fix(chrome): scroll a gesture point into view and dispatch trusted input

getBoundingClientRect keeps reporting elements the growing document pushed below the emulated viewport, and input coordinates are viewport-relative, so a click below the fold was hit-tested to the document root and the step read as an action that landed. every gesture now scrolls the point back in and reports ErrGestureUndelivered when nothing is under it; a selector that names no node reports ErrSelectorMatchedNothing rather than waiting. swipe dispatches a real touch stream instead of page-synthesized pointer events, scroll is a wheel so its distance is exact rather than a fling, and the second tap of a double tap carries click count 2 so dblclick actually fires.

internal/driver/chrome/driver.go

@@ -11,6 +11,7 @@ import (
 	"sync"
 	"time"
 
+	"github.com/chromedp/cdproto/cdp"
 	"github.com/chromedp/cdproto/input"
 	"github.com/chromedp/cdproto/network"
 	"github.com/chromedp/cdproto/page"
@@ -201,14 +202,86 @@ func (d *Driver) Terminate(_ context.Context) error {
 	return nil
 }
 
+// pointInViewScript scrolls a point the caller took from the hierarchy back
+// inside the viewport and reports where to dispatch at, plus whether anything
+// is there to receive it.
+//
+// The emulated viewport is sized once at launch, but getBoundingClientRect goes
+// on reporting elements the growing document has pushed below it, so the two
+// disagree the moment an app adds content. Input coordinates are
+// viewport-relative: a click below the fold is hit-tested to the document root,
+// which delivers it to <html> and never to the element the caller named. No
+// error is raised on any layer, so the step reads as an action that landed and
+// changed nothing.
+//
+// Only a point outside the viewport is moved, so a gesture that already had a
+// reachable target dispatches exactly where it did before.
+const pointInViewScript = `
+(function(x, y) {
+  const root = document.scrollingElement || document.documentElement;
+  let shiftX = 0, shiftY = 0;
+  if (x < 0 || x >= window.innerWidth) shiftX = Math.round(x - window.innerWidth / 2);
+  if (y < 0 || y >= window.innerHeight) shiftY = Math.round(y - window.innerHeight / 2);
+  if (shiftX || shiftY) {
+    const fromX = root.scrollLeft, fromY = root.scrollTop;
+    root.scrollLeft = fromX + shiftX;
+    root.scrollTop = fromY + shiftY;
+    shiftX = root.scrollLeft - fromX;
+    shiftY = root.scrollTop - fromY;
+  }
+  const atX = x - shiftX, atY = y - shiftY;
+  return [atX, atY, document.elementFromPoint(atX, atY) ? 1 : 0];
+})(%d, %d)`
+
+// pointInView returns the point to dispatch a gesture at for the point the
+// caller named, having scrolled it into view. It fails with
+// driver.ErrGestureUndelivered when no scroll can put an element under it.
+func pointInView(runCtx context.Context, x, y int) (int, int, error) {
+	var point [3]int
+	script := fmt.Sprintf(pointInViewScript, x, y)
+	if err := chromedp.Run(runCtx, chromedp.Evaluate(script, &point)); err != nil {
+		return 0, 0, err
+	}
+	if point[2] == 0 {
+		return 0, 0, fmt.Errorf(
+			"%w: (%d,%d)",
+			driver.ErrGestureUndelivered,
+			x,
+			y,
+		)
+	}
+	return point[0], point[1], nil
+}
+
 func (d *Driver) Tap(ctx context.Context, x, y int) error {
 	runCtx, cancel := d.runCtx(ctx)
 	defer cancel()
+	atX, atY, err := pointInView(runCtx, x, y)
+	if err != nil {
+		return err
+	}
 	return chromedp.Run(runCtx,
-		chromedp.MouseClickXY(float64(x), float64(y)),
+		chromedp.MouseClickXY(float64(atX), float64(atY)),
 	)
 }
 
