@gjsify/adwaita-web 0.41.0 → 0.43.0

package/src/elements/row-activation.ts

@@ -0,0 +1,92 @@
+// Tab stop and Enter/Space activation for the rows that libadwaita makes activatable.
+//
+// THE INCIDENT
+//
+// Measured in Firefox on an `<adw-preferences-group>` holding an activatable
+// `<adw-action-row>`, an `<adw-button-row>` and an `<adw-expander-row>`: Tab reached NONE
+// of them. Every one reported `tabIndex` `-1` and `role` `null` while all three activated
+// on click, so the whole group was operable by mouse and invisible to a keyboard. The only
+// things Tab did reach inside a group were the NATIVE controls a row happens to contain —
+// `<adw-switch-row>`'s bare checkbox and `<adw-combo-row>`'s `<select>` — i.e. the row was
+// never the thing that took focus, even where upstream says it is.
+//
+// GTK gives this free and the web renderer does not inherit it: every one of these extends
+// `GtkListBoxRow`, which is focusable by construction and activates on Enter and Space.
+//
+// WHICH ROWS, AND WHY EACH
+//
+// Derived from the C, not chosen — a row that is a tab stop it should not be turns every
+// static label into a stop, and one that is not becomes unreachable:
+//
+//   - `<adw-action-row>` only while `activatable`. A plain row is a label.
+//   - `<adw-button-row>` always: `adw-button-row.c:31` says "AdwButtonRow is always
+//     activatable", and the upstream template hardcodes it.
+//   - `<adw-expander-row>` always, on its HEADER — `adw_expander_row_grab_focus`
+//     (adw-expander-row.c:123) forwards focus to `priv->action_row`, so the header row is
+//     the focus target and the disclosure control at once.
+//   - `<adw-switch-row>` on the ROW, and its slider stops being a tab stop:
+//     `adw_switch_row_init` sets `gtk_widget_set_can_focus (self->slider, FALSE)`
+//     (adw-switch-row.c:159) and `gtk_list_box_row_set_activatable (…, TRUE)` (:160).
+//
+// `<adw-combo-row>` is deliberately NOT here. Its control is a native `<select>` stretched
+// over the row: already a tab stop, already a combobox to the accessibility tree, already
+// arrow-navigable, and already `disabled` at one option or fewer — which is exactly
+// `gtk_list_box_row_set_activatable (…, n_items > 1)` (adw-combo-row.c:194) without a line
+// of our own. Making the row the tab stop would take focus off it and put a SECOND
+// combobox in the tree.
+//
+// The key set is `GtkListBoxRow`'s: Enter and Space. Space is prevented because the
+// browser would otherwise scroll the page under the row the user just activated.
+
+/** What a row needs to become the keyboard's idea of one activatable thing. */
+export interface AdwRowActivationInit {
+    /** The element that becomes the tab stop and receives the keys. */
+    row: HTMLElement;
+    /** Re-read on every {@link AdwRowActivation.sync}: is the row activatable right now? */
+    activatable: () => boolean;
+    /**
+     * What Enter and Space do. Defaults to `row.click()` so a row keeps ONE activation
+     * path — the click handler it already has — rather than growing a second one that can
+     * drift from it.
+     */
+    activate?: () => void;
+}
+
+export interface AdwRowActivation {
+    /** Re-derive the tab stop after anything that can change `activatable()`. */
+    sync(): void;
+}
+
+/**
+ * Make `row` a tab stop while it is activatable, and activate it on Enter and Space.
+ *
+ * The listener is the element's OWN and is bound once, without teardown: it survives a
+ * re-parent and is collected with the element, so there is nothing here for
+ * `check-adwaita-connect-rebind.mjs` to hold. Only bindings that reach outside the element
+ * need releasing and re-arming.
+ */
+export function attachRowActivation(init: AdwRowActivationInit): AdwRowActivation {
+    const { row, activatable } = init;
+    const activate = init.activate ?? (() => row.click());
+
+    row.addEventListener('keydown', (event: KeyboardEvent) => {
+        // A child control owns its own keys: a `<select>`, an entry's `<input>` and a
+        // suffix button all use Enter or Space for something else, and swallowing them
+        // here would break the control to activate the row it sits in.
