Measured descriptive leaves without per-target layout

gpui_kit_semantics::MeasuredLeafBatch is for descriptions of many caller-owned geometries, not controls. It uses one layout element and a real native Group, publishing native children with the existing A11ySubtreeBuilder authority. Diagnostics go to the installed, window-scoped semantic coordinator. No second registry, renderer, clip engine, input route or retained leaf cache is added.

Supply current-frame window geometry

use gpui::{Bounds, Styled, point, px, size};
use gpui_kit_semantics::{MeasuredLeaf, MeasuredLeafBatch, MeasuredLeafRole};

let targets = MeasuredLeafBatch::new("plot.descriptions", |area, _, _| {
    let measured = Bounds::new(
        area.origin + point(px(3.5), px(4.5)),
        size(area.size.width / 4.0, px(7.5)),
    );
    vec![MeasuredLeaf::new("plot.mark.alpha", MeasuredLeafRole::Image, measured)
        .text("Alpha")
        .value("42 units")
        .selected(true)
        .read_only(true)]
})
.diagnostic_parent("plot") // host separately publishes this diagnostic parent
.absolute()
.size_full();

As with ordinary semantic_in, install semantics once and begin the semantic frame at the window root. Diagnostics require an active SemanticCoordinator arm; native accessibility works without an arm or semantics installation. The measurement callback runs at most once in current prepaint after layout. When both outputs are inactive it does not run, so painting, picking and required state updates must not depend on callback side effects.

Bounds are untransformed logical window coordinates, not batch-relative offsets. Translate local geometry by the callback's area.origin. Keep caller projection precision until converting into Pixels. The callback owns visibility selection and geometry/viewport intersection, including holes, offscreen data, opacity and retiring geometry. The batch cannot infer these from rectangles.

MeasuredLeafRole accepts only Image and Text. Text, literal description, value, read-only state and diagnostic selection are supported. Credential-shaped text, description and value use the same redaction as ordinary semantics. Image/Text native selected remains unset, matching Semantic conventions. There is no general NodeSpec conversion or API accepting focus, actions, ranges, checks, live regions or non-topological relationships; those belong on ordinary controls.

Identity, ancestry and clipping are explicit

• Use stable business-derived leaf ids, never indices into the visible list. Duplicate ids inside a batch panic rather than silently dropping a native leaf. The caller must also keep diagnostic ids unique within its window.
• The batch's stable mounted id plus each leaf key derive its native identity. Reordering, changing properties and removing other leaves keep that identity. Moving the batch to another mounted ancestry changes the native identity.
• Native leaves are direct children of the batch's real Group, beneath the enclosing mounted native parent. diagnostic_parent only supplies diagnostic parentage; it neither adds a diagnostic owner nor reparents native nodes.
• Diagnostics preserve the ordinary transformed-logical unclipped convention. Native bounds additionally receive GPUI's physical scaling and conservative rectangular/rounded ancestor clip. Rounded corners are not exact native shapes.
• While its owner is published, a fully clipped native leaf remains as a zero-area description at the clip edge. A fully clipped nonempty owner omits its native subtree. Omitted leaves, empty output or unmount disappear in the next completed frame; nothing retains prior targets by id.
• Per-leaf Taffy layout is absent, including its device-pixel snapping. Caller fractional geometry can therefore differ from prior absolute Div bounds by layout quantization. Do not round the geometry to imitate the old layout. Compare exact properties/keys, independently calculated geometry and rendered pixels as separate assertions.
• The batch installs no hitboxes, focus or input handlers. Its optional Styled container paint is ordinary GPUI style paint; the leaves themselves paint no primitives and cannot intercept parent pointer capture.

The public native inspection JSON now includes optional bounds with physical x0, y0, x1, y1. These are the committed AccessKit coordinates, not a second calculation. This additive debug field changes no native publication behavior.

Mounted verification and reproduction

cargo test -p gpui-box-kit-semantics
cargo test -p gpui-box --lib
cargo clippy -p gpui-box-kit-semantics -p gpui-box --lib --tests -- -D warnings
cargo run -p xtask -- api generate
cargo run -p xtask -- api check
cargo run -p xtask -- dependencies check
cargo test -p gpui-box-kit-semantics --lib measured_leaf_work_breakdown -- --ignored --nocapture --test-threads=1
cargo test --release -p gpui-box-kit-semantics --lib measured_leaf_work_breakdown -- --ignored --nocapture --test-threads=1

Mounted tests compare ordinary semantic Divs and batches on exactly representable coordinates, then independently verify nonzero current-frame origin/resize, fractional visual transforms, native physical scale, clipped leaves and rounded corner conservatism. With visual scale 1.25 around (−10,−5), native scale 2 and ancestor clip x=10..152.5, y=15..75 logical, one leaf's diagnostic rectangle is (150,56.25,16.25,11.25), while its native physical rectangle is (x0=300,y0=112.5,x1=305,y1=135). Another fully clipped leaf remains native (20,150,20,150), and a zero-size point stays (142.5,112.5,142.5,112.5). A leaf inside the conservative rectangle but outside its rounded corner remains native (22.5,32.5,25,35), proving the intentionally conservative contract.

Additional tests cover business-key reorder/subsets, changed readout properties, redaction, reattachment, empty output, unmount, duplicate-key refusal, independent output activation and diagnostic disarming. A compile-fail test rejects a Button role. A mounted parent still receives pointer capture through the descriptive batch even when native accessibility is active without semantics installation.

For 0, 1 and 10,000 targets, every output combination has identical fixture request-layout/prepaint/paint counts: 3/3/3 (entity/root plus one batch). The ordinary 10k Div fixture uses 10,002/10,002/10,002. Both-off measurement invocations are zero; otherwise exactly the current count is measured. Native and diagnostic leaf counts are independently asserted outside benchmark timing.

The CPU benchmark uses independent windows for every old/batch × native × diagnostic case, closes and verifies empty windows between cases, and records five warm refresh_windows/run_until_parked redraws. Each timed iteration must advance the frame index and exhibit its expected element count. JSON export, snapshot/count assertions and input construction checks are outside the timer; render-time key/property construction remains included. Geometry is a 10-row, source-ordered dense fixture of 10×10 rectangles. No native GPU or serialization occurs inside the timer. Release uses the repository's thin-LTO/codegen-units=1 profile, not an isolated substitute profile.

CPU samples on Linux x86-64, 2026-09-13

Base: 480336c3, Rust 1.97.1. Entries are median microseconds [minimum..maximum] of five warm frames. These are paired fixtures in this orb, not a comparison against another worker's elapsed times. Variation between modes is not additive overhead.

Both benchmark profiles passed all mode counts and every timed frame-index and layout-count assertion. The off/off batch does not construct leaf geometry or properties; active modes necessarily remain O(n) output work and O(n) temporary storage, with one fixed layout element. The fixture has no paintable leaf primitives, so it does not attribute all whole-visualization work to semantics.

Validation passed: 31 semantics library tests (one benchmark ignored), one compile-fail doctest, all 626 GPUI library tests, library/test Clippy with -D warnings, API generation/check, and dependency-authority check. Catalog generation produced current API/developer indexes locally; generated index overwrites are deliberately omitted from the integration patches.

Platform screen-reader acceptance and full renderer/headless integration belong to the coordinator's native lanes and geography integration. This Linux TestPlatform evidence establishes CPU work and committed accessibility data, not device touch, native presentation or FPS. Generated catalogs must be regenerated after integration rather than overwritten with another checkout's generated files.