+// requireSelectorMatch reports driver.ErrSelectorMatchedNothing when the
+// selector names no node on the page right now. chromedp.Click waits instead,
+// so without this the caller hears a deadline (or nothing at all) for an action
+// that had no target.
+func requireSelectorMatch(runCtx context.Context, target, selector string) error {
+	var nodes []*cdp.Node
+	if err := chromedp.Run(runCtx,
+		chromedp.Nodes(target, &nodes, chromedp.BySearch, chromedp.AtLeast(0)),
+	); err != nil {
+		return err
+	}
+	if len(nodes) == 0 {
+		return fmt.Errorf("%w: %q", driver.ErrSelectorMatchedNothing, selector)
+	}
+	return nil
+}
+
 func (d *Driver) TapSelector(ctx context.Context, selector string) error {
 	runCtx, cancel := d.runCtx(ctx)
 	defer cancel()
@@ -218,6 +291,9 @@ func (d *Driver) TapSelector(ctx context.Context, selector string) error {
 		// reject it loudly if it isn't a valid CSS selector.
 		target = selector
 	}
+	if err := requireSelectorMatch(runCtx, target, selector); err != nil {
+		return err
+	}
 	if isXPath {
 		return chromedp.Run(runCtx, chromedp.Click(target, chromedp.NodeVisible, chromedp.BySearch))
 	}
@@ -229,16 +305,55 @@ func (d *Driver) TapSelector(ctx context.Context, selector string) error {
 // primitive, so the gesture is two taps with this gap.
 const doubleTapGap = 50 * time.Millisecond
 
+// DoubleTap resolves the point once and dispatches both taps there: resolving
+// per tap would scroll the second one away from the element the first hit.
 func (d *Driver) DoubleTap(ctx context.Context, x, y int) error {
-	return webDoubleTap(ctx, func() error { return d.Tap(ctx, x, y) })
+	runCtx, cancel := d.runCtx(ctx)
+	defer cancel()
+	atX, atY, err := pointInView(runCtx, x, y)
+	if err != nil {
+		return err
+	}
+	return webDoubleTap(ctx, func(clickCount int) error {
+		return chromedp.Run(
+			runCtx,
+			chromedp.MouseClickXY(
+				float64(atX),
+				float64(atY),
+				chromedp.ClickCount(clickCount),
+			),
+		)
+	})
 }
 
 func (d *Driver) DoubleTapSelector(ctx context.Context, selector string) error {
-	return webDoubleTap(ctx, func() error { return d.TapSelector(ctx, selector) })
+	runCtx, cancel := d.runCtx(ctx)
+	defer cancel()
+	target, isXPath, err := TranslateStringSelector(selector)
+	if err != nil {
+		target = selector
+	}
+	if err := requireSelectorMatch(runCtx, target, selector); err != nil {
+		return err
+	}
+	options := []chromedp.QueryOption{chromedp.NodeVisible}
+	if isXPath {
+		options = append(options, chromedp.BySearch)
+	}
+	return webDoubleTap(ctx, func(clickCount int) error {
+		if clickCount < 2 {
+			return chromedp.Run(runCtx, chromedp.Click(target, options...))
+		}
+		return chromedp.Run(runCtx, chromedp.DoubleClick(target, options...))
+	})
 }
 
-func webDoubleTap(ctx context.Context, tap func() error) error {
-	if err := tap(); err != nil {
+// webDoubleTap dispatches the pair a browser reads as one double click. Blink
+// raises dblclick off the click count the second event carries, so two taps
+// that both say "first click" arrive at a dblclick handler as two ordinary
+// clicks and the gesture never happens at all.
+func webDoubleTap(ctx context.Context, tap func(clickCount int) error) error {
+	if err := tap(1); err != nil {
 		return err
 	}
 	timer := time.NewTimer(doubleTapGap)
@@ -248,7 +363,7 @@ func webDoubleTap(ctx context.Context, tap func() error) error {
 		return ctx.Err()
 	case <-timer.C:
 	}
-	return tap()
+	return tap(2)
 }
 
 func (d *Driver) InputText(callerCtx context.Context, text string) error {
@@ -307,31 +422,69 @@ func (d *Driver) EraseText(callerCtx context.Context, _ int) error {
 	)
 }
 