+        if (event.target !== row) return;
+        if (!activatable()) return;
+        if (event.key !== 'Enter' && event.key !== ' ') return;
+        event.preventDefault();
+        activate();
+    });
+
+    const sync = () => {
+        // `removeAttribute`, not `tabIndex = -1`: a row that is not activatable is not
+        // focusable at all upstream, and writing `-1` would make every static label
+        // programmatically focusable instead.
+        if (activatable()) row.tabIndex = 0;
+        else row.removeAttribute('tabindex');
+    };
+    sync();
+    return { sync };
+}

package/src/elements/swipe-drag.ts

@@ -0,0 +1,450 @@
+// Dragging a swipeable widget with the pointer — the DOM half of `AdwSwipeTracker`.
+//
+// THE INCIDENT
+//
+// Measured in Firefox on a bare `<adw-carousel>` with three 440 px pages: a 300 px mouse
+// drag across it moved `position` 0 → 0 and `scrollLeft` 0 → 0. Nothing. A touchpad
+// two-finger swipe moved 0 → 2 and writing `scrollLeft = 440` moved `position` to 1, so
+// the strip was scrollable and the position feedback worked — the gesture was simply not
+// implemented. Upstream it is not optional: `AdwCarousel` turns it on in `init`
+// (`adw_swipe_tracker_set_allow_mouse_drag (self->tracker, TRUE)`, adw-carousel.c:1209)
+// and exposes it as a public property whose default is TRUE (`allow-mouse-drag`, :1091),
+// documented "If the value is FALSE, dragging is only available on touch."
+//
+// WHAT THIS DOES *NOT* TOUCH, and why that is the compatible answer
+//
+// Touch and touchpad swiping already work, because the widget is a real scroll container
+// with `scroll-snap-type` and the browser owns those gestures — momentum, rubber-band and
+// snapping included. GTK is likewise the platform there. Taking them over to run the
+// tracker on every touchmove would replace a native gesture with an imitation of one, so
+// this module handles the ONE input the browser gives no gesture for: a held mouse button.
+//
+// The consequence to know about is a real divergence and it is ledgered, not hidden: GTK
+// runs its touchpad scroll through the same tracker, so `allow-long-swipes: false` bounds
+// a touchpad flick to the adjacent page there and does not here — measured, a two-finger
+// flick crossed two pages.
+//
+// WHY POINTER EVENTS AND NOT touch/mouse PAIRS
+//
+// `pointerType` is a fact about the event; `'ontouchstart' in window` is a fact about the
+// device. A laptop with a touchscreen is both, so a listener set chosen from the device —
+// the shape `@ribajs/extras`' touch-events service uses, ported from jQuery Mobile — picks
+// one path for a machine that needs both. One set of pointer listeners and a check on
+// `pointerType` cannot get that wrong.
+//
+// THE CLAIM IS A STATE MACHINE, NOT A BOOLEAN, and compressing it into one cost three
+// behaviours. `drag_update_cb` (adw-swipe-tracker.c:640-750) runs in this order, and the
+// order is the point:
+//
+//   1. `append_to_history` — UNCONDITIONAL (:670), ahead of everything below. A move made
+//      before the gesture is claimed still counts toward the velocity.
+//   2. at STATE_NONE, the axis is checked and the gesture is DENIED on the wrong one
+//      (:673) — on the FIRST move, at any distance, not only once 16 px are travelled.
+//   3. at STATE_PENDING, past the threshold, four more refusals before claiming: outside
+//      the swipe area, the wrong axis again (:715), a strip whose first and last snap
+//      point are the same (:720), and overshooting either end with overshoot disabled
+//      (:725-739) — which `AdwCarousel` always has.
+//   4. only at STATE_SCROLLING does `gesture_update` apply the delta (:749).
+//
+// Skipping 3 is what made a backwards drag on page 0 swallow the click of the button
+// under the cursor: the gesture claimed, clamped to the bound so nothing moved, and then
+// suppressed the click anyway. Skipping 1 cost a whole event — with the first record
+// discarded as the clock-starter, a two-move flick measured zero velocity and settled
+// back instead of paging.
+//
+// HOW A DRAG BEATS A CLICK
+//
+// GTK does not blocklist widgets under the pointer: `should_suppress_drag`
+// (adw-swipe-tracker.c:551) only refuses window handles, and a button inside a carousel
+// loses its click because the drag gesture CLAIMS the event sequence once it passes
+// `DRAG_THRESHOLD_DISTANCE`. This mirrors that: every `pointerdown` is tracked, nothing is
+// pre-excluded, and the claim at 16 px is what suppresses the `click` that would otherwise
+// follow. {@link NATIVELY_DRAGGABLE} is the one place a blocklist is unavoidable — a
+// `<select>` or a range slider runs its own drag inside the browser, where there is no
+// sequence to lose the race for.
