1.6.2

Fixed

• A focused WInput no longer turns into a flat grey box when a route transition is pushed in above it. _clearanceBelowCaret runs during build and asked RenderEditable for the pixel-snapped caret rect. That snap ends in _snapToPhysicalPixel → localToGlobal, which walks every ancestor's applyPaintTransform, and RenderFractionalTranslation.applyPaintTransform reads size with no layout guard while RenderBox.size throws rather than asserting. A slide route transition builds exactly that object, and during the frame it is inserted it has not been laid out. Measured on an iPhone in a release build: StateError: Bad state: RenderBox was not laid out: RenderFractionalTranslation, thrown from build, which replaces the subtree with an ErrorWidget. Debug never reaches it, which is why no simulator run and no test had ever seen it.

  The clearance now comes from getEndpointsForSelection, which reads _textPainter, preferredLineHeight and _paintOffset and never leaves the render object. getLocalRectForCaret was the obvious alternative and is the thing being replaced: it is the one that snaps. For a collapsed selection the endpoint answers the line bottom rather than the caret bottom, a sub-pixel difference against a getter already documented as generous by the top padding and compared by its consumer with a 1px threshold; all seven pixel assertions in w_input_scroll_into_view_test.dart are unchanged by it.

  An invalid selection now reads as offset 0 rather than bailing out to the full field height, since offset 0 is where an empty field's caret sits and the full height is the over-reserve this getter exists to avoid. No test covers that branch: instrumenting the getter shows the selection arriving as collapsed(offset: 0) on every frame after a focused field is cleared, so the framework normalises it before build reads it and the path could not be constructed. The regression test is named for what it asserts rather than for that.

  The crash itself has no reproducer either, and the test file says so first: a hand-driven SlideTransition and a real Navigator.push with an autofocusing field both pass with the defect in place, because on a device the focus change and the transition's insertion land in the same frame while in a widget test autofocus fires a frame later, by which time the ancestor is laid out and the unsafe read succeeds. What holds the fix is the stack trace and the API contract. (lib/src/widgets/w_input.dart, test/widgets/w_input_caret_during_transition_test.dart) (#212)

---