+// Swipe drags a finger across the page as a trusted touch stream. Events
+// synthesized in the page carry isTrusted false: they reach a handler that
+// happens to listen, but never enter the input pipeline that scrolls, honours
+// touch-action or resolves a gesture.
 func (d *Driver) Swipe(ctx context.Context, fromX, fromY, toX, toY int, duration time.Duration) error {
 	runCtx, cancel := d.runCtx(ctx)
 	defer cancel()
-	millis := max(duration.Milliseconds(), 50)
-	script := fmt.Sprintf(`
-(function() {
-  const el = document.elementFromPoint(%d, %d);
-  if (!el) return;
-  const steps = Math.max(1, Math.floor(%d / 16));
-  const dx = (%d - %d) / steps;
-  const dy = (%d - %d) / steps;
-  el.dispatchEvent(new PointerEvent('pointerdown', {clientX: %d, clientY: %d, bubbles: true}));
-  for (let i = 1; i <= steps; i++) {
-    el.dispatchEvent(new PointerEvent('pointermove', {clientX: %d + dx*i, clientY: %d + dy*i, bubbles: true}));
+	atX, atY, err := pointInView(runCtx, fromX, fromY)
+	if err != nil {
+		return err
 	}
-  el.dispatchEvent(new PointerEvent('pointerup', {clientX: %d, clientY: %d, bubbles: true}));
-})();`,
-		fromX, fromY,
-		millis,
-		toX, fromX, toY, fromY,
-		fromX, fromY,
-		fromX, fromY,
-		toX, toY,
+	toX, toY = toX-(fromX-atX), toY-(fromY-atY)
+	steps := max(int(duration.Milliseconds())/16, 1)
+	actions := []chromedp.Action{touchAt(input.TouchStart, atX, atY)}
+	for i := 1; i <= steps; i++ {
+		actions = append(actions, touchAt(input.TouchMove,
+			atX+(toX-atX)*i/steps, atY+(toY-atY)*i/steps))
+	}
+	actions = append(
+		actions,
+		input.DispatchTouchEvent(input.TouchEnd, []*input.TouchPoint{}),
 	)
-	return chromedp.Run(runCtx, chromedp.Evaluate(script, nil))
+	return chromedp.Run(runCtx, actions...)
+}
+
+func touchAt(kind input.TouchType, x, y int) *input.DispatchTouchEventParams {
+	return input.DispatchTouchEvent(
+		kind,
+		[]*input.TouchPoint{{X: float64(x), Y: float64(y)}},
+	)
+}
+
+// Scroll moves the content under the point with a trusted wheel, which is how a
+// browser scrolls.
+//
+// A finger drag scrolls too, but it ends in a fling whose distance follows the
+// release velocity: five identical 240 px drags moved the page 354 to 616 px,
+// so two runs of one seed would explore different screens. A wheel delta lands
+// exactly, and chains from the element under the point out to its scrollable
+// ancestors, which is what scrolling a named container means. The drag stays as
+// Swipe, the verb for the gestures only a finger reaches.
+func (d *Driver) Scroll(
+	ctx context.Context,
+	fromX, fromY, toX, toY int,
+	_ time.Duration,
+) error {
+	runCtx, cancel := d.runCtx(ctx)
+	defer cancel()
+	atX, atY, err := pointInView(runCtx, fromX, fromY)
+	if err != nil {
+		return err
+	}
+	wheel := input.DispatchMouseEvent(input.MouseWheel, float64(atX), float64(atY)).
+		WithDeltaX(float64(fromX - toX)).
+		WithDeltaY(float64(fromY - toY))
+	// The wheel is applied off the CDP round trip, so without this the next
+	// read races it: a step could observe the page before its own scroll, and
+	// the pending scroll then lands during the following one.
+	return chromedp.Run(runCtx, wheel, chromedp.Evaluate(
+		`new Promise(done => requestAnimationFrame(() => requestAnimationFrame(done)))`,
+		nil,
+		awaitPromise,
+	))
 }
 
 func (d *Driver) PressKey(ctx context.Context, key string) error {
@@ -347,6 +500,10 @@ func (d *Driver) PressKey(ctx context.Context, key string) error {
 func (d *Driver) LongPress(ctx context.Context, x, y int) error {
 	runCtx, cancel := d.runCtx(ctx)
 	defer cancel()
+	x, y, err := pointInView(runCtx, x, y)
+	if err != nil {
+		return err
+	}
 	script := fmt.Sprintf(`
 (function() {
   const el = document.elementFromPoint(%d, %d);

internal/driver/chrome/driver_test.go

@@ -9,11 +9,14 @@ import (
 	"net"
 	"net/http"
 	"net/http/httptest"
+	"os"
 	"strings"
 	"testing"
 	"time"
 
