diff --git a/src/scene/mod.rs b/src/scene/mod.rs
index 2e0a5ca6c..b71239da1 100644
--- a/src/scene/mod.rs
+++ b/src/scene/mod.rs
@@ -408,14 +408,14 @@ pub struct Scene {
     /// Maps block_handle → (entity_handle.value() → sort_handle.value()).
     /// Replaces the O(objects) linear scan inside `wires_for_block()` with an O(1) lookup.
     sort_cache: RefCell<Option<(u64, HashMap<Handle, HashMap<u64, u64>>)>>,
-    /// Cached hatch fill models, keyed by (geometry_epoch, camera_generation).
-    /// The camera term is needed because `synced_hatch_models` now view-culls
-    /// via the quadtree — a key that ignored camera would return stale culled
-    /// lists across pan/zoom.
-    hatch_cache: RefCell<Option<((u64, u64), Arc<Vec<HatchModel>>)>>,
-    /// Cached wipeout fill models, keyed by (geometry_epoch, camera_generation).
-    /// Same reasoning as `hatch_cache`.
-    wipeout_cache: RefCell<Option<((u64, u64), Arc<Vec<HatchModel>>)>>,
+    /// Cached hatch fill models, keyed by geometry_epoch. View culling
+    /// is handled at draw time via `hatch_skip_flags` in the pipeline,
+    /// not at build time — that lets the GPU buffer stay stable across
+    /// pan/zoom while still skipping out-of-view hatches.
+    hatch_cache: RefCell<Option<(u64, Arc<Vec<HatchModel>>)>>,
+    /// Cached wipeout fill models, keyed by geometry_epoch. Same
+    /// reasoning as `hatch_cache`.
+    wipeout_cache: RefCell<Option<(u64, Arc<Vec<HatchModel>>)>>,
     /// Cached image models, keyed by geometry_epoch. Images do their own
     /// per-frame culling in the GPU pipeline (vp_scissor); no camera key
     /// needed here.
@@ -1061,32 +1061,30 @@ impl Scene {
     }
 
     pub(super) fn hatch_models_arc(&self) -> Arc<Vec<HatchModel>> {
-        let key = (self.geometry_epoch, self.camera_generation);
         {
             let cache = self.hatch_cache.borrow();
-            if let Some((cached_key, ref arc)) = *cache {
-                if cached_key == key {
+            if let Some((cached_epoch, ref arc)) = *cache {
+                if cached_epoch == self.geometry_epoch {
                     return Arc::clone(arc);
                 }
             }
         }
         let arc = Arc::new(self.synced_hatch_models());
-        *self.hatch_cache.borrow_mut() = Some((key, Arc::clone(&arc)));
+        *self.hatch_cache.borrow_mut() = Some((self.geometry_epoch, Arc::clone(&arc)));
         arc
     }
 
     pub(super) fn wipeout_models_arc(&self) -> Arc<Vec<HatchModel>> {
-        let key = (self.geometry_epoch, self.camera_generation);
         {
             let cache = self.wipeout_cache.borrow();
-            if let Some((cached_key, ref arc)) = *cache {
-                if cached_key == key {
+            if let Some((cached_epoch, ref arc)) = *cache {
+                if cached_epoch == self.geometry_epoch {
                     return Arc::clone(arc);
                 }
             }
         }
         let arc = Arc::new(self.wipeout_models());
-        *self.wipeout_cache.borrow_mut() = Some((key, Arc::clone(&arc)));
+        *self.wipeout_cache.borrow_mut() = Some((self.geometry_epoch, Arc::clone(&arc)));
         arc
     }
 
@@ -2482,33 +2480,16 @@ impl Scene {
                 .unwrap_or(false)
         };
 
