fix: exclude the map's gesture detector from semantics

[wcjord]

Fixes #2176.

The mechanism

MapInteractiveViewer builds RawGestureDetector(gestures: _gestures, ...) and leaves excludeFromSemantics at its default false. _gestures contains a TapGestureRecognizer and a LongPressGestureRecognizer, so Flutter wraps the child in _GestureSemantics and publishes SemanticsAction.tap / longPress on a semantics node covering the detector's whole hit area — i.e. the entire map.

Two consequences once the semantics tree is enabled (SemanticsBinding.instance.ensureSemantics(), or automatically when assistive technology is detected):

1. onTap reports the map centre (#2176). A semantics-driven activation doesn't carry a pointer position, so the framework synthesises one at the node's centre and the recognizer fires with kind: unknown. Since the node is the size of the map, that centre is the map centre — which is exactly the reported symptom, and why kind changes from mouse to unknown. Long-press is affected in principle for the same reason; drag recognizers publish no semantics action, which is why panning/zooming were unaffected and only onTap looked broken.

2. On web it blankets any HtmlElementView over the map. Flutter web gives a tappable semantics node pointer-events: all. Platform views are not part of the semantics tree, so a map-sized tappable node sits above any <iframe> / <video> composited over the map and swallows its mouse events — the embed renders and plays but its own controls never respond (a YouTube embed never even shows its control bar, because it never receives mousemove).

I confirmed (2) in a Flutter web app by measuring document.elementFromPoint over a playing embed: the top element was the map's flt-semantics node, viewport-sized, pointer-events: all, every ancestor correctly none. Setting pointer-events: none on that one node — changing nothing else — immediately restored the embed's full controls.

The change

excludeFromSemantics: true on that RawGestureDetector. The map's gestures are pointer gestures, not semantic actions: a map-sized "tap" target isn't a meaningful screen-reader affordance, and its only current effect is to mis-report the tap position. Markers and other children keep their own semantics, so nothing meaningfully activatable loses a node.

Testing

• New regression test in test/gestures/map_interactive_viewer_test.dart: with semantics enabled, a bare FlutterMap publishes no tap/longPress semantics action. It fails on master and passes with this change.
• Full suite green locally (118/118).

Happy to adjust the approach — if you'd rather keep a semantic affordance on the map and fix the position synthesis instead, say so and I'll rework it.

AI disclosure

Per CONTRIBUTING: this was written with AI assistance (Claude). I've reviewed the diagnosis and the diff, ran the tests locally, and take responsibility for the change.