 	"github.com/chromedp/chromedp"
+	"github.com/priyanshujain/sanderling/internal/driver"
+	"github.com/priyanshujain/sanderling/internal/hierarchy"
 )
 
 // TestLaunch_ClearStateWipesStorageForTheTargetOrigin covers the CLI's default
@@ -1084,3 +1087,426 @@ func TestHierarchy_CarriesEveryMarkupAttribute(t *testing.T) {
 		t.Errorf(`attrs["data-testid"] = %q, want "account-card"`, got)
 	}
 }
+
+// growingPage serves a page whose content starts shorter than one screen and
+// grows past it when its first button is tapped, which is what a chat, feed or
+// ledger does as a run drives it.
+const growingPage = `<body style="margin:0">
+<button id="grow" style="height:80px">grow</button>
+<div id="status">idle</div>
+<div id="rest"></div>
+<script>
+document.getElementById('grow').addEventListener('click', function() {
+  document.getElementById('rest').innerHTML =
+    '<div style="height:900px"></div>' +
+    '<button id="below" style="height:60px">below</button>';
+  document.getElementById('below').addEventListener('click', function() {
+    document.getElementById('status').textContent = 'below tapped';
+  });
+});
+</script></body>`
+
+// TestTap_ActuatesAnElementBelowTheLaunchViewport pins the invariant the whole
+// web path rests on: an element the driver reports as present and clickable can
+// be acted on. The emulated viewport is sized once at launch, so content the
+// app adds afterwards lies below it while getBoundingClientRect keeps reporting
+// where it is; a click dispatched there is hit-tested to the document root and
+// the element never sees it, with no error anywhere.
+func TestTap_ActuatesAnElementBelowTheLaunchViewport(t *testing.T) {
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(growingPage))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+
+	tapByID := func(id string) {
+		t.Helper()
+		dump, err := d.Hierarchy(ctx)
+		if err != nil {
+			t.Fatalf("Hierarchy: %v", err)
+		}
+		tree, err := hierarchy.Parse(dump)
+		if err != nil {
+			t.Fatalf("Parse: %v", err)
+		}
+		element := tree.Find("id:" + id)
+		if element == nil {
+			t.Fatalf("%s is not in the dump", id)
+		}
+		if !element.Clickable {
+			t.Fatalf("%s is not reported clickable", id)
+		}
+		x, y := element.Bounds.Center()
+		if err := d.Tap(ctx, x, y); err != nil {
+			t.Fatalf("Tap %s at (%d,%d): %v", id, x, y, err)
+		}
+	}
+	tapByID("grow")
+	tapByID("below")
+
+	dump, err := d.Hierarchy(ctx)
+	if err != nil {
+		t.Fatalf("Hierarchy: %v", err)
+	}
+	tree, err := hierarchy.Parse(dump)
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	status := tree.Find("id:status")
+	if status == nil {
+		t.Fatal("status is not in the dump")
+	}
+	if status.Text != "below tapped" {
+		t.Errorf(
+			"status = %q, want %q: the tap reached no element",
+			status.Text,
+			"below tapped",
+		)
+	}
+}
+
+// TestTap_ReportsAGestureThatReachesNoElement covers the half of the same bug
+// that no scrolling can fix: a page that cannot scroll leaves the point out of
+// reach, and the caller has to hear about it rather than read a clean run.
+func TestTap_ReportsAGestureThatReachesNoElement(t *testing.T) {
+	const page = `<body style="margin:0;overflow:hidden">
+<div style="height:80px">top</div>
+<div id="rest" style="height:900px;overflow:hidden"></div>
+<style>html{overflow:hidden}</style></body>`
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(page))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	err := d.Tap(ctx, 100, 5000)
+	if !errors.Is(err, driver.ErrGestureUndelivered) {
+		t.Fatalf(
+			"Tap far below an unscrollable page: err = %v, want ErrGestureUndelivered",
+			err,
+		)
+	}
+}
+
+// TestTapSelector_ReportsASelectorThatMatchesNothing covers the by-selector
+// half of a step that reads as dispatched and did nothing: the node the
+// selector names is not on the page, so the click has no target at all.
+func TestTapSelector_ReportsASelectorThatMatchesNothing(t *testing.T) {
+	const page = `<body style="margin:0">
+<button id="present">here</button>
+<div id="status">none</div>
+<script>
+  document.getElementById('present').addEventListener('click', function () {
+    document.getElementById('status').textContent = 'present tapped';
+  });
+</script></body>`
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(page))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+
