//! `raw` is the one that says whose bug it is: no wgpu is involved, so
+//! there is nothing for wgpu or a caller to fix in its drop order. `keep`
+//! is the fix -- hold one `wgpu::Instance` for the process, which is what
+//! wgpu asks for anyway. `iris/tests/mask_sdf.rs` does exactly that.
+//!
+//! `VK_LOADER_DISABLE_DYNAMIC_LIBRARY_UNLOADING=1` also makes every mode
+//! exit cleanly, which is the confirmation that the unload is the
+//! mechanism -- but it is an environment variable every caller would have
+//! to remember, so it belongs in this comment rather than in a script.
+
+use ash::vk;
+use pollster::block_on;
+use wgpu::*;
+
+fn main() {
+    let mode = std::env::args().nth(1).unwrap_or_else(|| "thread".into());
+    let body = match mode.as_str() {
+        "main" => return wgpu_open_and_close(false),
+        "thread" => || wgpu_open_and_close(false),
+        "keep" => || wgpu_open_and_close(true),
+        "raw" => raw_vulkan_open_and_close,
+        other => panic!("unknown mode {other:?}: main | thread | keep | raw"),
+    };
+    std::thread::spawn(body).join().expect("the spawned thread");
+    // Not reached when the thread's exit takes the process with it.
+    eprintln!("thread joined");
+}
+
+/// A wgpu instance and device, opened and closed. `keep_instance` is the
+/// fix under test: everything else still drops normally.
+fn wgpu_open_and_close(keep_instance: bool) {
+    let instance = Instance::default();
+    let adapter =
+        block_on(instance.request_adapter(&RequestAdapterOptions::default())).expect("no adapter");
+    let info = adapter.get_info();
+    eprintln!(
+        "adapter: {} ({:?}, {})",
+        info.name, info.backend, info.driver
+    );
+    let (device, queue) =
+        block_on(adapter.request_device(&DeviceDescriptor::default())).expect("no device");
+
+    drop(queue);
+    drop(device);
+    drop(adapter);
+    if keep_instance {
+        std::mem::forget(instance);
+    } else {
+        drop(instance);
+    }
+    eprintln!("wgpu closed");
+}
+
+/// The same shape with no wgpu in it at all, which is what makes this a
+/// loader/driver bug rather than a wgpu one.
+fn raw_vulkan_open_and_close() {
+    unsafe {
+        let entry = ash::Entry::load().expect("vulkan loader");
+        let app = vk::ApplicationInfo::default().api_version(vk::make_api_version(0, 1, 1, 0));
+        let instance = entry
+            .create_instance(
+                &vk::InstanceCreateInfo::default().application_info(&app),
+                None,
+            )
+            .expect("instance");
+        let phys = instance.enumerate_physical_devices().expect("devices")[0];
+        let priorities = [1.0f32];
+        let queues = [vk::DeviceQueueCreateInfo::default()
+            .queue_family_index(0)
+            .queue_priorities(&priorities)];
+        let device = instance
+            .create_device(
+                phys,
+                &vk::DeviceCreateInfo::default().queue_create_infos(&queues),
+                None,
+            )
+            .expect("device");
+        device.destroy_device(None);
+        instance.destroy_instance(None);
+    }
+    eprintln!("raw vulkan closed");
+}

rigs/gpu-probe/src/main.rs

@@ -79,6 +79,21 @@ fn main() {
             "  {:?} {} ({:?})",
             info.backend, info.name, info.device_type
         );
+        // Compute is a *downlevel* capability, not a feature: Vulkan
+        // grants it to any 1.0 device, and GLES only from ES 3.1. So
+        // "can iris use a compute pass here" is this flag on every
+        // adapter iris might fall back to, not just the preferred one.
+        let down = adapter.get_downlevel_capabilities();
+        let limits = adapter.limits();
+        println!(
+            "    compute shaders: {}  (shader model {:?})",
+            down.flags.contains(DownlevelFlags::COMPUTE_SHADERS),
+            down.shader_model
+        );
+        println!(
+            "    max compute invocations/workgroup: {}, workgroup storage: {} bytes",
+            limits.max_compute_invocations_per_workgroup, limits.max_compute_workgroup_storage_size
+        );
     }
 
     let Some(adapter) = pollster::block_on(instance.request_adapter(&RequestAdapterOptions {

