List selection visible on every backend; win32 layout propagates up, canvases grow, pickers report their selection

[nicolas-maman]

Four defects found by looking at the screens on Windows, three of them with a cross-backend angle.

A selected list row is visible on every backend

.aui-row-selected was painted by GTK4's stylesheet and, on AppKit, Win32 and UIKit, only reported to the driver: a listbox's selection could be read by a spec and not seen by a user. Each backend now paints its system's selection tint over the row's own ground and takes it off when the class leaves:

• Win32: the user's accent as a tint the labels stay legible on (w32_selection_ground: a quarter over white on a light system, a good third over the dark ground on a dark one), through w32_own_ground so WM_CTLCOLORSTATIC gives the labels the legible text colour automatically;
• AppKit: unemphasizedSelectedContentBackgroundColor in aeui_apply_state_bg (keeps the app's label colours legible in both appearances);
• UIKit: tertiarySystemFillColor, the row's own colour kept and restored.

win32 lays the containers above a changed stack out again

A child added to a nested stack (a listbox's rows go into the each container inside the app's column) changes that stack's natural size, and the column has to place what follows it lower. The layout request stopped at the nested stack: the column kept the container at the height it had when empty, and 200 rows were drawn over the buttons under it (listbox_demo after "Add 200"). w32_request_layout now climbs to every container whose size can change, stopping after a scrollview or a stack pinned in height — the bounds set_hidden's synchronous climb already used. rebuild_bench unchanged (0.88–0.95s on screen, 0.63–0.67s headless).

win32: a canvas grows with its window, and on_resize fires

canvas_create's width and height are the natural size on GTK4 (expand past content size) and AppKit (priority 150); win32 stored them as the pin width()/height()/canvas_size set, so a canvas created 80×80 stayed 80×80 in a 700px window and canvas_on_resize never fired. The natural size is the measure's answer now; a pin is still a pin. tests/resizecb_demo passes on Windows (it was on the "known local failures" list — it was a bug).

win32: a picker's text is its selection; a programmatic selection fires on_change

The driver's text for a picker was always empty (a combo box has no window text to cache); the selected item is cached after every change. CB_SETCURSEL sends no CBN_SELCHANGE, so picker_set_selected invokes the closure itself when the index changed, as GTK4, AppKit and UIKit do. tests/picker passes on Windows (also on the known-failures list — also a bug).

Verified natively

Runtime test; 74 headless specs against main (the remaining local failures — gpuview_demo, tumbling_cube, vg_tooltip — are the GL-less headless run and the AETHER_UI_CHROME/AETHER_UI_TOOLTIP env gates ci.sh sets, not defects); 21 canvas specs; captures of listbox_demo with rows added (buttons back below the list) and a row selected, then moved with Down (the highlight follows).