Cheaper id-map pass, and buildings that read right from below

[apresmoi]

Three fixes to how @glyphcss/maps draws a city: one for speed, two for what a building actually looks like when you stand next to it.

The id-map pass skips what it can't paint. computeOcclusionIds rasterised every polygon in the scene, including the ones Polygon.hidden had already culled and the cull runs the paint loop itself skips. A hidden polygon that claims a cell is worse than a wasted cycle — it punches every other layer out of a cell nothing then fills. computeOcclusionIds 6.84 → 4.69 ms on a globe overview and 6.95 → 0.68 ms at street level, where 85% of the walls are culled; on /maps with the probe, 32.23 → 28.51 ms/frame, 31.0 → 35.0 fps.

A raised band gets an underside. glyphMapVectorMesh emitted a cap and walls and no floor, so a band standing off its own ground was a hollow shell from below: the cap is back-face culled, the far walls go with it, and you look straight through. The Berlin Fernsehturm is eight stacked render_min_height bands and seven of them were open. A raised band now gets a floor wound opposite its own cap face; a building standing on the ground still gets none, since that floor is buried.

A facade is refused on a mast. Window bays are a claim that the wall is a storey stack, and no tile schema can tell us whether it is — the live OpenFreeMap building layer carries exactly render_height, render_min_height, hide_3d and colour, so building=tower never arrives. The shape separates instead: across three real z14 city tiles every genuine building is at or under 7.8 times taller than its footprint is wide and the first mast is 8.5, so GLYPH_MAP_FACADE_MAX_SLENDERNESS is 8. Houston's towers keep their windows, the Fernsehturm loses its.

Reasoning and measurements in docs/design/maps.md and docs/design/performance.md, contracts in AGENTS.md. See the checks on this PR for pnpm test and pnpm build.