+	missCtx, missCancel := context.WithTimeout(ctx, 10*time.Second)
+	defer missCancel()
+	if err := d.TapSelector(missCtx, "id:absent"); !errors.Is(err, driver.ErrSelectorMatchedNothing) {
+		t.Fatalf("TapSelector on an absent element: err = %v, want ErrSelectorMatchedNothing", err)
+	}
+	if err := d.DoubleTapSelector(missCtx, "id:absent"); !errors.Is(err, driver.ErrSelectorMatchedNothing) {
+		t.Fatalf("DoubleTapSelector on an absent element: err = %v, want ErrSelectorMatchedNothing", err)
+	}
+	if err := d.TapSelector(ctx, "id:present"); err != nil {
+		t.Fatalf("TapSelector on the element that is there: %v", err)
+	}
+	dump, err := d.Hierarchy(ctx)
+	if err != nil {
+		t.Fatalf("Hierarchy: %v", err)
+	}
+	tree, err := hierarchy.Parse(dump)
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	if status := tree.Find("id:status"); status == nil || status.Text != "present tapped" {
+		t.Fatalf("status = %+v, want the tap on the present element to have landed", status)
+	}
+}
+
+const doubleClickPage = `<body style="margin:0">
+<div id="target" style="width:200px;height:60px">edit me</div>
+<div id="status">none</div>
+<script>
+  var box = document.getElementById('target');
+  var report = document.getElementById('status');
+  var clicks = 0;
+  box.addEventListener('click', function () { clicks++; });
+  box.addEventListener('dblclick', function () {
+    report.textContent = 'edited after ' + clicks + ' clicks';
+  });
+</script></body>`
+
+func doubleClickStatus(t *testing.T, d *Driver, ctx context.Context) string {
+	t.Helper()
+	dump, err := d.Hierarchy(ctx)
+	if err != nil {
+		t.Fatalf("Hierarchy: %v", err)
+	}
+	tree, err := hierarchy.Parse(dump)
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+	status := tree.Find("id:status")
+	if status == nil {
+		t.Fatal("status is not in the dump")
+	}
+	return status.Text
+}
+
+// TestDoubleTap_ReachesADoubleClickHandler pins a gesture the web driver had no
+// way to deliver. Blink raises dblclick off the click count the second event
+// carries, so a pair that both said "first click" arrived as two ordinary
+// clicks: every double-click affordance on the web (an editable list row, a
+// canvas, a table cell) was unreachable, with no error on any layer.
+func TestDoubleTap_ReachesADoubleClickHandler(t *testing.T) {
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(doubleClickPage))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	if err := d.DoubleTap(ctx, 100, 30); err != nil {
+		t.Fatalf("DoubleTap: %v", err)
+	}
+	if status := doubleClickStatus(t, d, ctx); status != "edited after 2 clicks" {
+		t.Errorf(
+			"status = %q, want %q: the pair never read as one double click",
+			status,
+			"edited after 2 clicks",
+		)
+	}
+}
+
+// TestDoubleTapSelector_ReachesADoubleClickHandler covers the same gesture on
+// the path the runner takes when the action names its target rather than a
+// point.
+func TestDoubleTapSelector_ReachesADoubleClickHandler(t *testing.T) {
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(doubleClickPage))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	if err := d.DoubleTapSelector(ctx, "id:target"); err != nil {
+		t.Fatalf("DoubleTapSelector: %v", err)
+	}
+	if status := doubleClickStatus(t, d, ctx); status != "edited after 2 clicks" {
+		t.Errorf(
+			"status = %q, want %q: the pair never read as one double click",
+			status,
+			"edited after 2 clicks",
+		)
+	}
+}
+
+// gesturesServer serves the fixture both gesture tests measure against: a
+// document taller than the emulated viewport, a scrollable container inside it,
+// and a row that dismisses on a horizontal drag.
+func gesturesServer(t *testing.T) *httptest.Server {
+	t.Helper()
+	body, err := os.ReadFile("testdata/gestures.html")
+	if err != nil {
+		t.Fatal(err)
+	}
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write(body)
+		}),
+	)
+	t.Cleanup(server.Close)
+	return server
+}
+
+// TestScroll_MovesThePageAndAScrollableContainer covers the verb the runner
+// lowers every Scroll action onto. Script-dispatched pointer events are
+// untrusted and a browser never scrolls on them, so the web Scroll used to
+// leave scrollY and every scrollTop exactly where they were while reporting a
+// step that ran. The repeat also pins the distance: a run that scrolls a