+
+import { ADW_SWIPE_DRAG_THRESHOLD, SwipeTracker } from '@gjsify/adwaita-core';
+
+/**
+ * Controls that drag themselves, inside the browser, with no sequence for a claim to win.
+ *
+ * GTK needs no such list (see the header); this is the DOM's missing half of gesture
+ * claiming and nothing more. A `<button>` is deliberately ABSENT: losing its click to a
+ * deliberate drag is the upstream behaviour.
+ */
+const NATIVELY_DRAGGABLE = 'select, textarea, input[type="range"], [contenteditable=""], [contenteditable="true"]';
+
+/** Which way the widget pages, and therefore which axis claims a drag. */
+export type AdwSwipeOrientation = 'horizontal' | 'vertical';
+
+export interface AdwSwipeDragInit {
+    /** The element the listeners sit on. Items rebuilt under it stay covered. */
+    host: HTMLElement;
+    orientation: AdwSwipeOrientation;
+    /** Re-read on every press: may a mouse drag start right now? */
+    enabled: () => boolean;
+    /** Page pitch in px. The raw pixel delta is divided by it to reach progress. */
+    pitch: () => number;
+    /** Snap points, ascending, as `adw_swipeable_get_snap_points` returns them. */
+    points: () => readonly number[];
+    /** Progress now — the gesture is seeded from it and bounded by it. */
+    progress: () => number;
+    /** `AdwCarousel:allow-long-swipes`. */
+    allowLongSwipes: () => boolean;
+    /**
+     * `adw_swipeable_get_cancel_progress` — where a CANCELLED gesture goes.
+     *
+     * Not optional by accident: the tracker's own fallback is where the gesture began,
+     * and `AdwCarousel`'s answer is `get_closest_snap_point` — the point nearest where
+     * the finger is NOW (adw-carousel.c:1294-1299). The two differ for exactly the
+     * gesture that matters, one cancelled past the halfway mark: upstream commits to the
+     * page under the cursor, the fallback yanks it back.
+     */
+    cancelProgress: () => number;
+    /** Show this progress. Called on every move that the gesture owns. */
+    onUpdate: (progress: number) => void;
+    /**
+     * Whether a gesture may be claimed at all right now, on top of the refusals this
+     * module makes itself. `AdwCarousel` needs none — its `interactive` and
+     * `allow-mouse-drag` are {@link enabled} — but `is_in_swipe_area` lives here for the
+     * widgets that implement `get_swipe_area`, which the carousel does not.
+     */
+    claimable?: (start: { x: number; y: number }) => boolean;
+    /** Settle on this progress. `velocity` is raw px/ms, for an animation duration. */
+    onEnd: (progress: number, velocity: number) => void;
+}
+
+/** What a claimed gesture had to override on the host, so releasing it can put it back. */
+interface FrozenStyles {
+    scrollSnapType: string;
+    scrollBehavior: string;
+    userSelect: string;
+}
+
+/**
+ * Make `host` draggable with a mouse, deciding where it lands the way `AdwSwipeTracker`
+ * does.
+ *
+ * The listener is the element's own and is bound once, without teardown: it survives a
+ * re-parent and is collected with the element.
+ */
+/** What a caller gets back, so it can tell whether a gesture owns the widget. */
+export interface AdwSwipeDrag {
+    /**
+     * Whether a gesture is currently claimed.
+     *
+     * A widget needs this to keep its OTHER inputs out of the offset while a drag owns
+     * it. Measured on `<adw-carousel>`: one wheel notch during a claimed drag ran the
+     * whole wheel path and jumped `scrollLeft` 80 → 440 in a frame, which the next move
+     * then yanked back. Upstream's `scroll_cb` is ungated too, but there the external
+     * request is a spring the next frame quietly overrides — here `claim()` sets
+     * `scroll-behavior: auto`, which turns it into an instant jump.