rigs/virtgpu-probe/virtgpu-probe.c

@@ -0,0 +1,133 @@
+// What this VM's virtio-gpu actually offers, asked of the kernel and the
+// driver rather than inferred from the host's qemu command line.
+//
+//     cc -O2 -o virtgpu-probe virtgpu-probe.c -I/usr/include/libdrm -ldrm -lvulkan
+//     ./virtgpu-probe
+//
+// Written 2026-09-08 for the question "Venus keeps causing problems, is
+// there something to do with qemu instead" (docs/RUST.md, "Venus went
+// away for an hour"). Three things, each of which was guessed wrong at
+// least once before being measured:
+//
+//   1. Which capsets the host offers. Capset 6 (DRM) is the "native
+//      context" one -- RADV running in the guest against a passed-through
+//      DRM context instead of Venus proxying every Vulkan call. Whether
+//      it is available is a host-side fact this is the only way to read
+//      from in here.
+//   2. Whether the Vulkan device has compute. It does, and assuming it
+//      did not sent one investigation down the wrong path: the "no
+//      compute" finding on record is about the *Android emulator's*
+//      SwiftShader GL path, a different machine entirely.
+//   3. Whether plain Vulkan device teardown crashes on this adapter. It
+//      does not -- which is what makes wgpu's teardown SIGSEGV wgpu's and
+//      not the driver's, and is the kind of claim that is worthless
+//      without the negative half.
+//
+// C rather than a Rust crate on purpose: two of the three are ioctl and
+// loader questions that a wgpu-shaped rig cannot ask, and this needs to
+// keep working when the thing under suspicion is wgpu itself.
+#include <errno.h>
+#include <fcntl.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+
+#include <drm/virtgpu_drm.h>
+#include <vulkan/vulkan.h>
+#include <xf86drm.h>
+
+static const char *capset_name(int id) {
+    switch (id) {
+    case 1: return "VIRGL";
+    case 2: return "VIRGL2";
+    case 3: return "GFXSTREAM_VULKAN";
+    case 4: return "VENUS";
+    case 5: return "CROSS_DOMAIN";
+    case 6: return "DRM (native context)";
+    default: return "unknown";
+    }
+}
+
+static int capsets(void) {
+    int fd = open("/dev/dri/renderD128", O_RDWR);
+    if (fd < 0) {
+        printf("capsets: cannot open /dev/dri/renderD128: %s\n", strerror(errno));
+        return 1;
+    }
+    struct drm_virtgpu_getparam gp;
+    uint64_t mask = 0;
+    memset(&gp, 0, sizeof gp);
+    gp.param = VIRTGPU_PARAM_SUPPORTED_CAPSET_IDs;
+    gp.value = (uint64_t)(uintptr_t)&mask;
+    int rc = drmIoctl(fd, DRM_IOCTL_VIRTGPU_GETPARAM, &gp);
+    close(fd);
+    if (rc) {
+        // Not a virtio-gpu at all, or a kernel without the param: say
+        // which, rather than printing an empty list that reads like "the
+        // host offers nothing".
+        printf("capsets: SUPPORTED_CAPSET_IDs unavailable: %s\n", strerror(errno));
+        return 1;
+    }
+    printf("capsets: bitmask 0x%llx\n", (unsigned long long)mask);
+    for (int i = 1; i <= 8; i++)
+        if (mask & (1ull << i)) printf("  %d: %s\n", i, capset_name(i));
+    if (!(mask & (1ull << 6)))
+        printf("  (no capset 6: native context needs host-side virglrenderer + qemu support)\n");
+    return 0;
+}
+
+static int devices(void) {
+    VkInstanceCreateInfo ici = {.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO};
+    VkInstance inst;
+    if (vkCreateInstance(&ici, NULL, &inst) != VK_SUCCESS) {
+        printf("vulkan: no instance -- the loader found no usable ICD\n");
+        return 1;
+    }
+    uint32_t n = 0;
+    vkEnumeratePhysicalDevices(inst, &n, NULL);
+    if (n == 0) {
+        // The exact state this VM was in for an hour on 2026-09-08.
+        printf("vulkan: instance ok but ZERO devices -- the host refused a context\n");
+        vkDestroyInstance(inst, NULL);
+        return 1;
+    }
+    VkPhysicalDevice pd[8];
+    if (n > 8) n = 8;
+    vkEnumeratePhysicalDevices(inst, &n, pd);
+    for (uint32_t i = 0; i < n; i++) {
+        VkPhysicalDeviceProperties p;
+        vkGetPhysicalDeviceProperties(pd[i], &p);
+        printf("vulkan: %s (api %u.%u.%u)\n", p.deviceName, VK_VERSION_MAJOR(p.apiVersion),
+               VK_VERSION_MINOR(p.apiVersion), VK_VERSION_PATCH(p.apiVersion));
+        printf("  compute: %u invocations/workgroup, size %u,%u,%u, %u bytes shared\n",
+               p.limits.maxComputeWorkGroupInvocations, p.limits.maxComputeWorkGroupSize[0],
+               p.limits.maxComputeWorkGroupSize[1], p.limits.maxComputeWorkGroupSize[2],
+               p.limits.maxComputeSharedMemorySize);
+    }
+
+    // The negative half of "wgpu's teardown crashes on Venus": five
+    // devices and the instance, created and destroyed the plain way.
+    float prio = 1.0f;
+    VkDeviceQueueCreateInfo q = {.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
+                                 .queueFamilyIndex = 0, .queueCount = 1, .pQueuePriorities = &prio};
+    VkDeviceCreateInfo dci = {.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
+                              .queueCreateInfoCount = 1, .pQueueCreateInfos = &q};
+    for (int i = 0; i < 5; i++) {
+        VkDevice dev;
+        if (vkCreateDevice(pd[0], &dci, NULL, &dev) != VK_SUCCESS) {
+            printf("teardown: device %d could not be created\n", i);
+            return 1;
+        }
+        vkDestroyDevice(dev, NULL);
+    }
+    vkDestroyInstance(inst, NULL);
+    printf("teardown: 5 devices + instance created and destroyed cleanly\n");
+    return 0;
+}
+
+int main(void) {
+    int bad = capsets();
+    bad |= devices();
+    return bad;
+}

