Merge PR #1001: GPU allocation stability and Greek locale

locales/ar-SA/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = إغلاق

locales/bg-BG/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Затвори

locales/cs-CZ/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Zavřít

locales/de-DE/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Schließen

locales/en-US/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Close

locales/es-ES/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Cerrar

locales/fi-FI/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Sulje

locales/fr-FR/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Fermer

locales/hi-IN/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = बंद करें

locales/hu-HU/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Bezárás

locales/it-IT/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Chiudi

locales/ja-JP/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = 閉じる

locales/ko-KR/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = 닫기

locales/nl-NL/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Sluiten

locales/pl-PL/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Zamknij

locales/pt-BR/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Fechar

locales/ru-RU/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Замкнуть

locales/tr-TR/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = Kapat

locales/zh-CN/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = 关闭

locales/zh-TW/opencadstudio.ftl

@@ -20,6 +20,7 @@ language =
     .polish = Polski
     .chinese-traditional = 繁體中文
     .bulgarian = Български
+    .greek = Ελληνικά
 
 action =
     .close = 關閉

src/i18n.rs

@@ -65,10 +65,12 @@ pub enum Language {
     PlPl,
     #[serde(rename = "zh-TW")]
     ZhTw,
+    #[serde(rename = "el-GR")]
+    ElGr,
 }
 
 impl Language {
-    pub const ALL: [Language; 21] = [
+    pub const ALL: [Language; 22] = [
         Language::System,
         Language::EnUs,
         Language::BgBg,
@@ -78,6 +80,7 @@ impl Language {
         Language::FrFr,
         Language::FiFi,
         Language::DeDe,
+        Language::ElGr,
         Language::HuHu,
         Language::ItIt,
         Language::JaJp,
@@ -115,6 +118,7 @@ impl Language {
             Language::HuHu => vec!["hu-HU".parse().expect("valid locale")],
             Language::PlPl => vec!["pl-PL".parse().expect("valid locale")],
             Language::ZhTw => vec!["zh-TW".parse().expect("valid locale")],
+            Language::ElGr => vec!["el-GR".parse().expect("valid locale")],
         }
     }
 
@@ -141,6 +145,7 @@ impl Language {
             Language::HuHu => crate::tr!("language", "hungarian"),
             Language::PlPl => crate::tr!("language", "polish"),
             Language::ZhTw => crate::tr!("language", "chinese-traditional"),
+            Language::ElGr => crate::tr!("language", "greek"),
         }
     }
 }

src/scene/pipeline/face3d_gpu.rs

@@ -20,7 +20,6 @@
 
 use crate::scene::model::wire_model::WireModel;
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 pub(crate) fn planar_solid_faces_view(wire: &WireModel, view_dir: glam::Vec3) -> bool {
     if !wire.fill_is_2d_solid || wire.fill_is_3d {
@@ -187,6 +186,7 @@ impl Face3DGpu {
     ///   legacy planar SOLID interior while keeping other 2-D overlays.
     pub fn from_wires(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         face3d_wires: &[WireModel],
         all_wires: &[WireModel],
         keep_3d_mesh_fills: bool,
@@ -302,6 +302,7 @@ impl Face3DGpu {
 
         let (block_chunks_3d, block_chunks_2d) = upload_block_chunks(
             device,
+            queue,
             all_wires,
             keep_3d_mesh_fills,
             show_2d_solid_fills,
@@ -309,8 +310,8 @@ impl Face3DGpu {
             depth_map,
         );
         Self {
-            chunks_3d: upload_chunks(device, &verts_3d, "face3d.vbuf.3d"),
-            chunks_2d: upload_chunks(device, &verts_2d, "face3d.vbuf.2d"),
+            chunks_3d: upload_chunks(device, queue, &verts_3d, "face3d.vbuf.3d"),
+            chunks_2d: upload_chunks(device, queue, &verts_2d, "face3d.vbuf.2d"),
             block_chunks_3d,
             block_chunks_2d,
         }
@@ -319,6 +320,7 @@ impl Face3DGpu {
 
 fn upload_block_chunks(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     wires: &[WireModel],
     keep_3d_mesh_fills: bool,
     show_2d_solid_fills: bool,
@@ -376,11 +378,9 @@ fn upload_block_chunks(
         if vertices.is_empty() {
             continue;
         }
-        let budget = (device.limits().max_buffer_size as usize / 10) * 9;
-        let max_vertices = ((budget / std::mem::size_of::<Face3DVertex>()).max(3) / 3) * 3;
-        let max_instances = ((device.limits().max_buffer_size as usize / 10) * 9
-            / std::mem::size_of::<Face3DInstance>())
-            .max(1);
+        // Triangle lists: chunks must stay a multiple of 3.
+        let max_vertices = super::gpu_budget::max_elements_grouped::<Face3DVertex>(device, 3);
+        let max_instances = super::gpu_budget::max_elements::<Face3DInstance>(device);
         let instances: Vec<Face3DInstance> = group
             .iter()
             .filter_map(|wire| {
@@ -417,21 +417,25 @@ fn upload_block_chunks(
             .chunks(max_instances)
             .map(|chunk| {
                 (
-                    device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                        label: Some("face3d.block.instances"),
-                        contents: bytemuck::cast_slice(chunk),
-                        usage: wgpu::BufferUsages::VERTEX,
-                    }),
+                    super::gpu_upload::upload_buffer(
+                        device,
+                        queue,
+                        "face3d.block.instances",
+                        chunk,
+                        wgpu::BufferUsages::VERTEX,
+                    ),
                     chunk.len() as u32,
                 )
             })
             .collect();
         for vertex_chunk in vertices.chunks(max_vertices) {
-            let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("face3d.block.vertices"),
-                contents: bytemuck::cast_slice(vertex_chunk),
-                usage: wgpu::BufferUsages::VERTEX,
-            });
+            let vertex_buffer = super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "face3d.block.vertices",
+                vertex_chunk,
+                wgpu::BufferUsages::VERTEX,
+            );
             for (instance_buffer, instance_count) in &instance_buffers {
                 let chunk = BlockFace3DChunk {
                     vertex_buffer: vertex_buffer.clone(),
@@ -456,21 +460,22 @@ fn upload_block_chunks(
 /// range never truncates.
 fn upload_chunks(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     verts: &[Face3DVertex],
     label: &'static str,
 ) -> Vec<Face3DChunk> {
-    let budget = (device.limits().max_buffer_size as usize / 10) * 9; // 10% headroom
-    let vsize = std::mem::size_of::<Face3DVertex>();
     // Round down to a multiple of 3 so triangles never straddle chunks.
-    let max_verts = ((budget / vsize).max(3) / 3) * 3;
+    let max_verts = super::gpu_budget::max_elements_grouped::<Face3DVertex>(device, 3);
     verts
         .chunks(max_verts)
         .map(|c| Face3DChunk {
-            vertex_buffer: device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some(label),
-                contents: bytemuck::cast_slice(c),
-                usage: wgpu::BufferUsages::VERTEX,
-            }),
+            vertex_buffer: super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                label,
+                c,
+                wgpu::BufferUsages::VERTEX,
+            ),
             vertex_count: c.len() as u32,
         })
         .collect()

