perf(alpr): read anchors from the terrain globe and sample outside the paint

The overlay took up to six scene.sampleHeight depth renders inside one
paint, and kept the answers on entities, so every refresh and every
re-enable sampled the same cameras again. On a terrain globe the anchor now
comes from globe.getHeight; answers are kept per camera id until the map
source changes; and what sampling remains runs after the paint, one per
frame.

Co-Authored-By: Claude Fable 5.2 <[email protected]>

docs/CURRENT-STATE.md

@@ -1513,10 +1513,14 @@ tile streamed in; floors resolve before new badges appear (bounded wait), and a
 badge whose cell has not resolved clamps until it does, then moves in place.
 Overlay anchors start on the same floor and write marker properties only on
 change. Once the camera is still and tiles have settled, each overlay anchor is
-replaced once by the rendered surface under it (two depth samples per paint,
-accepted only within -30/+40 m of the floor cell), so badges do not float or
-sink on slopes; wedge corners follow their own floor cells instead of staying
-level with the apex. Unknown bearings never acquire a wedge. The selected
+replaced once by the rendered surface under it (accepted only within -30/+40 m
+of the floor cell), so badges do not float or sink on slopes; wedge corners
+follow their own floor cells instead of staying level with the apex. The
+surface is read outside the overlay paint, one camera per frame: a terrain
+globe answers from its loaded mesh (`globe.getHeight`) and only photoreal
+tiles are depth-sampled (`scene.sampleHeight`). Answers are kept per camera
+until the map source changes, so a refresh or turning the layer off and on
+samples nothing again. Unknown bearings never acquire a wedge. The selected
 card's `DIRECTION` wraps after rounding to whole degrees, so a mapped bearing
 of 359.5° or more reads `0°` rather than `360°`. The row reads
 `N nearby · M on screen`; when every loaded camera is outside the view it adds