ObjectHorizonSweep

THE SCHEDULE FOR AN OBJECT-GEOMETRY HORIZON SWEEP. A sensor bolted to a mast does not see a clean horizon either, and the thing in its way is not terrain — it is the platform the sensor is standing on. A radar on a ship's foremast is blanked aft by its own funnels, hangar and after mast; STK draws that as a green surface carved out of the sensor volume and refuses access through it ("Constraining Access Based on Visibility"). TerrainHorizonSweep answers the same question for terrain by walking RANGE along each azimuth and reading heights. That schedule is wrong for an object. Terrain is a height field sampled at a position; a superstructure is a closed solid that a ray either hits or misses, and the sampling primitive available for it is a pick, not a height read. So this module walks ELEVATION, not range: THE MASK IS A THRESHOLD IN ELEVATION. Along one azimuth, a ray fired at high elevation clears the structure and a ray fired at low elevation hits it. Because the platform is a solid attached below/around the mount, that transition is what the mask angle IS. The sweep therefore fires a coarse descending ELEVATION LADDER until the first hit, then BISECTS the bracket between the last miss and that first hit. Each halving buys one bit, so the cost is `ladder + ceil(log2(step / tolerance))` picks per azimuth instead of the hundreds a uniform elevation scan at the same tolerance would need — and picks are expensive (each one is an offscreen render). AZIMUTHS TERMINATE ON THE BOUNDING SPHERE. The platform is entirely inside one sphere. Nothing outside that sphere is the platform, so: reach = |center - mount| + radius (no hit is possible past it) topRise = (center.up - mount.up) + radius (nothing can be higher) bound(d) = atan(topRise / d), d <= reach; -infinity beyond Once that bound has fallen to the mask already accumulated for an azimuth, no further sample on it can matter. Unlike the terrain sweep — whose Everest ceiling keeps open-ocean azimuths alive for hundreds of kilometres — an object's bound collapses within metres, which is exactly why the sweep is affordable at pick prices. Everything here is PURE ARITHMETIC. Nothing in this module touches a Scene, a Model or a Ray: a caller supplies hit/miss verdicts from whatever native sampler it has (`Scene.pickFromRay` is the one OrbPro ships), and this module decides where to sample next, when to stop, and how the resulting table becomes sensor geometry. That separation is what lets the guarantees be asserted without a GL context. The output is deliberately the SAME elevation-by-azimuth table TerrainHorizonSweep feeds: one elevation per azimuth bin, azimuth measured from north and increasing with the index, elevation in radians and positive above the mount's horizontal. ObjectHorizonSweep.directions converts it to the custom-sensor boundary, so an object mask and a terrain mask are interchangeable inputs to a `Sensor` of `Type.CUSTOM`.