src/scene/pipeline/gpu_budget.rs

@@ -0,0 +1,62 @@
+//! One VRAM-aware budget for every GPU buffer this renderer allocates.
+//!
+//! `wgpu::Limits::max_buffer_size` is an **API validation limit**, not a
+//! statement about physical memory: NVIDIA reports several GB of it on a 2 GB
+//! card. Sizing a chunk from that limit therefore produces buffers larger than
+//! the whole device, the allocation fails, and the frame dies — issue #203
+//! (`mesh_gpu`), #358 (`face3d_gpu`) and the block-wire path were all the same
+//! bug rediscovered.
+//!
+//! `mesh_gpu.rs` was the only site that clamped the result. This module makes
+//! that clamp the single shared rule so a new upload path cannot reintroduce
+//! the mistake by copying one of the unclamped expressions.
+
+use iced::wgpu;
+
+/// Upper bound on any single buffer, regardless of what the device claims it
+/// would validate. Chosen to match the clamp `mesh_gpu` has used since #203:
+/// large enough that chunk overhead stays negligible, small enough that a
+/// failed allocation is recoverable on a low-VRAM card.
+const HARD_CAP_BYTES: usize = 32 * 1024 * 1024;
+
+/// 10% headroom below the device's validation limit, matching the `/ 10 * 9`
+/// the individual sites used before they shared this helper.
+fn device_ceiling(device: &wgpu::Device) -> usize {
+    (device.limits().max_buffer_size as usize / 10) * 9
+}
+
+/// Test/debug override, in MiB. Raising it above the real VRAM is the
+/// cheapest way to force an out-of-memory path and check that the renderer
+/// degrades the frame instead of aborting.
+fn cap_bytes() -> usize {
+    static CAP: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
+    *CAP.get_or_init(|| {
+        std::env::var("OCS_GPU_CHUNK_MIB")
+            .ok()
+            .and_then(|value| value.trim().parse::<usize>().ok())
+            .filter(|mib| *mib > 0)
+            .map(|mib| mib * 1024 * 1024)
+            .unwrap_or(HARD_CAP_BYTES)
+    })
+}
+
+/// Largest byte size a single buffer built by this renderer may take.
+pub fn buffer_budget(device: &wgpu::Device) -> usize {
+    device_ceiling(device).min(cap_bytes())
+}
+
+/// How many `T` fit in one buffer. Never returns 0, so callers can pass the
+/// result straight to `slice::chunks`, which panics on a zero chunk size.
+pub fn max_elements<T>(device: &wgpu::Device) -> usize {
+    (buffer_budget(device) / std::mem::size_of::<T>().max(1)).max(1)
+}
+
+/// `max_elements` rounded down to a whole number of `group` elements, for
+/// vertex data whose primitives must not straddle a chunk boundary: 3 for a
+/// triangle list, 6 for the quad expansion the wire shaders use.
+///
+/// Always returns at least one whole group.
+pub fn max_elements_grouped<T>(device: &wgpu::Device, group: usize) -> usize {
+    let group = group.max(1);
+    (max_elements::<T>(device) / group).max(1) * group
+}

src/scene/pipeline/gpu_upload.rs

@@ -0,0 +1,116 @@
+//! Buffer creation that survives an out-of-memory device.
+//!
+//! `wgpu::util::DeviceExt::create_buffer_init` — and any
+//! `mapped_at_creation: true` followed by `get_mapped_range_mut()` — maps the
+//! new buffer unconditionally. When the allocation fails wgpu hands back an
+//! *invalid* buffer, and mapping an invalid buffer panics:
+//!
+//! ```text
+//! thread 'main' panicked at wgpu-29.0.4/src/backend/wgpu_core.rs:2253:18:
+//!   Error in Buffer::get_mapped_range: Validation Error
+//!   Caused by: Buffer with 'block_wire.vertices' label is invalid
+//! ```
+//!
+//! That panic unwinds through iced's main-thread redraw, trips the `wgpu-hal`
+//! swapchain assertion in a destructor, and turns into a **non-unwinding
+//! abort** — the process dies with the user's unsaved drawing in it.
+//!
+//! Writing through the queue instead never maps. A failed allocation produces
+//! a validation error on the `write_buffer`, which reaches the handler
+//! installed by `install_gpu_error_handler` and degrades the frame, which is
+//! what that handler was written for: "A bad frame must degrade, not end the
+//! session."
+
+use iced::wgpu;
+
+/// Create a buffer and fill it without ever mapping it.
+///
+/// `COPY_DST` is added to `usage` because the contents arrive via
+/// `Queue::write_buffer`. Callers pass the same usage flags they would have
+/// given `create_buffer_init`.
+///
+/// Empty `data` still produces a one-element buffer: wgpu rejects zero-sized
+/// buffers, and every draw path here already guards its instance/vertex count
+/// before issuing a draw, so a placeholder allocation is harmless.
+pub fn upload_buffer<T: bytemuck::Pod>(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    label: &str,
+    data: &[T],
+    usage: wgpu::BufferUsages,
+) -> wgpu::Buffer {
+    let bytes: &[u8] = bytemuck::cast_slice(data);
+    upload_bytes(device, queue, label, bytes, std::mem::size_of::<T>(), usage)
+}
+
+/// `upload_buffer` for data already flattened to bytes. `min_size` is the
+/// placeholder size used when `bytes` is empty.
+pub fn upload_bytes(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    label: &str,
+    bytes: &[u8],
+    min_size: usize,
+    usage: wgpu::BufferUsages,
+) -> wgpu::Buffer {
+    const ALIGN: usize = wgpu::COPY_BUFFER_ALIGNMENT as usize;
+    let logical = bytes.len().max(min_size).max(ALIGN);
+    // `write_buffer` copies whole `COPY_BUFFER_ALIGNMENT` units, so both the
+    // allocation and the source slice are rounded up to it. Every vertex and
+    // instance type here is `#[repr(C)]` over 4-byte fields, so the padding
+    // branch below is a safety net rather than a routine cost.
+    let size = logical.div_ceil(ALIGN) * ALIGN;
+    let buffer = device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some(label),
+        size: size as u64,
+        usage: usage | wgpu::BufferUsages::COPY_DST,
+        mapped_at_creation: false,
+    });
+    if !bytes.is_empty() {
+        if bytes.len().is_multiple_of(ALIGN) {
+            queue.write_buffer(&buffer, 0, bytes);
+        } else {
+            let mut padded = bytes.to_vec();
+            padded.resize(bytes.len().div_ceil(ALIGN) * ALIGN, 0);
+            queue.write_buffer(&buffer, 0, &padded);
+        }
+    }
+    buffer
+}
+
+/// Allocate `size` bytes and write `data` into the front of it.
+///
+/// For buffers deliberately larger than their initial contents — the wire
+/// arena reserves headroom so later edits patch in place instead of
+/// reallocating. Same no-mapping guarantee as [`upload_buffer`]; `usage` must
+/// already contain `COPY_DST` for the later patches, and it is added here
+/// regardless.
+pub fn alloc_with_prefix<T: bytemuck::Pod>(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    label: &str,
+    size: u64,
+    data: &[T],
+    usage: wgpu::BufferUsages,
+) -> wgpu::Buffer {
+    const ALIGN: u64 = wgpu::COPY_BUFFER_ALIGNMENT;
+    let size = size.max(ALIGN).div_ceil(ALIGN) * ALIGN;
+    let buffer = device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some(label),
+        size,
+        usage: usage | wgpu::BufferUsages::COPY_DST,
+        mapped_at_creation: false,
+    });
+    let bytes: &[u8] = bytemuck::cast_slice(data);
+    if !bytes.is_empty() {
+        let unit = ALIGN as usize;
+        if bytes.len().is_multiple_of(unit) {
+            queue.write_buffer(&buffer, 0, bytes);
+        } else {
+            let mut padded = bytes.to_vec();
+            padded.resize(bytes.len().div_ceil(unit) * unit, 0);
+            queue.write_buffer(&buffer, 0, &padded);
+        }
+    }
+    buffer
+}

src/scene/pipeline/hatch_gpu/mod.rs

@@ -167,7 +167,7 @@ impl HatchGpu {
                     .filter(|hatch| hatch.boundary.len() >= 3)
                     .cloned()
                     .collect();
-                *resident = StorageHatchBatch::build(device, &self.bind_group_layout, &renderable);
+                *resident = StorageHatchBatch::build(device, queue, &self.bind_group_layout, &renderable);
             }
             HatchBackend::Texture { resident, .. } => {
                 let renderable: Vec<HatchModel> = hatches
@@ -198,7 +198,7 @@ impl HatchGpu {
                     .filter(|hatch| hatch.boundary.len() >= 3)
                     .cloned()
                     .collect();
-                *preview = StorageHatchBatch::build(device, &self.bind_group_layout, &renderable);
+                *preview = StorageHatchBatch::build(device, queue, &self.bind_group_layout, &renderable);
             }
             HatchBackend::Texture { preview, .. } => {
                 let renderable: Vec<HatchModel> = hatches

src/scene/pipeline/hatch_gpu/storage.rs

@@ -3,7 +3,6 @@
 
 use crate::scene::model::hatch_model::{HatchModel, HatchPattern, PatFamily};
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 // ── GPU structs ────────────────────────────────────────────────────────────
 //
@@ -188,6 +187,7 @@ impl StorageHatchBatch {
     /// Builds whole-hatch chunks within the device's buffer limits.
     pub(super) fn build(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         bgl: &wgpu::BindGroupLayout,
         hatches: &[HatchModel],
     ) -> Vec<Self> {
@@ -196,9 +196,13 @@ impl StorageHatchBatch {
         }
 
         let limits = device.limits();
+        // Both device figures are API validation limits, not physical VRAM, so
+        // the shared budget caps them (`storage.rs` chunks are allocations, not
+        // just size checks — the `*_fits` predicates below are the checks).
         let chunk_limit = limits
             .max_buffer_size
-            .min(limits.max_storage_buffer_binding_size as u64);
+            .min(limits.max_storage_buffer_binding_size as u64)
+            .min(crate::scene::pipeline::gpu_budget::buffer_budget(device) as u64);
         let mut ranges = Vec::new();
         let mut start = 0;
         let mut used = [0u64; 7];
@@ -220,12 +224,13 @@ impl StorageHatchBatch {
         ranges.push(start..hatches.len());
         ranges
             .into_iter()
-            .filter_map(|range| Self::build_chunk(device, bgl, &hatches[range]))
+            .filter_map(|range| Self::build_chunk(device, queue, bgl, &hatches[range]))
             .collect()
     }
 
     fn build_chunk(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         bgl: &wgpu::BindGroupLayout,
         hatches: &[HatchModel],
     ) -> Option<Self> {
@@ -496,42 +501,56 @@ impl StorageHatchBatch {
             return None;
         }
 
