fix(city3d): fade the city in as one solid layer, dithered, never blended

[NateIsern]

Problem

At a resting zoom between 15.5 and 16.5 (BUILDINGS_3D_FADE), every building and tree surface was blended at the fade's alpha. A pixel mixed the front surface with whatever the city drew behind it, so walls and streets showed through the blocks. The report came from Madrid at z16, night.

Fix: the zoom fade is a dither

The zoom fade now uses the same screen door as a level's crossfade (dithered in tsl/common.ts):

• Each drawn entry keeps its share of the transition's pixels, scaled by the zoom's: setEntryFade(entry, from × opacity, to × opacity).
• Complementary level crossfades still split exactly the zoom's share of the pixels between them.
• The buildings and trees are opaque again and keep their depth writes, so they composite as one solid layer.
• The noise is in screen space, so the pattern stays still while panning.
• The opacity uniform is gone from every material. Web and native share one path; cityTileView included.
• The fade window is still BUILDINGS_3D_FADE.

What it does and doesn't change: a fade still shows the map. At z16, half the screen's pixels draw the city and half draw the map. What can no longer happen is a pixel blending the city's front surface with a surface behind it (inner walls, or streets behind a block).

Why a dither and not an offscreen layer composited at one alpha

The offscreen option costs, on a phone:

• a colour + depth target the size of the screen, about 20 MB at 1080×2400;
• an extra full-screen pass every frame in the window, which matters for bandwidth on tile-based GPUs;
• handing depth back to MapLibre so its later 3D layers still occlude correctly, on web and again on WebGPU.

The dither costs none of these: its variants already exist and are warmed.

Measured

bench:map (production builds, pixel6a profile, real GPU RTX 5090). The machine was shared and loaded, so treat CPU numbers as noise; the GPU process crashed on a third pass.

Scene (in the fade window)	Draws, triangles	GPU p50, main → this PR
eixample z16.3	138, 3.02 M (identical)	4.6–4.8 → 1.6 ms
manhattan z16.2	119–120, 1.43 M (identical)	1.5–2.3 → 1.2 ms

There is no blending, and half the fragments are discarded.

Regression check (fails on main)

check-city3d-visible's nosky leg adds a mid-fade view at z16 (or --skies=fade alone):

• It shoots the city alone, every other layer hidden, over a black and then a white canvas backdrop.
• Each pixel's coverage is 1 − (white − black)/255.
• Under 2% of the city's pixels may be covered more than the fade's share and less than fully, which means a surface behind another shows through.

	Coverage per pixel (share of the view)	Pixels showing a surface behind another
main @52dde03	0: 17%, 0.5: 49%, 0.8: 20%, 0.9: 11%, 1: 3%	37%
This PR	0: 58%, 1: 42%	0.3%

Checks

• bun run lint and tsc pass.
• The frontend tests pass (1710).

🤖 Generated with Claude Code