feat(web): bring the app's 3D to the map: view button, bars, pillars, buildings, terrain and rotation

[khagele]

Closes #595

Stacked on #592 (the map on MapLibre). It also carries the app-side commits it builds on, cherry-picked: #586 (terrain, app/src/terrain.js, three commits) and #588 (a bar's tint, pillarTint). Those show in this diff until they merge; the web half is what this PR adds.

The problem

Since #465 the map runs on MapLibre and could pitch, but nothing on the site reached it: rotation and tilt were off, there was no hex-3d, points-3d or buildings-3d layer, no terrain and no control. A visitor who knows the app's 3D view could not get the same read of a cell's signal on the site, and the site is meant to be the superset of the app.

My changes in this PR

The app's 3D, whole (decision log docs/2026-09-06-map-in-3d.md):

• A view button in the map's control corner under the zoom and the compass (mapcore.js addButton, a MapLibre control group of one). Pressed, the camera eases to the app's pitch of 60; the hex cells stand as bars in the tint of their own cell (app: 3D bars don't render the tier colour their own flat cell does #412), the receptions as octagon pillars collapsed where they coincide (app: coincident 3D pillars z-fight against each other when receptions share a position #402), the buildings rise from the hosted style's own vector source. The layer mode stays the segmented control in the filter panel; which layers a view shows is maplayers.js, so Both in 3D draws pillars over flat cells, as the app does.
• 3D carries the terrain. Kasper's call (2026-09-06): the 3D view is the terrain switch. Hillshade and the relief mesh come with it at the exaggeration from Settings (steps 1, 2, 4, 7, 10, default 7, the app's control) and go with it. The app follows the same rule since this morning (feat(app): ship terrain on the AWS DEM, with a terrain button and an exaggeration setting #586, its terrain button went), so terrainPlan is read the same way on both surfaces.
• Rotation and tilt on: right- or ctrl-drag, two-finger turn and tilt. MapLibre's compass draws the bearing and the pitch and puts north back; it is drawn as a mask over currentColor, since the library's needle is a dark image on our dark surface.
• The view travels in the URL: ?view=3d is the layer state and on its own implies the fixed pitch; ?pitch= and ?bearing= are the camera and win when present. A flat north-up view keeps them out. ?exag= travels when off the default and persists like the theme.
• Shared modules whole: signal.js, maplayers.js, pointmarker.js, terrain.js are the app's files in web/, pinned byte for byte in parity.test.js. The web copy of signal.js used to be a hand-kept half.
• --ch-building added to both themes with the app's values, pinned per theme: undefined it was '', which MapLibre rejects and then skips the layer without a word.
• Changelog entry in both copies.

Three e2e tests fixed on the way, all pre-existing and all only visible under a full parallel run. Two were races in the tests: popover.spec.js read a panel's box once right after a resize (now polled, every edge still has to end up inside), and onboarding.spec.js:233 counted the callouts and then waited on the inline panel from a later placement pass (now one atomic snapshot). The third was a bug in the map, in its own commit: a guest deep-linking ?nodepos=1 got the "needs a verified member account" key from the role gate, which also unchecks the box, and the refresh the gate then asks for redrew from the box and took the key back 250 ms later. The key now answers the ask, held past the uncheck until the role can see the layer; the test arms a MutationObserver and forces a draw, so it fails every time without the fix (3 of 3, checked) instead of 2 in 6.

Verification

Suite	Result
web vitest	599 passed (35 files), eslint clean
web e2e --fail-on-flaky-tests	206 passed, five consecutive full runs exit 0 after the three fixes; the base (#592) 3 of 3
new view3d.spec.js	6 tests: the button both ways, the shared link, the fixed tilt, keyboard rotate and the compass, pillars and Both in 3D, the terrain following the view and the exaggeration

Browser, Playwright Chromium at 1100x700 against real OpenFreeMap tiles and the real DEM (Liège, relief and a city): 2D Both draws 40 points and 31 cells; the button eases to pitch 60, the mesh arrives once the DEM tiles are in (ready: true, mesh: true), 40 pillars over flat cells, buildings up, hillshade on; Hex draws 31 bars; three Shift+Arrow turns give bearing=-45 in the URL with the compass turned; Settings shows the exaggeration row with the five steps.

Decisions worth a look

• Pillars are rebuilt on every move, since their footprint floor is a pixel size; the app does it at 1 Hz.
• Age fade is not on the pillars: the site's points carry no time window the way the app's records do.

🤖 Generated with Claude Code