-        let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.vertex"),
-            contents: bytemuck::cast_slice(&verts),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
-        let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.index"),
-            contents: bytemuck::cast_slice(&indices),
-            usage: wgpu::BufferUsages::INDEX,
-        });
-        let placement_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.placements"),
-            contents: bytemuck::cast_slice(&placements),
-            usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
-        });
-        let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.instances"),
-            contents: bytemuck::cast_slice(&instances),
-            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
-        });
-        let family_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.families"),
-            contents: bytemuck::cast_slice(&families),
-            usage: wgpu::BufferUsages::STORAGE,
-        });
-        let dash_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.dashes"),
-            contents: bytemuck::cast_slice(&dashes),
-            usage: wgpu::BufferUsages::STORAGE,
-        });
+        let vertex_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.vertex",
+            &verts,
+            wgpu::BufferUsages::VERTEX,
+        );
+        let index_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.index",
+            &indices,
+            wgpu::BufferUsages::INDEX,
+        );
+        let placement_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.placements",
+            &placements,
+            wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+        );
+        let instance_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.instances",
+            &instances,
+            wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
+        );
+        let family_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.families",
+            &families,
+            wgpu::BufferUsages::STORAGE,
+        );
+        let dash_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.dashes",
+            &dashes,
+            wgpu::BufferUsages::STORAGE,
+        );
 
-        let visibility_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.visibility"),
-            contents: bytemuck::cast_slice(&visibility),
-            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
-        });
+        let visibility_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.visibility",
+            &visibility,
+            wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
+        );
 
         let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
             label: Some("hatch.bg"),

src/scene/pipeline/hatch_gpu/texture.rs

@@ -3,7 +3,6 @@
 
 use crate::scene::model::hatch_model::{HatchModel, HatchPattern};
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 /// Width (in texels) of the RGBA32F data texture. Height grows to fit; a hatch
 /// with N total texels uses ceil(N / WIDTH) rows.
@@ -199,16 +198,20 @@ impl TextureHatch {
             .into_iter()
             .map(|[x, y]| HatchVertex { pos: [x, y, 0.0], _pad: 0.0 })
             .collect();
-        let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.texture.vbuf"),
-            contents: bytemuck::cast_slice(&vertices),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
-        let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.texture.ibuf"),
-            contents: bytemuck::cast_slice(&indices),
-            usage: wgpu::BufferUsages::INDEX,
-        });
+        let vertex_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.texture.vbuf",
+            &vertices,
+            wgpu::BufferUsages::VERTEX,
+        );
+        let index_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.texture.ibuf",
+            &indices,
+            wgpu::BufferUsages::INDEX,
+        );
         let base = model
             .render_instance
             .map_or([0.0; 3], |instance| instance.translation);
@@ -231,11 +234,13 @@ impl TextureHatch {
                 }
             })
             .collect();
-        let placement_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.texture.placements"),
-            contents: bytemuck::cast_slice(&placements),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
+        let placement_buffer = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.texture.placements",
+            &placements,
+            wgpu::BufferUsages::VERTEX,
+        );
 
         // ── Gradient projection range (snapped-local space) ───────────────
         let base_mode = mode & 0xFF;
@@ -371,11 +376,13 @@ impl TextureHatch {
             dash_off,
             tex_width: width,
         };
-        let _uniform_buf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.texture.uniform"),
-            contents: bytemuck::bytes_of(&uniform_data),
-            usage: wgpu::BufferUsages::UNIFORM,
-        });
+        let _uniform_buf = crate::scene::pipeline::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.texture.uniform",
+            std::slice::from_ref(&uniform_data),
+            wgpu::BufferUsages::UNIFORM,
+        );
 
         let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
             label: Some("hatch.texture.bind_group1"),

src/scene/pipeline/image_gpu.rs

@@ -7,7 +7,6 @@
 
 use crate::scene::model::image_model::ImageModel;
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 use std::sync::Arc;
 
 // ── Vertex ────────────────────────────────────────────────────────────────
@@ -179,24 +178,24 @@ impl ImageGpu {
                 }
             })
             .collect();
-        let instance_buffer = Arc::new(device.create_buffer_init(
-            &wgpu::util::BufferInitDescriptor {
-                label: Some("image.instances"),
-                contents: bytemuck::cast_slice(&instances),
-                usage: wgpu::BufferUsages::VERTEX,
-            },
+        let instance_buffer = Arc::new(super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "image.instances",
+            &instances,
+            wgpu::BufferUsages::VERTEX,
         ));
         let params = ImageParams {
             opacity: model.opacity.clamp(0.0, 1.0),
             draw_depth: 0.0,
             _pad: [0.0; 2],
         };
-        let params_buf = Arc::new(device.create_buffer_init(
-            &wgpu::util::BufferInitDescriptor {
-                label: Some("image.params"),
-                contents: bytemuck::bytes_of(&params),
-                usage: wgpu::BufferUsages::UNIFORM,
-            },
+        let params_buf = Arc::new(super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "image.params",
+            std::slice::from_ref(&params),
+            wgpu::BufferUsages::UNIFORM,
         ));
 
         let mut output = Vec::with_capacity(x_tiles.len() * y_tiles.len());
@@ -264,12 +263,13 @@ impl ImageGpu {
                         },
                     ],
                 });
-                let vertex_buffer =
-                    device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                        label: Some("image.vbuf"),
-                        contents: bytemuck::cast_slice(&verts),
-                        usage: wgpu::BufferUsages::VERTEX,
-                    });
+                let vertex_buffer = super::gpu_upload::upload_buffer(
+                    device,
+                    queue,
+                    "image.vbuf",
+                    &verts,
+                    wgpu::BufferUsages::VERTEX,
+                );
                 output.push(Self {
                     vertex_buffer,
                     instance_buffer: Arc::clone(&instance_buffer),

src/scene/pipeline/mesh_gpu.rs

@@ -2,7 +2,6 @@
 
 use crate::scene::model::mesh_model::{MeshLodSet, MeshModel};
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 // ── Vertex layout ─────────────────────────────────────────────────────────
 
@@ -377,13 +376,15 @@ struct MeshBatchStubs {
 }
 
 impl MeshBatchStubs {
-    fn new(device: &wgpu::Device) -> Self {
+    fn new(device: &wgpu::Device, queue: &wgpu::Queue) -> Self {
         Self {
-            index: device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("mesh.batch.empty_ibuf"),
-                contents: bytemuck::cast_slice(&[0u32]),
-                usage: wgpu::BufferUsages::INDEX,
-            }),
+            index: super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "mesh.batch.empty_ibuf",
+                &[0u32],
+                wgpu::BufferUsages::INDEX,
+            ),
             vertex: device.create_buffer(&wgpu::BufferDescriptor {
                 label: Some("mesh.batch.empty_vbuf"),
                 size: std::mem::size_of::<MeshVertex>() as u64,
@@ -784,11 +785,13 @@ pub fn create_material_resources(
             [material.color_bleed_scale, 0.0, 0.0, 0.0]
         }),
     };
-    let params_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-        label: Some("mesh.material.params"),
-        contents: bytemuck::bytes_of(&params),
-        usage: wgpu::BufferUsages::UNIFORM,
-    });
+    let params_buffer = super::gpu_upload::upload_buffer(
+        device,
+        queue,
+        "mesh.material.params",
+        std::slice::from_ref(&params),
+        wgpu::BufferUsages::UNIFORM,
+    );
     MeshMaterialResources {
         diffuse_view,
         specular_view,
@@ -810,6 +813,7 @@ pub fn create_material_resources(
 