+     */
+    readonly dragging: boolean;
+}
+
+export function attachSwipeDrag(init: AdwSwipeDragInit): AdwSwipeDrag {
+    const tracker = new SwipeTracker({
+        points: init.points,
+        allowLongSwipes: init.allowLongSwipes,
+        cancelProgress: init.cancelProgress,
+    });
+
+    let pointerId: number | null = null;
+    let startX = 0;
+    let startY = 0;
+    let previousOffset = 0;
+    /** Where the pointer is now, for the second half of the swipe-area test. */
+    let lastX = 0;
+    let lastY = 0;
+    /** `ADW_SWIPE_TRACKER_STATE_*`: nothing pressed, a press that may become one, ours. */
+    let state: 'none' | 'pending' | 'scrolling' = 'none';
+    let frozen: FrozenStyles | null = null;
+    /**
+     * Set when a claimed drag ends, cleared by the click it eats — or by the next press,
+     * because a drag that ended OUTSIDE the host produces no click at all and a one-shot
+     * listener would then sit armed and swallow the next real one.
+     */
+    let suppressNextClick = false;
+
+    const axisOffset = (dx: number, dy: number) => (init.orientation === 'vertical' ? dy : dx);
+
+    /**
+     * Dragging LEFT must advance, so the offset is negated — `drag_update_cb`'s
+     * `offset = -offset`.
+     *
+     * LTR ONLY, deliberately. Upstream flips this sign for RTL through
+     * `adw_swipe_tracker_set_reversed` (`adw-carousel.c:455-465`), and taking a `reversed`
+     * flag here would be an afternoon's work — but the web carousel does not work in RTL
+     * at ALL yet: its offset model is `scrollLeft = position * distance`, and an RTL
+     * scroll container counts `scrollLeft` DOWN from 0, so even `scrollToPage(1)` moves
+     * nothing (measured). A reversed branch would compute a correct progress and write it
+     * to a container that ignores it. `status/open-todos.md` carries the whole gap; the
+     * flag belongs in the change that can test it.
+     */
+    const progressOffset = (dx: number, dy: number) => -axisOffset(dx, dy);
+
+    const release = () => {
+        // Listeners off and state cleared BEFORE the capture is released, because
+        // `releasePointerCapture` fires `lostpointercapture` synchronously and that is
+        // one of the endings this listens for — doing it in the other order re-enters
+        // `finish` in the middle of itself.
+        const captured = pointerId;
+        pointerId = null;
+        state = 'none';
+        tracker.reset();
+        window.removeEventListener('pointermove', onMove, true);
+        window.removeEventListener('pointerup', onFinish, true);
+        window.removeEventListener('pointercancel', onCancel, true);
+        window.removeEventListener('lostpointercapture', onLostCapture, true);
+
+        if (captured !== null && init.host.hasPointerCapture(captured)) {
+            init.host.releasePointerCapture(captured);
+        }
+        if (frozen !== null) {
+            init.host.style.scrollSnapType = frozen.scrollSnapType;
+            init.host.style.scrollBehavior = frozen.scrollBehavior;
+            init.host.style.userSelect = frozen.userSelect;
+            frozen = null;
+        }
+    };
+
+    /**
+     * `:720-739` — the refusals that happen at the threshold, after the axis and before
+     * the claim. Returns whether the gesture may be claimed at all.
+     */
+    const claimAllowed = (offset: number): boolean => {
+        const points = init.points();
+        if (points.length === 0) return false;
+        const first = points[0];
+        const last = points[points.length - 1];
+        // A strip with one snap point has nowhere to go (:720). The C cannot reach an
+        // EMPTY one — `adw_carousel_get_snap_points` returns `MAX (n, 1)` points
+        // (adw-carousel.c:1269) — so the length test above stands in for it.
+        if (Math.abs(first - last) < Number.EPSILON) return false;
+
+        const progress = init.progress();
+        // Overshooting an end with overshoot disabled, which AdwCarousel always is
+        // (:725-739). Without this a backwards drag on the first page claims, moves
+        // nothing, and still eats the click of whatever it started on.
+        const overshootingLower = offset < 0 && (Math.abs(progress - first) < Number.EPSILON || progress < first);
+        const overshootingUpper = offset > 0 && (Math.abs(progress - last) < Number.EPSILON || progress > last);
+        if (overshootingLower || overshootingUpper) return false;
+
+        // BOTH ends, as `is_in_swipe_area (start) || is_in_swipe_area (start + offset)`
+        // does (:709-712): a gesture that began outside the grabbable strip and has
+        // arrived inside it is still allowed.