+// different amount each time explores differently on the same seed.
+func TestScroll_MovesThePageAndAScrollableContainer(t *testing.T) {
+	server := gesturesServer(t)
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+
+	read := func(expression string) int {
+		t.Helper()
+		var value int
+		if err := chromedp.Run(d.tabCtx, chromedp.Evaluate(expression, &value)); err != nil {
+			t.Fatalf("evaluate %s: %v", expression, err)
+		}
+		return value
+	}
+	const pageScroll = `Math.round(window.scrollY)`
+	const containerScroll = `Math.round(document.getElementById("inner").scrollTop)`
+
+	if before := read(pageScroll); before != 0 {
+		t.Fatalf("scrollY before = %d, want 0", before)
+	}
+	var pageDistances []int
+	for range 3 {
+		if err := chromedp.Run(d.tabCtx, chromedp.Evaluate(`window.scrollTo(0, 0)`, nil)); err != nil {
+			t.Fatalf("reset: %v", err)
+		}
+		if err := d.Scroll(ctx, 195, 500, 195, 260, 300*time.Millisecond); err != nil {
+			t.Fatalf("Scroll: %v", err)
+		}
+		pageDistances = append(pageDistances, read(pageScroll))
+	}
+	if pageDistances[0] <= 0 {
+		t.Errorf(
+			"scrollY after = %d, want > 0: the page never scrolled",
+			pageDistances[0],
+		)
+	}
+	if pageDistances[0] != pageDistances[1] ||
+		pageDistances[1] != pageDistances[2] {
+		t.Errorf(
+			"scrollY over three identical scrolls = %v, want one distance",
+			pageDistances,
+		)
+	}
+
+	if err := chromedp.Run(d.tabCtx, chromedp.Evaluate(`window.scrollTo(0, 0)`, nil)); err != nil {
+		t.Fatalf("reset: %v", err)
+	}
+	containerY := read(
+		`Math.round(document.getElementById("inner").getBoundingClientRect().top + 100)`,
+	)
+	if before := read(containerScroll); before != 0 {
+		t.Fatalf("container scrollTop before = %d, want 0", before)
+	}
+	if err := d.Scroll(ctx, 195, containerY, 195, containerY-120, 300*time.Millisecond); err != nil {
+		t.Fatalf("Scroll in the container: %v", err)
+	}
+	if after := read(containerScroll); after <= 0 {
+		t.Errorf(
+			"container scrollTop after = %d, want > 0: the container never scrolled",
+			after,
+		)
+	}
+	if after := read(pageScroll); after != 0 {
+		t.Errorf(
+			"scrollY = %d, want 0: a scroll inside a container moved the page instead",
+			after,
+		)
+	}
+}
+
+// TestScroll_ReportsAGestureThatReachesNoElement keeps the scroll path on the
+// same footing as the tap path: a page that cannot bring the point into the
+// viewport has to say the gesture reached nothing rather than read as a step
+// that scrolled.
+func TestScroll_ReportsAGestureThatReachesNoElement(t *testing.T) {
+	const page = `<body style="margin:0;overflow:hidden">
+<div style="height:80px">top</div>
+<style>html{overflow:hidden}</style></body>`
+	server := httptest.NewServer(
+		http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+			w.Header().Set("Content-Type", "text/html")
+			_, _ = w.Write([]byte(page))
+		}),
+	)
+	defer server.Close()
+
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	err := d.Scroll(ctx, 100, 5000, 100, 4800, 300*time.Millisecond)
+	if !errors.Is(err, driver.ErrGestureUndelivered) {
+		t.Fatalf(
+			"Scroll far below an unscrollable page: err = %v, want ErrGestureUndelivered",
+			err,
+		)
+	}
+}
+
+// TestSwipe_DeliversATrustedDragToARowHandler covers what the manual says
+// sideways swipes are for. Script-dispatched pointer events carry isTrusted
+// false, which is the mark of a gesture the browser never routed: nothing in
+// the page's own input pipeline saw it, so scrolling, touch-action and any
+// handler that filters on trust behave as if the finger never moved.
+func TestSwipe_DeliversATrustedDragToARowHandler(t *testing.T) {
+	server := gesturesServer(t)
+	d := New()
+	defer d.Terminate(context.Background())
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer cancel()
+	if err := d.Launch(ctx, server.URL, false, nil); err != nil {
+		t.Fatalf("Launch: %v", err)
+	}
+	if err := d.Swipe(ctx, 300, 40, 100, 40, 300*time.Millisecond); err != nil {
+		t.Fatalf("Swipe: %v", err)
+	}
+	var status string
+	if err := chromedp.Run(d.tabCtx, chromedp.Evaluate(
+		`document.getElementById("status").textContent`, &status)); err != nil {
+		t.Fatalf("evaluate: %v", err)
+	}
+	if status != "dismissed left trusted" {
+		t.Errorf("row status = %q, want %q", status, "dismissed left trusted")
+	}
+}

