fix(pipeline): hatch frustum cull tests AABB in wrong space

`HatchGpu.world_aabb` was built by iterating `model.boundary` — but
those vertices are stored as f32 offsets from `model.world_origin`
(precision anchor for UTM-scale WCS). So the AABB was in
*origin-relative* space, not the local space `view_proj` projects
from. `aabb_offscreen` (Phase 2.3) then asked "is this rect near (0,0)
in NDC?" instead of "is this rect at the hatch's actual position?" —
which culled in-view hatches whenever the camera moved off the
origin and rendered out-of-view ones near it.

Sub-pixel LOD (`aabb_below_pixel`, Phase 3.3) survived because it
measures a projected diagonal, which is translation-invariant.

Fixes:
- Add `world_origin` back when building `HatchGpu.world_aabb`, so the
  stored rect is in absolute local space.
- 25 % margin on the viewport rect in `aabb_offscreen`, matching the
  wire path's `view_world_aabb` margin — keeps hatches at the edge
  rendered while panning.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/scene/pipeline/hatch_gpu.rs

@@ -349,11 +349,28 @@ impl HatchGpu {
             ],
         });
 
+        // Boundary vertices are stored as f32 offsets from
+        // `model.world_origin` (so f32 keeps precision at UTM-scale
+        // WCS). The pipeline-level frustum cull (Phase 2.3
+        // `aabb_offscreen`) projects through `view_proj`, which expects
+        // local-space (world_offset-subtracted) coordinates — the same
+        // space as `world_origin`. Add the origin back so `world_aabb`
+        // is in that absolute local space rather than relative to the
+        // anchor. (Sub-pixel LOD via `aabb_below_pixel` is unaffected:
+        // it measures a projected diagonal, invariant under translation.)
+        let ox = model.world_origin[0] as f32;
+        let oy = model.world_origin[1] as f32;
+        let world_aabb = if min_x.is_finite() && min_y.is_finite() {
+            [min_x + ox, min_y + oy, max_x + ox, max_y + oy]
+        } else {
+            [min_x, min_y, max_x, max_y]
+        };
+
         Self {
             vertex_buffer,
             bind_group,
             vp_scissor: model.vp_scissor,
-            world_aabb: [min_x, min_y, max_x, max_y],
+            world_aabb,
             _uniform_buf,
             _boundary_buf,
             _family_buf,

src/scene/pipeline/mod.rs

@@ -1327,9 +1327,16 @@ fn aabb_diagonal_pixels(
 }
 
 /// `true` when the world-XY AABB projects entirely outside the
-/// viewport rect. Phase 2.2 mesh-frustum cull. Equivalent to a 2D
+/// viewport rect (extended by `MARGIN_FRAC` to absorb pan inertia and
-/// bounding-box rejection test in NDC; same 4-corner projection used
+/// avoid edge pop-in). Phase 2.2 mesh-frustum / Phase 2.3 hatch +
-/// for LOD picking, so the extra cost is negligible.
+/// wipeout cull. Equivalent to a 2D bounding-box rejection test in
+/// NDC; uses the same 4-corner projection that LOD picking already
+/// does, so the extra cost is negligible.
+///
+/// IMPORTANT: the AABB must be in the same local space (world_offset
+/// subtracted) that `view_proj` expects. `HatchGpu.world_aabb` rebuilds
+/// the absolute local-space rect from `model.world_origin + boundary
+/// extents` for this reason; meshes already store an absolute rect.
 fn aabb_offscreen(
     aabb: [f32; 4],
     view_proj: glam::Mat4,
@@ -1360,8 +1367,13 @@ fn aabb_offscreen(
         if py < min_py { min_py = py; }
         if py > max_py { max_py = py; }
     }
-    // Reject if projected AABB sits fully to one side of the viewport.
+    // 25% pad on each side — matches `view_world_aabb` (wire path),
-    max_px < 0.0 || min_px > w || max_py < 0.0 || min_py > h
+    // keeps edge geometry rendered while panning before the next
+    // upload reaches the GPU.
+    const MARGIN_FRAC: f32 = 0.25;
+    let mx = w * MARGIN_FRAC;
+    let my = h * MARGIN_FRAC;
+    max_px < -mx || min_px > w + mx || max_py < -my || min_py > h + my
 }
 
 /// Return `true` when the world-XY AABB's screen-space size is below the