 pub fn create_material_bind_group_from_resources(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     layout: &wgpu::BindGroupLayout,
     resources: &MeshMaterialResources,
     material: Option<&crate::scene::model::material_model::MeshMaterial>,
@@ -825,11 +829,13 @@ pub fn create_material_bind_group_from_resources(
         advanced,
         flags,
     };
-    let surface_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-        label: Some("mesh.material.surface"),
-        contents: bytemuck::bytes_of(&surface),
-        usage: wgpu::BufferUsages::UNIFORM,
-    });
+    let surface_buffer = super::gpu_upload::upload_buffer(
+        device,
+        queue,
+        "mesh.material.surface",
+        std::slice::from_ref(&surface),
+        wgpu::BufferUsages::UNIFORM,
+    );
     device.create_bind_group(&wgpu::BindGroupDescriptor {
         label: Some("mesh.material.bind_group"),
         layout,
@@ -912,6 +918,7 @@ pub fn create_material_bind_group(
     let color = material.map_or([0.8, 0.8, 0.8, 1.0], |material| material.diffuse);
     create_material_bind_group_from_resources(
         device,
+        queue,
         layout,
         &resources,
         material,
@@ -936,6 +943,7 @@ pub fn upload_chunk_material_bind_groups(
                     let resources = create_material_resources(device, queue, Some(material));
                     chunk.material_bind_group = Some(create_material_bind_group_from_resources(
                         device,
+                        queue,
                         layout,
                         &resources,
                         chunk.material.as_ref(),
@@ -950,6 +958,7 @@ pub fn upload_chunk_material_bind_groups(
         };
         chunk.material_bind_group = Some(create_material_bind_group_from_resources(
             device,
+            queue,
             layout,
             resources,
             chunk.material.as_ref(),
@@ -1580,15 +1589,13 @@ pub fn build_mesh_batch_filtered(
     // Derive the caps from the real device limit and vertex size. The previous
     // fixed 6 M-vertex cap assumed 40 B/vertex, but `position_low` (RTE) grew
     // MeshVertex to 52 B, so 6 M × 52 B = 312 MB blew past the 256 MB cap.
-    let hard_budget = (device.limits().max_buffer_size as usize / 10) * 9;
-    let budget = hard_budget.min(32 * 1024 * 1024);
-    let vsize = std::mem::size_of::<MeshVertex>();
-    let max_verts = (budget / vsize).max(3);
+    let budget = super::gpu_budget::buffer_budget(device);
+    let max_verts = super::gpu_budget::max_elements::<MeshVertex>(device).max(3);
     // A triangle contributes three line segments (six standalone edge
     // vertices). Bound chunks by the largest buffer produced for wire meshes.
     let max_tris =
         (budget / (6 * std::mem::size_of::<MeshEdgeVertex>())).max(1);
-    let stubs = MeshBatchStubs::new(device);
+    let stubs = MeshBatchStubs::new(device, queue);
 
     let mut chunks = Vec::new();
     let mut verts: Vec<MeshVertex> = Vec::new();
@@ -1800,9 +1807,9 @@ pub fn build_mesh_batch_filtered(
                 .verts
                 .len()
                 .saturating_mul(std::mem::size_of::<MeshVertex>())
-                <= hard_budget
+                <= budget
             && part.indices.len().saturating_mul(std::mem::size_of::<u32>())
-                <= hard_budget
+                <= budget
             && part
                 .set
                 .instance_source
@@ -1812,7 +1819,7 @@ pub fn build_mesh_batch_filtered(
                         .edge_verts
                         .len()
                         .saturating_mul(std::mem::size_of::<MeshEdgeVertex>())
-                        <= hard_budget;
+                        <= budget;
                     let triangle_edges_fit = !part.include_edges
                         || !source.edge_verts.is_empty()
                         || part
@@ -1820,7 +1827,7 @@ pub fn build_mesh_batch_filtered(
                             .len()
                             .saturating_mul(2)
                             .saturating_mul(std::mem::size_of::<MeshEdgeVertex>())
-                            <= hard_budget;
+                            <= budget;
                     feature_edges_fit && triangle_edges_fit
                 })
             && part

src/scene/pipeline/mod.rs

@@ -1,5 +1,7 @@
 mod device_capabilities;
 pub mod face3d_gpu;
+pub mod gpu_budget;
+pub mod gpu_upload;
 pub mod hatch_gpu;
 pub mod wipeout_gpu;
 pub mod image_gpu;
@@ -14,7 +16,6 @@ pub mod wire_arena;
 pub mod wire_gpu;
 
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 use iced::{Rectangle, Size};
 
 pub use face3d_gpu::Face3DGpu;
@@ -2120,11 +2121,13 @@ impl Pipeline {
         // UV crop uniform: [uv_offset_x, uv_offset_y, uv_scale_x, uv_scale_y]
         // padded to 16 bytes (std140 vec2 alignment). Defaulted to the
         // identity crop (offset 0, scale 1) for the common on-canvas case.
-        let blit_uniform_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("blit.uniform_buffer"),
-            contents: bytemuck::cast_slice(&[0.0f32, 0.0, 1.0, 1.0]),
-            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
-        });
+        let blit_uniform_buffer = gpu_upload::upload_buffer(
+            device,
+            queue,
+            "blit.uniform_buffer",
+            &[0.0f32, 0.0, 1.0, 1.0],
+            wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+        );
 
         let blit_layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
             label: Some("blit.pipeline_layout"),
@@ -2340,6 +2343,7 @@ impl Pipeline {
     pub fn build_wire_buffers(
         &self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
     ) -> (
@@ -2382,6 +2386,7 @@ impl Pipeline {
             let refs: Vec<&WireModel> = wires[i..j].iter().collect();
             batches.extend(WireGpu::from_run_refs(
                 device,
+                queue,
                 &refs,
                 depth_map,
                 mesh_edge,
@@ -2391,6 +2396,7 @@ impl Pipeline {
         }
         let block_batches = BlockWireGpu::from_wires(
             device,
+            queue,
             &block_wires,
             depth_map,
             None,
@@ -2425,6 +2431,7 @@ impl Pipeline {
     pub fn upload_selected_wires(
         &mut self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         selected: &rustc_hash::FxHashSet<acadrust::Handle>,
         hovered: &rustc_hash::FxHashSet<acadrust::Handle>,
@@ -2494,6 +2501,7 @@ impl Pipeline {
         let mut gpu = if let Some(tint) = selected_tint {
             WireGpu::from_highlight_refs(
                 device,
+                queue,
                 &selected_regular,
                 tint,
                 depth_map,
@@ -2504,6 +2512,7 @@ impl Pipeline {
         };
         gpu.extend(WireGpu::from_highlight_refs(
             device,
+            queue,
             &hover_regular,
             WireModel::HOVER,
             depth_map,
@@ -2513,6 +2522,7 @@ impl Pipeline {
         let mut block_gpu = if let Some(tint) = selected_tint {
             BlockWireGpu::from_wires(
                 device,
+                queue,
                 &selected_blocks,
                 depth_map,
                 Some(tint),
@@ -2523,6 +2533,7 @@ impl Pipeline {
         };
         block_gpu.extend(BlockWireGpu::from_wires(
             device,
+            queue,
             &hover_blocks,
             depth_map,
             Some(WireModel::HOVER),
@@ -2553,6 +2564,7 @@ impl Pipeline {
     pub fn upload_text_highlight(
         &mut self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         selected: &rustc_hash::FxHashSet<acadrust::Handle>,
         hovered: &rustc_hash::FxHashSet<acadrust::Handle>,
@@ -2633,10 +2645,11 @@ impl Pipeline {
             }
         }
         self.text_highlight_vcount = out.len() as u32;
-        self.text_highlight_vbuf = text_gpu::upload_vertices(device, &out);
+        self.text_highlight_vbuf = text_gpu::upload_vertices(device, queue, &out);
         let mut block_gpu = if let Some(tint) = selected_tint {
             text_gpu::upload_block_vertex_refs(
                 device,
+                queue,
                 &selected_blocks,
                 depth_map,
                 Some(tint),
@@ -2646,6 +2659,7 @@ impl Pipeline {
         };
         block_gpu.extend(text_gpu::upload_block_vertex_refs(
             device,
+            queue,
             &hover_blocks,
             depth_map,
             Some(WireModel::HOVER),
@@ -2669,13 +2683,14 @@ impl Pipeline {
     pub fn upload_preview_wires(
         &mut self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
     ) {
         self.gpu_preview_wires = if wires.is_empty() {
             vec![]
         } else {
-            WireGpu::from_run(device, wires, depth_map, false, self.wire_const_bgl.as_ref())
+            WireGpu::from_run(device, queue, wires, depth_map, false, self.wire_const_bgl.as_ref())
         };
     }
 