+        if (init.claimable === undefined) return true;
+        return init.claimable({ x: startX, y: startY }) || init.claimable({ x: lastX, y: lastY });
+    };
+
+    const claim = () => {
+        state = 'scrolling';
+        // Captured HERE and not on the press. With capture set at `pointerdown`, the
+        // `click` that follows is dispatched at the CAPTURING element — so it never
+        // reaches the button the user actually pressed, and measured, every button inside
+        // a carousel stopped working. From the claim on it is harmless, because a claimed
+        // gesture suppresses that click on purpose.
+        if (pointerId !== null) init.host.setPointerCapture(pointerId);
+        // Snap and smooth scrolling both fight a drag: with them on, every write to
+        // `scrollLeft` is a request the browser re-animates and re-snaps, so the strip
+        // lags the cursor and lands somewhere the tracker did not choose. They go back
+        // on before the settle, which is what makes the release animate.
+        frozen = {
+            scrollSnapType: init.host.style.scrollSnapType,
+            scrollBehavior: init.host.style.scrollBehavior,
+            userSelect: init.host.style.userSelect,
+        };
+        init.host.style.scrollSnapType = 'none';
+        init.host.style.scrollBehavior = 'auto';
+        init.host.style.userSelect = 'none';
+    };
+
+    /**
+     * MOVE AND UP ARE LISTENED FOR ON THE WINDOW, and this is not belt-and-braces.
+     *
+     * Measured: with the listeners on the host and `setPointerCapture` alone, a drag that
+     * travelled past the left edge of the viewport never delivered `pointerup` to the
+     * track at all — `scroll-snap-type` stayed `none` and the strip sat at `scrollLeft`
+     * 387 of a 440 px page, forever, with the gesture still believing it was in flight.
+     * That is worse than the missing feature was. Capture is kept as well, because it is
+     * what keeps the moves targeted while the pointer is still inside.
+     */
+    const arm = () => {
+        window.addEventListener('pointermove', onMove, true);
+        window.addEventListener('pointerup', onFinish, true);
+        window.addEventListener('pointercancel', onCancel, true);
+        // The ending of last resort, for a CLAIMED gesture whose `pointerup` this never
+        // sees — the window losing focus mid-drag, an OS-level cancel. Without it such a
+        // gesture leaves the strip stopped mid-page with snapping switched off, which was
+        // measured and is worse than the feature being absent.
+        //
+        // It does NOT cover the host being removed from the document, which an earlier
+        // draft of this comment claimed: the implicit release dispatches
+        // `lostpointercapture` at the now-detached element, and an event dispatched in a
+        // detached tree never reaches `window`. Measured — the styles stayed frozen until
+        // a later `pointerup` cleaned up. The button check in `onMove` is the net that
+        // actually spans both, since a removed host stops delivering moves anyway.
+        window.addEventListener('lostpointercapture', onLostCapture, true);
+    };
+
+    init.host.addEventListener('pointerdown', (event: PointerEvent) => {
+        // FIRST, before any guard below can return. A cancelled gesture produces no click
+        // to consume the flag, so leaving it set until the next SUCCESSFUL press meant the
+        // next click was eaten instead — measured: cancel a drag, then click a `<textarea>`
+        // inside the carousel (which `NATIVELY_DRAGGABLE` refuses to drag) and the click
+        // never arrived.
+        suppressNextClick = false;
+
+        // A gesture still open under the SAME pointer means its end was never seen: close
+        // it, or one lost `pointerup` disables dragging for good. A DIFFERENT pointer is
+        // the opposite case and upstream is explicit about it — `drag_begin_cb` denies the
+        // NEW sequence and keeps the one in flight (adw-swipe-tracker.c:617-620). Tearing
+        // the live gesture down instead left the strip rubber-banding back with no
+        // `onEnd` and no tracker decision at all.
+        if (pointerId !== null) {
+            if (event.pointerId !== pointerId) return;
+            release();
+        }
+        // Mouse only. Touch and touchpad are the browser's own gestures, and a pen
+        // behaves like touch on every platform that has one.
