win32: opacity on a child widget fades it (tweened); screenshot reads the screen; curve tests theme-agnostic

[nicolas-maman]

Closes TODO.md's "win32 child-widget opacity (and the capture path that would prove it)" — both halves, with measurements.

What was wrong

• opacity(v) in a build block (text("…") { opacity(0.25) }) reaches aether_ui_set_opacity, which on win32 was top-level-only: the widget stayed opaque while GTK4 and AppKit faded it. style_opacity (the CSS path) had used a layered child window all along, so the two verbs disagreed on one backend.
• Nothing could prove any of it on Windows: the curve tests read "ink" (255 − brightness), which under a dark system theme rises as a label fades, so they reported "the label did not fade" against a fade that was there; and the driver's PrintWindow capture, mid-tween of a layered child, now and then returns a frame without the child.

What changed

Backend (backend/aether_ui_win32.c)

• One opacity path, w32_apply_opacity: model value recorded at once (what the driver reports), presentation tweened when a transition was declared, snapped otherwise — for a top-level and a child alike.
• A child given WS_EX_LAYERED before its top-level has ever been shown never paints (measured: absent from the screen, a forced redraw did not bring it back, the same alpha applied after the show fades it). So the alpha is owed while the window is unmapped and settled from WM_WINDOWPOSCHANGED / SWP_SHOWWINDOW, whichever path showed the window.
• A transition declaration makes the widget layered at full alpha up front (the way a CoreAnimation view is layer-backed before it animates): the switch to a layered window, not the alpha, is what cost the first frames of a tween.
• /screenshot reads the composed frame from the screen when the window is on screen and every corner and the centre belong to it (after raising it without activation); PrintWindow(PW_RENDERFULLCONTENT) otherwise — the headless case, still fine for still frames.

Tests

• test_easing_curve.sh / test_spring_curve.sh: contrast against the ground (mean distance from the frame's median) instead of ink; median-of-three over the samples and the middle of three start captures, so a single dropped frame is not a point on the curve. Header notes rewritten to what was measured.
• tests/win32/win32_runtime_test.c: child_opacity_is_owed_until_the_window_shows.

Measured (Windows 11, window on screen, screen grabs — not the driver)

• { opacity(0.25) } label: 38.9 mean brightness against 60.3 opaque; a { opacity(0.4) } button 39.0 against 50.0.
• 1200ms ease-out fade: label region 59.9 → 46.1 → 36.1 (before, 400ms in, settled).
• Spring: 56.2 → 49.7 → 43.8 → 32.8 at 320ms (past the settled 35.7) → 35.7 — the designed overshoot.
• Both curve tests pass on win32: ease-out early/late travel ratio ~5–6×; spring peak progress 1.10 at 360ms.
• 74 headless driver specs identical to main (same five box-local failures), 16 on-screen specs pass, runtime test passes.

CHANGELOG under [current]; TODO.md section closed.