@@ -2708,7 +2723,7 @@ impl Pipeline {
                 atlas.clear_dirty();
             }
         }
-        self.text_preview_vbuf = text_gpu::upload_vertices(device, verts);
+        self.text_preview_vbuf = text_gpu::upload_vertices(device, queue, verts);
         self.text_preview_vcount = verts.len() as u32;
     }
 
@@ -2719,6 +2734,7 @@ impl Pipeline {
     pub fn upload_silhouettes(
         &mut self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         sets: &[crate::scene::model::mesh_model::MeshLodSet],
         content_id: u64,
         view_dir: glam::Vec3,
@@ -2727,12 +2743,7 @@ impl Pipeline {
         let view = glam::DVec3::new(view_dir.x as f64, view_dir.y as f64, view_dir.z as f64)
             .normalize_or(glam::DVec3::NEG_Z);
         use crate::scene::pipeline::mesh_gpu::{SilhouetteInstance, SilhouetteVertex};
-        use wgpu::util::DeviceExt;
-        const SILHOUETTE_CHUNK_BYTES: u64 = 64 * 1024 * 1024;
-        let chunk_bytes = device
-            .limits()
-            .max_buffer_size
-            .min(SILHOUETTE_CHUNK_BYTES) as usize;
+        let chunk_bytes = gpu_budget::buffer_budget(device);
         let max_vertices = chunk_bytes / std::mem::size_of::<SilhouetteVertex>() / 2 * 2;
         if max_vertices < 2 {
             self.silhouette_chunks.clear();
@@ -2807,12 +2818,12 @@ impl Pipeline {
                         .chunks(max_instances.max(1))
                         .map(|instances| {
                             (
-                                device.create_buffer_init(
-                                    &wgpu::util::BufferInitDescriptor {
-                                        label: Some("mesh.silhouette.instances"),
-                                        contents: bytemuck::cast_slice(instances),
-                                        usage: wgpu::BufferUsages::VERTEX,
-                                    },
+                                gpu_upload::upload_buffer(
+                                    device,
+                                    queue,
+                                    "mesh.silhouette.instances",
+                                    instances,
+                                    wgpu::BufferUsages::VERTEX,
                                 ),
                                 instances.len() as u32,
                             )
@@ -2914,12 +2925,12 @@ impl Pipeline {
                 if vertex_count == 0 {
                     return;
                 }
-                let vertex_buffer = device.create_buffer_init(
-                    &wgpu::util::BufferInitDescriptor {
-                        label: Some("mesh.silhouette.vbuf"),
-                        contents: bytemuck::cast_slice(&verts[..vertex_count]),
-                        usage: wgpu::BufferUsages::VERTEX,
-                    },
+                let vertex_buffer = gpu_upload::upload_buffer(
+                    device,
+                    queue,
+                    "mesh.silhouette.vbuf",
+                    &verts[..vertex_count],
+                    wgpu::BufferUsages::VERTEX,
                 );
                 for (instance_buffer, instance_count) in &group.instance_buffers {
                     chunks.push(SilhouetteChunk {
@@ -2959,8 +2970,7 @@ impl Pipeline {
     /// so the mask pipeline stamps it straight into the stencil with `Invert`
     /// (even-odd fill → interior marked, any convexity). Empty input clears the
     /// boundary so the viewport renders unclipped (its render rectangle clips).
-    pub fn upload_clip_boundary(&mut self, device: &wgpu::Device, boundary_ndc: &[[f32; 2]]) {
-        use wgpu::util::DeviceExt;
+    pub fn upload_clip_boundary(&mut self, device: &wgpu::Device, queue: &wgpu::Queue, boundary_ndc: &[[f32; 2]]) {
         if boundary_ndc.len() < 3 {
             self.clip_boundary = None;
             return;
@@ -2976,11 +2986,13 @@ impl Pipeline {
             self.clip_boundary = None;
             return;
         }
-        let vbuf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("clip_boundary.vbuf"),
-            contents: bytemuck::cast_slice(&verts),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
+        let vbuf = gpu_upload::upload_buffer(
+            device,
+            queue,
+            "clip_boundary.vbuf",
+            &verts,
+            wgpu::BufferUsages::VERTEX,
+        );
         self.clip_boundary = Some((vbuf, (verts.len() / 2) as u32));
     }
 
@@ -3036,6 +3048,7 @@ impl Pipeline {
     pub fn upload_face3d(
         &mut self,
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         face3d_wires: &[WireModel],
         all_wires: &[WireModel],
         wireframe_only: bool,
@@ -3047,7 +3060,7 @@ impl Pipeline {
         // Edge buffer is always built from `face3d_wires`, so 3DFACE
         // outlines stay on the screen regardless of mode.
         self.gpu_face3d_edges =
-            WireGpu::from_batch(device, face3d_wires, depth_map, self.wire_const_bgl.as_ref());
+            WireGpu::from_batch(device, queue, face3d_wires, depth_map, self.wire_const_bgl.as_ref());
         // Fill buffer split: 3D quads + PolyfaceMesh / PolygonMesh face
         // tris go to `chunks_3d` (gated by `keep_3d_mesh_fills`);
         // 2D fills (text-LOD greek, MultiLeader background, dimension arrows) go to
@@ -3070,6 +3083,7 @@ impl Pipeline {
         } else {
             self.gpu_face3d_fill = Some(Face3DGpu::from_wires(
                 device,
+                queue,
                 face3d_wires,
                 all_wires,
                 keep_3d_mesh_fills,
@@ -3366,13 +3380,18 @@ impl Pipeline {
         self.hatch_gpu.upload_preview(device, queue, hatches);
     }
 
-    pub fn upload_wipeouts(&mut self, device: &wgpu::Device, wipeouts: &[HatchModel]) {
+    pub fn upload_wipeouts(
+        &mut self,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        wipeouts: &[HatchModel],
+    ) {
         let renderable: Vec<HatchModel> = wipeouts
             .iter()
             .filter(|h| h.boundary.len() >= 3)
             .cloned()
             .collect();
-        self.gpu_wipeouts = WipeoutGpu::from_models(device, &renderable, &self.wipeout_bgl1);
+        self.gpu_wipeouts = WipeoutGpu::from_models(device, queue, &renderable, &self.wipeout_bgl1);
     }
 
     pub fn upload_images(
@@ -3407,9 +3426,9 @@ impl Pipeline {
                 atlas.clear_dirty();
             }
         }
-        self.text_vbuf = text_gpu::upload_vertices(device, verts);
+        self.text_vbuf = text_gpu::upload_vertices(device, queue, verts);
         self.text_vcount = verts.len() as u32;
-        self.block_text_gpu = text_gpu::upload_block_vertices(device, wires, depth_map);
+        self.block_text_gpu = text_gpu::upload_block_vertices(device, queue, wires, depth_map);
         if let Some(started) = perf_started {
             let elapsed_ms = started.elapsed().as_secs_f64() * 1000.0;
             if elapsed_ms >= 1.0 {

src/scene/pipeline/text_gpu.rs

@@ -14,7 +14,6 @@
 #![allow(dead_code)]
 