+        if (event.pointerType !== 'mouse' || event.button !== 0) return;
+        if (!init.enabled()) return;
+        if ((event.target as Element | null)?.closest(NATIVELY_DRAGGABLE) !== null) return;
+
+        pointerId = event.pointerId;
+        startX = event.clientX;
+        startY = event.clientY;
+        lastX = event.clientX;
+        lastY = event.clientY;
+        previousOffset = 0;
+        state = 'pending';
+        // No capture yet — see `claim`. The moves that DECIDE the claim come from the
+        // window listeners instead, which is why they are armed here and not there: a
+        // 16 px threshold near an edge is otherwise unreachable.
+        arm();
+    });
+
+    function onMove(event: PointerEvent): void {
+        if (pointerId !== event.pointerId) return;
+
+        lastX = event.clientX;
+        lastY = event.clientY;
+        const dx = event.clientX - startX;
+        const dy = event.clientY - startY;
+        const offset = progressOffset(dx, dy);
+        const delta = offset - previousOffset;
+        previousOffset = offset;
+        // TRUNCATED to whole milliseconds, because the C's clock is a `guint32` and its
+        // `first_time == last_time` guard is a real low-pass filter: two events inside
+        // one millisecond yield NO velocity there. Left as a float, that guard is dead
+        // and two coalesced moves 0.1 ms apart read as tens of pages per second.
+        const time = Math.trunc(event.timeStamp);
+        // UNCONDITIONAL, and ahead of every refusal below — `append_to_history` at :670.
+        tracker.record(delta, time);
+
+        // NO BUTTON HELD means the press this is tracking ended somewhere unseen. Before
+        // the claim there is no pointer capture and therefore no `lostpointercapture`, so
+        // this is the only net that window: measured, one lost `pointerup` left the strip
+        // FOLLOWING a cursor with nothing pressed, snapping off, until the next click
+        // anywhere. Checked for every state, since a claimed gesture can lose its up too.
+        if ((event.buttons & 1) === 0) {
+            release();
+            return;
+        }
+
+        if (state === 'pending') {
+            // Whichever axis moved MORE decides whose gesture this is
+            // (`is_offset_vertical`, :661). Tested here and ONLY here: the C tests it at
+            // STATE_NONE, on the first move at any distance (:673), and again inside the
+            // threshold block (:715) — never at STATE_SCROLLING. One test in the pending
+            // branch covers both, because `dx`/`dy` are cumulative from the press.
+            //
+            // Hoisting it above this branch, as a first draft did, re-tested it on every
+            // claimed move — and `dx`/`dy` being cumulative means a drag that comes BACK
+            // toward its start ("changed my mind, with a little vertical wander") flips
+            // the dominant axis and cancelled itself mid-gesture.
+            const wantsVertical = Math.abs(dy) > Math.abs(dx);
+            if (wantsVertical !== (init.orientation === 'vertical')) {
+                release();
+                return;
+            }
+            if (Math.hypot(dx, dy) < ADW_SWIPE_DRAG_THRESHOLD) return;
+            if (!claimAllowed(offset)) {
+                release();
+                return;
+            }
+            // `gesture_prepare` is the C's first-move step and sets `initial_progress`
+            // (:202-204). It runs here rather than at `pointerdown` because until a move
+            // arrives there is no direction, and until the axis passes there is no
+            // gesture to prepare.
+            tracker.prepare(init.progress());
+            claim();
+            // From the claim, not from the press: the threshold travel is what BOUGHT
+            // the gesture and must not also be applied as progress (`prev_offset = offset`
+            // at :744), or the content jumps 16 px the instant it follows the cursor.
+            previousOffset = offset;
+            return;
+        }
+
+        if (state !== 'scrolling') return;
+        // The standard signal that this move is consumed. Deliberately NOT credited with
+        // the two things it usually is: text selection is prevented by the `user-select`
+        // in `claim`, and a native image or link drag begins from `dragstart` BEFORE the
+        // claim, so this line has never run by then — measured, an image dragged inside a
+        // page is never refused here. Kept because consuming the move is what it is for,
+        // not because a specific default was observed to need it.
+        event.preventDefault();
+        init.onUpdate(tracker.advance(delta, init.pitch()));
+    }
+
+    function finish(event: PointerEvent, cancelled: boolean): void {
+        if (pointerId !== event.pointerId) return;
+        if (state !== 'scrolling') {
+            release();
+            return;
+        }
+        if (cancelled) tracker.cancel();
+        const { to, velocity } = tracker.end(Math.trunc(event.timeStamp));
+        release();
+        // A deliberate drag must not also activate whatever it ended on — the DOM's
+        // stand-in for GTK's claimed event sequence.
