Stop snapshotting the page; own the wheel gesture; one-row phone nav

Selection markers, tab panels, brand changes and theme changes no longer use
a view transition. The page snapshot could not blur what sits behind frosted
glass, so every frosted surface went clear for the length of the change.
nv.transition now measures the marker before and after, and animates the
marker itself (FLIP); arriving content animates on its own; colours fade on
their own surfaces. Verified in Firefox: no view transition is ever started.

Time picker: wheels take every pointer type (mouse, pen, touch) with
touch-action: none, so a phone, a browser's phone mode and a desktop behave
alike. The rubber-band is capped at about 40% of a row (a whole row made the
neighbour look selected). Each wheel has its own frame slot, so one wheel can
no longer swallow another's last scroll event. Switching 24H / AM / PM
happens when the wheel settles, and the hours animate into the new layout.

Carousel arrows are frosted glass, as anything floating over content should be.
At phone width the top bar keeps to one row and its link list scrolls.

GUIDELINES.md records the new cross-engine rules.

check.py, build.py and build.py --check pass.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

GUIDELINES.md

@@ -350,7 +350,7 @@ have watched it reach.**
    exactly and costs no frame budget. Provide a fallback where the feature is missing — either the plain state, or the same curve from script (as the wheels do for Firefox) — and test input differences between engines: wheel notches arrive in pixels in Chromium and in *lines* in Firefox, so normalise `deltaMode` before using `deltaY`.
 10. **Local motion for local change.** Changing a month moves the month, not the
     page. Reserve page-wide view transitions for changes that are about the
-    page: appearance, route, a selection travelling between regions.
+    page: a route change, never appearance or selection — see 14.4.1.
 11. **Restraint wins ties.** When continuity and restraint conflict, animate less.
 
 ---
@@ -537,12 +537,15 @@ tune. This is a first-class goal, bounded by sections 3.1 and 11.
 
 Each of these was a real defect, found by driving Firefox with real input.
 
-1. **Never wrap an appearance change in a page-wide view transition.** The
-   snapshot cannot blur what sits behind frosted glass, so for its whole
-   duration every frosted bar, menu and Glaze pane goes clear and shows the page
-   sharply through itself. Let surfaces transition their own colours. Selection
-   markers still use a short view transition; keep it brief, and prefer a
-   per-element animation for anything inside frosted chrome.
+1. **Never wrap a change in a page-wide view transition.** The snapshot cannot
+   blur what sits behind frosted glass, so for its whole duration every frosted
+   bar, menu and Glaze pane goes clear and shows the page sharply through
+   itself. This applied to theme changes, brand changes *and* selection markers,
+   and was only noticed on the third. Let surfaces transition their own colours;
+   move a selection marker by animating the marker itself (measure, change,
+   animate the difference); animate arriving content on its own. `nv.transition`
+   is built this way; do not reintroduce `document.startViewTransition` around
+   anything on a page that has glass.
 2. **Do not use padding to make a scroll container's last item reachable.**
    Firefox leaves the container's bottom padding out of its scrollable area, so
    the final rows were unreachable and a short list could not scroll at all. Use
@@ -567,10 +570,23 @@ Each of these was a real defect, found by driving Firefox with real input.
 8. **Same-page links travel.** `scroll-behavior: smooth` under
    `prefers-reduced-motion: no-preference`; a jump to a section is the textbook
    change a user could have watched.
-9. **Know your test rig.** A headless browser without a GPU does not render
+9. **Own a gesture completely or not at all.** A control that is half native
+   scrolling and half script behaves differently in every engine, in a desktop
+   browser's phone mode, and on a real phone. The wheels take every pointer —
+   mouse, pen and touch — with `touch-action: none` and run one code path, so
+   what was tested is what runs everywhere.
+10. **An effect must not look like a value.** A rubber-band stretch of a whole
+    row makes the neighbouring row sit under the lens and read as a different
+    selection. Cap give at a fraction of a row (about 40%).
+11. **One animation-frame slot per element, never per component.** A slot shared
+    between a picker's wheels dropped the last scroll event of one wheel and
+    left its rows and value a step behind.
+12. **Know your test rig.** A headless browser without a GPU does not render
    `backdrop-filter`, reports no hover capability (so hover-only controls such
    as scroller arrows disappear), and may not deliver scroll events reliably.
-   Treat those as limits of the rig, not of the system, and confirm them on a
+   It also cannot pan on synthesized touch (not even the page itself), so any
+   behaviour that depends on native touch scrolling cannot be verified there.
+   Treat these as limits of the rig, not of the system, and confirm them on a
    real display.
 
 ### 14.3 Documentation