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 use crate::scene::text::glyph_quads::GlyphQuad;
 use crate::scene::text::sdf_atlas::GlyphAtlas;
@@ -462,15 +461,17 @@ pub fn create_pipelines(
 
 pub fn upload_block_vertices(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     wires: &[crate::scene::model::wire_model::WireModel],
     depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
 ) -> Vec<BlockTextGpu> {
     let refs: Vec<&crate::scene::model::wire_model::WireModel> = wires.iter().collect();
-    upload_block_vertex_refs(device, &refs, depth_map, None)
+    upload_block_vertex_refs(device, queue, &refs, depth_map, None)
 }
 
 pub fn upload_block_vertex_refs(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     wires: &[&crate::scene::model::wire_model::WireModel],
     depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
     tint: Option<[f32; 4]>,
@@ -516,7 +517,7 @@ pub fn upload_block_vertex_refs(
         } else {
             source.text_verts.as_slice()
         };
-        let Some(vertex_buffer) = upload_vertices(device, vertices) else {
+        let Some(vertex_buffer) = upload_vertices(device, queue, vertices) else {
             continue;
         };
         let instances: Vec<BlockTextInstance> = group
@@ -541,15 +542,16 @@ pub fn upload_block_vertex_refs(
                 })
             })
             .collect();
-        let max_instances = ((device.limits().max_buffer_size as usize / 10) * 9
-            / std::mem::size_of::<BlockTextInstance>())
-            .max(1);
+        let max_instances =
+            super::gpu_budget::max_elements::<BlockTextInstance>(device);
         for chunk in instances.chunks(max_instances) {
-            let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("block_text.instances"),
-                contents: bytemuck::cast_slice(chunk),
-                usage: wgpu::BufferUsages::VERTEX,
-            });
+            let instance_buffer = super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "block_text.instances",
+                chunk,
+                wgpu::BufferUsages::VERTEX,
+            );
             out.push(BlockTextGpu {
                 vertex_buffer: vertex_buffer.clone(),
                 instance_buffer,
@@ -562,15 +564,21 @@ pub fn upload_block_vertex_refs(
 }
 
 /// Upload a finished vertex list; `None` if empty.
-pub fn upload_vertices(device: &wgpu::Device, verts: &[TextVertex]) -> Option<wgpu::Buffer> {
+pub fn upload_vertices(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    verts: &[TextVertex],
+) -> Option<wgpu::Buffer> {
     if verts.is_empty() {
         return None;
     }
-    Some(device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-        label: Some("text.vbuf"),
-        contents: bytemuck::cast_slice(verts),
-        usage: wgpu::BufferUsages::VERTEX,
-    }))
+    Some(super::gpu_upload::upload_buffer(
+        device,
+        queue,
+        "text.vbuf",
+        verts,
+        wgpu::BufferUsages::VERTEX,
+    ))
 }
 
 // ── Tests ─────────────────────────────────────────────────────────────────

src/scene/pipeline/viewcube.rs

@@ -1078,43 +1078,54 @@ pub struct ViewCubePipeline {
 
 impl ViewCubePipeline {
     pub fn new(device: &wgpu::Device, queue: &wgpu::Queue, format: wgpu::TextureFormat) -> Self {
-        use wgpu::util::DeviceExt;
         let (verts, idxs) = build_geometry();
         let line_verts = surface_edge_lines(&verts, &idxs);
         let (ring_verts, ring_idxs) = build_ring_geometry();
         let ring_line_verts = surface_edge_lines(&ring_verts, &ring_idxs);
         let cube_px = VIEWCUBE_PX;
-        let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("vc.vb"),
-            contents: bytemuck::cast_slice(&verts),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
-        let index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("vc.ib"),
-            contents: bytemuck::cast_slice(&idxs),
-            usage: wgpu::BufferUsages::INDEX,
-        });
-        let line_vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("vc.line_vb"),
-            contents: bytemuck::cast_slice(&line_verts),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
-        let ring_vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("vc.ring_vb"),
-            contents: bytemuck::cast_slice(&ring_verts),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
-        let ring_index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("vc.ring_ib"),
-            contents: bytemuck::cast_slice(&ring_idxs),
-            usage: wgpu::BufferUsages::INDEX,
-        });
+        let vertex_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "vc.vb",
+            &verts,
+            wgpu::BufferUsages::VERTEX,
+        );
+        let index_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "vc.ib",
+            &idxs,
+            wgpu::BufferUsages::INDEX,
+        );
+        let line_vertex_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "vc.line_vb",
+            &line_verts,
+            wgpu::BufferUsages::VERTEX,
+        );
+        let ring_vertex_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "vc.ring_vb",
+            &ring_verts,
+            wgpu::BufferUsages::VERTEX,
+        );
+        let ring_index_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "vc.ring_ib",
+            &ring_idxs,
+            wgpu::BufferUsages::INDEX,
+        );
         let ring_line_vertex_buffer =
-            device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("vc.ring_line_vb"),
-                contents: bytemuck::cast_slice(&ring_line_verts),
-                usage: wgpu::BufferUsages::VERTEX,
-            });
+            super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "vc.ring_line_vb",
+                &ring_line_verts,
+                wgpu::BufferUsages::VERTEX,
+            );
         let uniform_buffer = device.create_buffer(&wgpu::BufferDescriptor {
             label: Some("vc.ub"),
             size: std::mem::size_of::<CubeUniforms>() as u64,
@@ -1291,11 +1302,13 @@ impl ViewCubePipeline {
                 ],
             });
         let composite_uniform_buffer =
-            device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("vc.composite_uniform"),
-                contents: bytemuck::cast_slice(&[1.0f32, 1.0, 0.0, 0.0]),
-                usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
-            });
+            super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "vc.composite_uniform",
+                &[1.0f32, 1.0, 0.0, 0.0],
+                wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            );
         let composite_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
             label: Some("vc.composite_sampler"),
             mag_filter: wgpu::FilterMode::Nearest,

src/scene/pipeline/wipeout_gpu.rs

@@ -17,7 +17,6 @@ use crate::scene::model::hatch_model::{
     FillPlane, HatchModel, HatchPattern, PatFamily, MAX_HATCH_BOUNDARY_VERTS,
 };
 use iced::wgpu;
-use iced::wgpu::util::DeviceExt;
 
 // ── Constants ──────────────────────────────────────────────────────────────
 
@@ -157,6 +156,7 @@ pub struct WipeoutGpu {
 impl WipeoutGpu {
     pub fn from_models(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         models: &[HatchModel],
         bgl1: &wgpu::BindGroupLayout,
     ) -> Vec<Self> {
@@ -176,11 +176,16 @@ impl WipeoutGpu {
         }
         groups
             .into_iter()
-            .map(|group| Self::new(device, &group, bgl1))
+            .map(|group| Self::new(device, queue, &group, bgl1))
             .collect()
     }
 
-    fn new(device: &wgpu::Device, models: &[&HatchModel], bgl1: &wgpu::BindGroupLayout) -> Self {
+    fn new(
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        models: &[&HatchModel],
+        bgl1: &wgpu::BindGroupLayout,
+    ) -> Self {
         let model = models[0];
         // ── Decode pattern mode ──────────────────────────────────────────
         let (mode, color2, grad_cos, grad_sin) = match &model.pattern {
@@ -264,11 +269,13 @@ impl WipeoutGpu {
                 _pad: 0.0,
             },
         ];
-        let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.vbuf"),
-            contents: bytemuck::cast_slice(&quad),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
+        let vertex_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.vbuf",
+            &quad,
+            wgpu::BufferUsages::VERTEX,
+        );
 
         // ── Gradient: projection range (in snapped-local space) ──────────
         let (grad_min, grad_range) = if mode == 2 {
@@ -354,21 +361,27 @@ impl WipeoutGpu {
         let family_batch = build_family_batch(&model.pattern);
 
         // ── GPU buffers ───────────────────────────────────────────────────
-        let _uniform_buf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.uniforms"),
-            contents: bytemuck::bytes_of(&uniform_data),
-            usage: wgpu::BufferUsages::UNIFORM,
-        });
-        let _boundary_buf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.boundary"),
-            contents: bytemuck::bytes_of(&boundary_data),
-            usage: wgpu::BufferUsages::UNIFORM,
-        });
-        let _family_buf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("hatch.families"),
-            contents: bytemuck::bytes_of(&family_batch),
-            usage: wgpu::BufferUsages::UNIFORM,
-        });
+        let _uniform_buf = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.uniforms",
+            std::slice::from_ref(&uniform_data),
+            wgpu::BufferUsages::UNIFORM,
+        );
+        let _boundary_buf = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.boundary",
+            std::slice::from_ref(&boundary_data),
+            wgpu::BufferUsages::UNIFORM,
+        );
+        let _family_buf = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "hatch.families",
+            std::slice::from_ref(&family_batch),
+            wgpu::BufferUsages::UNIFORM,
+        );
 
         let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
             label: Some("hatch.bind_group1"),
@@ -419,11 +432,13 @@ impl WipeoutGpu {
                 }
             })
             .collect();
-        let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-            label: Some("wipeout.instances"),
-            contents: bytemuck::cast_slice(&placements),
-            usage: wgpu::BufferUsages::VERTEX,
-        });
+        let instance_buffer = super::gpu_upload::upload_buffer(
+            device,
+            queue,
+            "wipeout.instances",
+            &placements,
+            wgpu::BufferUsages::VERTEX,
+        );
         let mut world_bounds = [
             f32::INFINITY,
             f32::INFINITY,