-        // Phase 2.1 — narrow the per-hatch visibility scan to candidates
-        // the quadtree says intersect the view. The map below still does
-        // the heavy lifting (model clone + render_style + selection tint)
-        // so cutting the pre-filter from O(N_hatches) → O(visible_hatches)
-        // is the main win.
-        let view_candidates: Option<std::collections::HashSet<Handle>> =
-            self.view_world_aabb().map(|local_view| {
-                let [ox, oy, _] = self.world_offset;
-                let view_wcs: [f64; 4] = [
-                    local_view[0] as f64 + ox,
-                    local_view[1] as f64 + oy,
-                    local_view[2] as f64 + ox,
-                    local_view[3] as f64 + oy,
-                ];
-                let tree = self.entity_index();
-                tree.query_rect(view_wcs).into_iter().collect()
-            });
-
+        // synced_hatch_models is cached on geometry_epoch and the GPU
+        // upload is keyed on geometry_epoch only (see render.rs — hatch
+        // buffers are "static"). Don't view-cull here; the per-frame
+        // skip flag in compute_hatch_lod handles frustum + sub-pixel
+        // culling at draw time, which keeps the GPU upload set stable
+        // across pan/zoom.
         let mut models: Vec<HatchModel> = self
             .hatches
             .iter()
             .filter(|(&handle, _)| {
-                if let Some(set) = view_candidates.as_ref() {
-                    if !set.contains(&handle) {
-                        return false;
-                    }
-                }
                 let Some(entity) = self.document.get_entity(handle) else {
                     return true;
                 };
@@ -2593,11 +2574,6 @@ impl Scene {
             if fills.is_empty() {
                 continue;
             }
-            if let Some(set) = view_candidates.as_ref() {
-                if !set.contains(&common.handle) {
-                    continue;
-                }
-            }
             if common.invisible || layer_hidden(&common.layer) {
                 continue;
             }
@@ -2643,29 +2619,17 @@ impl Scene {
         } else {
             self.world_offset
         };
-        // Phase 2.1 — quadtree pre-filter for wipeouts in Model layout.
-        let view_candidates: Option<std::collections::HashSet<Handle>> =
-            self.view_world_aabb().map(|local_view| {
-                let [ox, oy, _] = self.world_offset;
-                let view_wcs: [f64; 4] = [
-                    local_view[0] as f64 + ox,
-                    local_view[1] as f64 + oy,
-                    local_view[2] as f64 + ox,
-                    local_view[3] as f64 + oy,
-                ];
-                let tree = self.entity_index();
-                tree.query_rect(view_wcs).into_iter().collect()
-            });
+        // No per-frame view-cull here: GPU wipeout buffer upload is
+        // gated on geometry_epoch only (see render.rs), so any cull at
+        // build time would freeze the visible subset at the geometry
+        // epoch boundary and never re-evaluate as the user pans. The
+        // pipeline's `wipeout_skip_flags` (compute_wipeout_lod) does
+        // the per-frame skip at draw time instead.
         let mut models = Vec::new();
         for entity in self.document.entities() {
             let EntityType::Wipeout(wo) = entity else {
                 continue;
             };
-            if let Some(set) = view_candidates.as_ref() {
-                if !set.contains(&wo.common.handle) {
-                    continue;
-                }
-            }
             if entity.common().invisible {
                 continue;
             }
diff --git a/src/scene/pipeline/mod.rs b/src/scene/pipeline/mod.rs
index 8bc00cd9e..358352a12 100644
--- a/src/scene/pipeline/mod.rs
+++ b/src/scene/pipeline/mod.rs
@@ -65,6 +65,10 @@ pub struct Pipeline {
     hatch_skip_flags: Vec<bool>,
     /// Wipeout fills — rendered after wires in a separate pass.
     gpu_wipeouts: Vec<HatchGpu>,
+    /// Per-wipeout draw-time skip flag (Phase 2.3 frustum cull). `true`
+    /// when the wipeout's projected AABB sits entirely outside the
+    /// viewport rect. Recomputed by `compute_wipeout_lod`.
+    wipeout_skip_flags: Vec<bool>,
     /// Pixel scissor rects [x, y, w, h] for viewport-clipped wipeouts. Recomputed each frame.
     wipeout_pixel_scissors: Vec<Option<[u32; 4]>>,
     gpu_images: Vec<ImageGpu>,
@@ -622,6 +626,7 @@ impl Pipeline {
             hatch_pixel_scissors: vec![],
             hatch_skip_flags: vec![],
             gpu_wipeouts: vec![],
+            wipeout_skip_flags: vec![],
             wipeout_pixel_scissors: vec![],
             gpu_images: vec![],
             image_pixel_scissors: vec![],
@@ -676,7 +681,11 @@ impl Pipeline {
         self.hatch_skip_flags = self
             .gpu_hatches
             .iter()
-            .map(|h| aabb_below_pixel(h.world_aabb, view_proj, clip_w, clip_h, 2.0))
+            .map(|h| {
+                // Phase 3.3 sub-pixel LOD skip OR Phase 2.3 frustum skip.
+                aabb_below_pixel(h.world_aabb, view_proj, clip_w, clip_h, 2.0)
+                    || aabb_offscreen(h.world_aabb, view_proj, clip_w, clip_h)
+            })
             .collect();
     }
 
@@ -689,6 +698,18 @@ impl Pipeline {
             .collect();
     }
 
+    /// Per-frame wipeout frustum-skip flag (Phase 2.3). Mirrors
+    /// `compute_hatch_lod`'s frustum branch. No sub-pixel skip:
+    /// wipeouts mask, so dropping a sub-pixel one wouldn't be wrong
+    /// but also wouldn't pay off — they're usually few.
+    pub fn compute_wipeout_lod(&mut self, view_proj: glam::Mat4, clip_w: u32, clip_h: u32) {
+        self.wipeout_skip_flags = self
+            .gpu_wipeouts
+            .iter()
+            .map(|h| aabb_offscreen(h.world_aabb, view_proj, clip_w, clip_h))
+            .collect();
+    }
+
     /// Recompute pixel scissor rects for viewport-clipped raster images.
     pub fn compute_image_scissors(&mut self, view_proj: glam::Mat4, clip_w: u32, clip_h: u32) {
         self.image_pixel_scissors = self
@@ -1098,6 +1119,9 @@ impl Pipeline {
             pass.set_bind_group(0, &self.uniform_bind_group, &[]);
             let mut scissor_active = false;
             for (i, wipeout) in self.gpu_wipeouts.iter().enumerate() {
+                if self.wipeout_skip_flags.get(i).copied().unwrap_or(false) {
+                    continue;
+                }
                 match self.wipeout_pixel_scissors.get(i) {
                     Some(Some([x, y, w, h])) => {
                         pass.set_scissor_rect(*x, *y, *w, *h);
diff --git a/src/scene/render.rs b/src/scene/render.rs
index 571086860..48fb97fcd 100644
--- a/src/scene/render.rs
+++ b/src/scene/render.rs
@@ -148,6 +148,7 @@ impl shader::Primitive for Primitive {
         pipeline.compute_wipeout_scissors(self.uniforms.view_proj, clip_size.width, clip_size.height);
         pipeline.compute_image_scissors(self.uniforms.view_proj, clip_size.width, clip_size.height);
         pipeline.compute_hatch_lod(self.uniforms.view_proj, clip_size.width, clip_size.height);
+        pipeline.compute_wipeout_lod(self.uniforms.view_proj, clip_size.width, clip_size.height);
         pipeline.compute_mesh_lod(self.uniforms.view_proj, clip_size.width, clip_size.height);
         if self.show_viewcube {
             pipeline.viewcube.upload(