server/src/main.rs

@@ -120,6 +120,35 @@ struct Args {
     throwaway_sessions: bool,
 }
 
+/// Where this run keeps its certificates: `--certs`, else the XDG default.
+/// One reader because `--enroll-link` returns before the rest of startup
+/// gets there, and a link minted against a different directory's CA than
+/// the server presents is a handshake failure with nothing on screen
+/// saying why.
+fn certs_dir(certs: &Option<std::path::PathBuf>) -> std::path::PathBuf {
+    certs
+        .clone()
+        .unwrap_or_else(|| config_home("ai-app").join("certs"))
+}
+
+/// The CA every enrollment link carries (`wg_app_link::enroll::ca_param`),
+/// so an app that was not built on this machine can still pin it -- the
+/// iris client is cross-compiled in a VM and run against this server.
+///
+/// `--enroll-link` reads it before the server has been anywhere near
+/// `certs::ensure`, so the file may genuinely not exist yet; the message
+/// says what makes it exist rather than reporting a bare ENOENT.
+fn read_ca(certs_dir: &std::path::Path) -> Result<String> {
+    let path = certs_dir.join("ca.pem");
+    std::fs::read_to_string(&path).with_context(|| {
+        format!(
+            "no CA certificate at {} -- start ai-server once so it generates one, \
+             or point --certs at the directory that has it",
+            path.display()
+        )
+    })
+}
+
 #[tokio::main]
 async fn main() -> Result<()> {
     // Both rustls crypto providers are in the dependency graph (ureq brings
@@ -160,7 +189,13 @@ async fn main() -> Result<()> {
         )?;
         println!(
             "{}",
-            enroll::enrollment_uri("aiapp", bind_ip, args.port, &token)
+            enroll::enrollment_uri(
+                "aiapp",
+                bind_ip,
+                args.port,
+                &token,
+                Some(&read_ca(&certs_dir(&args.certs))?)
+            )?
         );
         return Ok(());
     }
@@ -216,9 +251,7 @@ async fn main() -> Result<()> {
     // Before the interface check below, deliberately: the certificates are also
     // what the phone app embeds at build time, so they need to be obtainable on
     // a machine whose tunnel isn't up yet. The leaf is reissued on every start.
-    let certs_dir = args
-        .certs
-        .unwrap_or_else(|| config_home("ai-app").join("certs"));
+    let certs_dir = certs_dir(&args.certs);
     let certificates = wg_app_link::certs::ensure("ai-app", &certs_dir, &netif::local_addresses())
         .with_context(|| format!("failed to prepare certificates in {}", certs_dir.display()))?;
     if certificates.ca_is_new {
@@ -252,7 +285,13 @@ async fn main() -> Result<()> {
         if rotating {
             tracing::info!("rotated the enrolled token; the previous one is now invalid");
         }
-        enroll::print_enrollment("aiapp", bind_ip, args.port, &token)?;
+        enroll::print_enrollment(
+            "aiapp",
+            bind_ip,
+            args.port,
+            &token,
+            Some(&read_ca(&certs_dir)?),
+        )?;
     }
 
     let tls_config = axum_server::tls_rustls::RustlsConfig::from_pem_file(