+        suppressNextClick = true;
+        init.onEnd(to, velocity);
+    }
+
+    function onFinish(event: PointerEvent): void {
+        finish(event, false);
+    }
+
+    function onCancel(event: PointerEvent): void {
+        finish(event, true);
+    }
+
+    function onLostCapture(event: PointerEvent): void {
+        // Settled, not cancelled: the pointer is gone, so whatever the user did they are
+        // done doing it, and dropping them back where they started would undo a drag
+        // they can see on screen.
+        finish(event, false);
+    }
+
+    // CAPTURE phase and permanent, rather than a one-shot armed at the end of a drag: a
+    // listener that is only ever removed by the click it is waiting for outlives a drag
+    // that produced none.
+    init.host.addEventListener(
+        'click',
+        (event: Event) => {
+            if (!suppressNextClick) return;
+            suppressNextClick = false;
+            event.stopPropagation();
+            event.preventDefault();
+        },
+        { capture: true },
+    );
+
+    return {
+        get dragging(): boolean {
+            return state === 'scrolling';
+        },
+    };
+}

package/src/empty-sections.spec.ts

@@ -0,0 +1,81 @@
+// The `hidden` derivation itself, driven in both directions.
+//
+// `connect-lifecycle.spec.ts` drives the WIDGETS: it appends into every container a
+// consumer can reach and requires the child to be laid out. That half alone does not
+// hold the invariant — replacing the whole derivation with `section.hidden = false`
+// left the full adwaita-web browser suite green, because nothing anywhere asserted that
+// an EMPTY section is hidden or that it goes back to hidden when its last child leaves.
+// An empty bar that re-enters the flow is what that mutant ships.
+//
+// Driven against the helper rather than a widget on purpose: the two elements whose
+// sections are private fields (`<adw-alert-dialog>`, `<adw-status-page>`) have no
+// container getter, so the widget driver cannot see them, and giving them one to satisfy
+// a test would be public API invented for the test.
+import { describe, expect, it } from '@gjsify/unit';
+
+import { bindEmptySections } from './empty-sections.js';
+
+/** A section mounted in the document, since `hidden` is only observable where there is layout. */
+function section(): HTMLElement {
+    const el = document.createElement('div');
+    document.body.appendChild(el);
+    return el;
+}
+
+/** One microtask checkpoint — where a MutationObserver callback lands. */
+const settle = () => Promise.resolve();
+
+export const AdwEmptySectionsTest = async () => {
+    await describe('bindEmptySections', async () => {
+        await it('hides a section that starts empty', async () => {
+            const el = section();
+            bindEmptySections(el);
+            const hidden = el.hidden;
+            el.remove();
+            expect(`empty: hidden=${hidden}`).toBe('empty: hidden=true');
+        });
+
+        await it('shows it when a child lands and hides it again when the last one leaves', async () => {
+            const el = section();
+            bindEmptySections(el);
+
+            const child = document.createElement('span');
+            el.appendChild(child);
+            await settle();
+            const filled = el.hidden;
+
+            child.remove();
+            await settle();
+            const emptied = el.hidden;
+
+            el.remove();
+            expect(`filled: hidden=${filled} | emptied: hidden=${emptied}`).toBe(
+                'filled: hidden=false | emptied: hidden=true',
+            );
+        });
+
+        await it('counts element children, not child nodes', async () => {
+            const el = section();
+            bindEmptySections(el);
+            // The whitespace an indented `<div>\n  </div>` carries is a text node, and a
+            // section full of indentation is still an empty section. This is also why the
+            // CSS `:empty` that this looks like cannot express the rule.
+            el.appendChild(document.createTextNode('\n    '));
+            await settle();
+            const hidden = el.hidden;
+            el.remove();
+            expect(`whitespace only: hidden=${hidden}`).toBe('whitespace only: hidden=true');
+        });
+
+        await it('derives each section it is handed independently', async () => {
+            const empty = section();
+            const filled = section();
+            filled.appendChild(document.createElement('span'));
+            bindEmptySections(empty, filled);
+            const report = `empty=${empty.hidden} filled=${filled.hidden}`;
+            empty.remove();
+            filled.remove();
+            expect(report).toBe('empty=true filled=false');
+        });
+    });
+};