internal/driver/chrome/testdata/gestures.html

@@ -0,0 +1,52 @@
+<!doctype html>
+<html>
+<head>
+<style>
+  :root { --frame-w: 390; --frame-h: 640; }
+  body { margin: 0; font: 16px sans-serif; }
+  #row { height: 80px; background: #eef; touch-action: pan-y; }
+  #inner { height: 200px; overflow-y: scroll; border: 2px solid #333; }
+  #inner div { height: 60px; }
+  #filler div { height: 60px; border-bottom: 1px solid #ccc; }
+</style>
+</head>
+<body>
+<div id="row" data-testid="row">swipe me</div>
+<div id="status" data-testid="status">idle</div>
+<div id="inner" data-testid="inner"></div>
+<div id="filler"></div>
+<script>
+  const inner = document.getElementById('inner');
+  for (let i = 0; i < 30; i++) {
+    const item = document.createElement('div');
+    item.textContent = 'inner ' + i;
+    inner.appendChild(item);
+  }
+  const filler = document.getElementById('filler');
+  for (let i = 0; i < 40; i++) {
+    const item = document.createElement('div');
+    item.textContent = 'page ' + i;
+    filler.appendChild(item);
+  }
+  const row = document.getElementById('row');
+  const status = document.getElementById('status');
+  let startX = 0, startY = 0, tracking = false;
+  row.addEventListener('pointerdown', function(e) {
+    startX = e.clientX;
+    startY = e.clientY;
+    tracking = true;
+    row.setPointerCapture(e.pointerId);
+  });
+  row.addEventListener('pointermove', function(e) {
+    if (!tracking) return;
+    const dx = e.clientX - startX, dy = e.clientY - startY;
+    if (Math.abs(dx) <= 40 || Math.abs(dx) <= Math.abs(dy)) return;
+    tracking = false;
+    status.textContent =
+      (dx < 0 ? 'dismissed left' : 'dismissed right') +
+      (e.isTrusted ? ' trusted' : ' untrusted');
+  });
+  row.addEventListener('pointerup', function() { tracking = false; });
+</script>
+</body>
+</html>