src/scene/pipeline/wire_arena.rs

@@ -255,48 +255,34 @@ fn make_const_bg(
 
 fn alloc_inst_initialized(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     cap: u64,
     data: &[WireInstance],
 ) -> wgpu::Buffer {
-    let buffer = device.create_buffer(&wgpu::BufferDescriptor {
-        label: Some("wire_arena.ibuf"),
-        size: cap * std::mem::size_of::<WireInstance>() as u64,
-        usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
-        mapped_at_creation: true,
-    });
-    if !data.is_empty() {
-        let bytes = bytemuck::cast_slice(data);
-        let mut mapped = buffer
-            .slice(..bytes.len() as u64)
-            .get_mapped_range_mut();
-        mapped.copy_from_slice(bytes);
-        drop(mapped);
-    }
-    buffer.unmap();
-    buffer
+    super::gpu_upload::alloc_with_prefix(
+        device,
+        queue,
+        "wire_arena.ibuf",
+        cap * std::mem::size_of::<WireInstance>() as u64,
+        data,
+        wgpu::BufferUsages::VERTEX,
+    )
 }
 
 fn alloc_const_initialized(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     cap: u64,
     data: &[WireConst],
 ) -> wgpu::Buffer {
-    let buffer = device.create_buffer(&wgpu::BufferDescriptor {
-        label: Some("wire_arena.cbuf"),
-        size: cap * std::mem::size_of::<WireConst>() as u64,
-        usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
-        mapped_at_creation: true,
-    });
-    if !data.is_empty() {
-        let bytes = bytemuck::cast_slice(data);
-        let mut mapped = buffer
-            .slice(..bytes.len() as u64)
-            .get_mapped_range_mut();
-        mapped.copy_from_slice(bytes);
-        drop(mapped);
-    }
-    buffer.unmap();
-    buffer
+    super::gpu_upload::alloc_with_prefix(
+        device,
+        queue,
+        "wire_arena.cbuf",
+        cap * std::mem::size_of::<WireConst>() as u64,
+        data,
+        wgpu::BufferUsages::STORAGE,
+    )
 }
 
 fn blank_const() -> WireConst {
@@ -328,7 +314,7 @@ impl WireArena {
     /// path).
     pub fn build(
         device: &wgpu::Device,
-        _queue: &wgpu::Queue,
+        queue: &wgpu::Queue,
         wires: &[&WireModel],
         depth_map: &FxHashMap<u64, [f32; 2]>,
         const_bgl: &wgpu::BindGroupLayout,
@@ -468,8 +454,8 @@ impl WireArena {
             .max(MIN_CONST_CAP)
             .min(MAX_CONSTS)) as u32;
         let upload_started = iced::time::Instant::now();
-        let inst_buf = alloc_inst_initialized(device, inst_cap as u64, &instances);
-        let const_buf = alloc_const_initialized(device, const_cap as u64, &consts_cpu);
+        let inst_buf = alloc_inst_initialized(device, queue, inst_cap as u64, &instances);
+        let const_buf = alloc_const_initialized(device, queue, const_cap as u64, &consts_cpu);
         let upload_ms = upload_started.elapsed().as_secs_f64() * 1000.0;
         let const_bind_group = make_const_bg(device, const_bgl, &const_buf);
         if perf {
@@ -934,25 +920,18 @@ fn can_resize_packed_terminal_slab(
 
 fn alloc_packed_initialized(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     cap: u64,
     data: &[PackedWireInstance],
 ) -> wgpu::Buffer {
-    let buffer = device.create_buffer(&wgpu::BufferDescriptor {
-        label: Some("wire_arena.packed.ibuf"),
-        size: cap * std::mem::size_of::<PackedWireInstance>() as u64,
-        usage: wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
-        mapped_at_creation: true,
-    });
-    if !data.is_empty() {
-        let bytes = bytemuck::cast_slice(data);
-        let mut mapped = buffer
-            .slice(..bytes.len() as u64)
-            .get_mapped_range_mut();
-        mapped.copy_from_slice(bytes);
-        drop(mapped);
-    }
-    buffer.unmap();
-    buffer
+    super::gpu_upload::alloc_with_prefix(
+        device,
+        queue,
+        "wire_arena.packed.ibuf",
+        cap * std::mem::size_of::<PackedWireInstance>() as u64,
+        data,
+        wgpu::BufferUsages::VERTEX,
+    )
 }
 
 fn visible_ranges(
@@ -998,6 +977,7 @@ fn visible_ranges(
 impl PackedWireArena {
     fn build(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[&WireModel],
         depth_map: &FxHashMap<u64, [f32; 2]>,
         mesh_edge: bool,
@@ -1078,7 +1058,7 @@ impl PackedWireArena {
         let inst_cap = ((inst_tail as u64 * HEADROOM_NUM / HEADROOM_DEN)
             .max(MIN_INST_CAP)
             .min(MAX_PACKED_INSTANCES)) as u32;
-        let inst_buf = alloc_packed_initialized(device, inst_cap as u64, &instances);
+        let inst_buf = alloc_packed_initialized(device, queue, inst_cap as u64, &instances);
 
         Some(Self {
             inst_buf,
@@ -1351,7 +1331,7 @@ impl PersistentWireArena {
             )?)
         } else {
             PersistentWireArenaKind::Packed(PackedWireArena::build(
-                device, wires, depth_map, mesh_edge,
+                device, queue, wires, depth_map, mesh_edge,
             )?)
         };
         Some(Self { inner })

src/scene/pipeline/wire_gpu.rs

@@ -28,34 +28,20 @@
 use crate::scene::model::wire_model::WireModel;
 use iced::wgpu;
 
-/// Allocate a VERTEX buffer with `mapped_at_creation` and write `data` directly
-/// into the mapped slice. Skips the intermediate staging copy that
-/// `create_buffer_init` performs and avoids holding a second `Vec` worth of
-/// memory during upload — meaningful on cold open where wire buffers can run
-/// into the hundreds of MB.
-fn instance_buffer_mapped<T: bytemuck::Pod>(
+/// Allocate a VERTEX buffer and fill it through the queue.
+///
+/// This used to map at creation to skip the staging copy. Mapping panics on
+/// the invalid buffer an out-of-memory device returns, and that panic aborts
+/// the process from inside iced's redraw, so the staging copy is the price of
+/// a session that survives a failed allocation. Chunking keeps the transient
+/// staging cost bounded. See `super::gpu_upload`.
+fn instance_buffer<T: bytemuck::Pod>(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     label: &str,
     data: &[T],
 ) -> wgpu::Buffer {
-    let bytes: &[u8] = bytemuck::cast_slice(data);
-    // wgpu rejects size-0 buffers; the renderer already guards `instance_count`
-    // before issuing a draw, so a placeholder allocation is fine here.
-    let size = bytes.len().max(std::mem::size_of::<T>()) as u64;
-    let buf = device.create_buffer(&wgpu::BufferDescriptor {
-        label: Some(label),
-        size,
-        usage: wgpu::BufferUsages::VERTEX,
-        mapped_at_creation: true,
-    });
-    {
-        let mut view = buf.slice(..).get_mapped_range_mut();
-        if !bytes.is_empty() {
-            view.slice(..bytes.len()).copy_from_slice(bytes);
-        }
-    }
-    buf.unmap();
-    buf
+    super::gpu_upload::upload_buffer(device, queue, label, data, wgpu::BufferUsages::VERTEX)
 }
 
 // ── Instance layout ───────────────────────────────────────────────────────
@@ -684,12 +670,12 @@ pub(crate) fn wire_draw_depth(
 impl BlockWireGpu {
     pub fn from_wires(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[&WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
         color_override: Option<[f32; 4]>,
         const_bgl: &wgpu::BindGroupLayout,
     ) -> Vec<Self> {
-        use wgpu::util::DeviceExt;
 
         let mut slots: rustc_hash::FxHashMap<(u64, bool), usize> =
             rustc_hash::FxHashMap::default();
@@ -740,16 +726,42 @@ impl BlockWireGpu {
                 };
                 vertices.extend_from_slice(&[vertex; 6]);
             }
-            let vertex_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("block_wire.vertices"),
-                contents: bytemuck::cast_slice(&vertices),
-                usage: wgpu::BufferUsages::VERTEX,
-            });
-            let const_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-                label: Some("block_wire.const"),
-                contents: bytemuck::bytes_of(&constant),
-                usage: wgpu::BufferUsages::UNIFORM,
-            });
+            // Split vertex data into buffers small enough to allocate on
+            // low-VRAM devices. Same scheme as the solid-mesh chunking in
+            // `mesh_gpu.rs` (#203) and the fill chunking in `face3d_gpu.rs`
+            // (#358); the block-wire path was never converted, so a
+            // block-heavy drawing produced one multi-hundred-MB buffer.
+            // Round down to a multiple of 6: each segment expands to six
+            // vertices and must not straddle two chunks.
+            let max_verts =
+                super::gpu_budget::max_elements_grouped::<BlockWireVertex>(device, 6);
+            let vertex_chunks: Vec<(wgpu::Buffer, u32)> = vertices
+                .chunks(max_verts)
+                .map(|chunk| {
+                    (
+                        super::gpu_upload::upload_buffer(
+                            device,
+                            queue,
+                            "block_wire.vertices",
+                            chunk,
+                            wgpu::BufferUsages::VERTEX,
+                        ),
+                        chunk.len() as u32,
+                    )
+                })
+                .collect();
+            // Tiny, but size is irrelevant here: once the device is out of
+            // memory every allocation fails, including a one-struct uniform,
+            // and create_buffer_init would map the invalid buffer and panic.
+            // This is the allocation that killed the process after the vertex
+            // buffers had already degraded gracefully.
+            let const_buffer = super::gpu_upload::upload_buffer(
+                device,
+                queue,
+                "block_wire.const",
+                std::slice::from_ref(&constant),
+                wgpu::BufferUsages::UNIFORM,
+            );
             let const_bind_group = std::sync::Arc::new(device.create_bind_group(
                 &wgpu::BindGroupDescriptor {
                     label: Some("block_wire.const.bg"),
@@ -788,22 +800,29 @@ impl BlockWireGpu {
                     })
                 })
                 .collect();
-            let max_instances = ((device.limits().max_buffer_size as usize / 10) * 9
-                / std::mem::size_of::<BlockWireInstance>())
-                .max(1);
-            for chunk in instances.chunks(max_instances) {
-                out.push(Self {
-                    vertex_buffer: vertex_buffer.clone(),
-                    instance_buffer: instance_buffer_mapped(
-                        device,
-                        "block_wire.instances",
-                        chunk,
-                    ),
-                    vertex_count: vertices.len() as u32,
-                    instance_count: chunk.len() as u32,
-                    is_3d_mesh_edge: mesh_edge,
-                    const_bind_group: const_bind_group.clone(),
-                });
+            let max_instances = super::gpu_budget::max_elements::<BlockWireInstance>(device);
+            // Build each instance buffer once and reuse it across vertex
+            // chunks, rather than re-uploading it per chunk.
+            let instance_chunks: Vec<(wgpu::Buffer, u32)> = instances
+                .chunks(max_instances)
+                .map(|chunk| {
+                    (
+                        instance_buffer(device, queue, "block_wire.instances", chunk),
+                        chunk.len() as u32,
+                    )
+                })
+                .collect();
+            for (vertex_buffer, vertex_count) in &vertex_chunks {
+                for (instance_buffer, instance_count) in &instance_chunks {
+                    out.push(Self {
+                        vertex_buffer: vertex_buffer.clone(),
+                        instance_buffer: instance_buffer.clone(),
+                        vertex_count: *vertex_count,
+                        instance_count: *instance_count,
+                        is_3d_mesh_edge: mesh_edge,
+                        const_bind_group: const_bind_group.clone(),
+                    });
+                }
             }
         }
         out
@@ -816,18 +835,20 @@ impl BlockWireGpu {
 /// each chunk.
 fn build_const_bind_group(
     device: &wgpu::Device,
+    queue: &wgpu::Queue,
     bgl: &wgpu::BindGroupLayout,
     consts: &[WireConst],
 ) -> std::sync::Arc<wgpu::BindGroup> {
-    use wgpu::util::DeviceExt;
     // wgpu rejects zero-sized buffers; pad with one zeroed record when empty.
     let one = [<WireConst as bytemuck::Zeroable>::zeroed()];
     let data: &[WireConst] = if consts.is_empty() { &one } else { consts };
-    let buf = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-        label: Some("wire_const.buf"),
-        contents: bytemuck::cast_slice(data),
-        usage: wgpu::BufferUsages::STORAGE,
-    });
+    let buf = super::gpu_upload::upload_buffer(
+        device,
+        queue,
+        "wire_const.buf",
+        data,
+        wgpu::BufferUsages::STORAGE,
+    );
     let bg = device.create_bind_group(&wgpu::BindGroupDescriptor {
         label: Some("wire_const.bg"),
         layout: bgl,
@@ -845,6 +866,7 @@ impl WireGpu {
     /// of a large selected polyline or block before packing the overlay.
     pub fn from_highlight_refs(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[&WireModel],
         color: [f32; 4],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
@@ -871,12 +893,13 @@ impl WireGpu {
             if instances.is_empty() {
                 return Vec::new();
             }
-            let bind_group = build_const_bind_group(device, const_bgl, &consts);
+            let bind_group = build_const_bind_group(device, queue, const_bgl, &consts);
             return instances
                 .chunks(MAX_INSTANCES)
                 .map(|chunk| Self {
-                    instance_buffer: instance_buffer_mapped(
+                    instance_buffer: instance_buffer(
                         device,
+                        queue,
                         "wire.highlight.ibuf",
                         chunk,
                     ),
@@ -904,8 +927,9 @@ impl WireGpu {
         instances
             .chunks(MAX_PACKED_INSTANCES)
             .map(|chunk| Self {
-                instance_buffer: instance_buffer_mapped(
+                instance_buffer: instance_buffer(
                     device,
+                    queue,
                     "wire.highlight.compat.ibuf",
                     chunk,
                 ),
@@ -923,6 +947,7 @@ impl WireGpu {
     /// uses chunked resident buffers.
     pub fn from_run_refs(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[&WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
         mesh_edge: bool,
@@ -951,8 +976,9 @@ impl WireGpu {
             return instances
                 .chunks(MAX_INSTANCES)
                 .map(|chunk| Self {
-                    instance_buffer: instance_buffer_mapped(
+                    instance_buffer: instance_buffer(
                         device,
+                        queue,
                         "wire.run.hybrid.compat.ibuf",
                         chunk,
                     ),
@@ -988,11 +1014,11 @@ impl WireGpu {
         if instances.is_empty() {
             return Vec::new();
         }
-        let bind_group = build_const_bind_group(device, const_bgl, &consts);
+        let bind_group = build_const_bind_group(device, queue, const_bgl, &consts);
         instances
             .chunks(MAX_INSTANCES)
             .map(|chunk| Self {
-                instance_buffer: instance_buffer_mapped(device, "wire.run.hybrid.ibuf", chunk),
+                instance_buffer: instance_buffer(device, queue, "wire.run.hybrid.ibuf", chunk),
                 first_instance: 0,
                 instance_count: chunk.len() as u32,
                 is_3d_mesh_edge: mesh_edge,
@@ -1013,6 +1039,7 @@ impl WireGpu {
     /// image for transparent / coincident wires.
     pub fn from_run(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
         mesh_edge: bool,
@@ -1047,11 +1074,11 @@ impl WireGpu {
             if instances.is_empty() {
                 return vec![];
             }
-            let bg = build_const_bind_group(device, const_bgl, &consts);
+            let bg = build_const_bind_group(device, queue, const_bgl, &consts);
             return instances
                 .chunks(MAX_INSTANCES)
                 .map(|chunk| {
-                    let buf = instance_buffer_mapped(device, "wire.run.ibuf", chunk);
+                    let buf = instance_buffer(device, queue, "wire.run.ibuf", chunk);
                     Self {
                         instance_buffer: buf,
                         first_instance: 0,
@@ -1084,7 +1111,7 @@ impl WireGpu {
         instances
             .chunks(MAX_PACKED_INSTANCES)
             .map(|chunk| {
-                let buf = instance_buffer_mapped(device, "wire.run.compat.ibuf", chunk);
+                let buf = instance_buffer(device, queue, "wire.run.compat.ibuf", chunk);
                 Self {
                     instance_buffer: buf,
                     first_instance: 0,
@@ -1101,6 +1128,7 @@ impl WireGpu {
     /// per-instance.
     pub fn from_batch(
         device: &wgpu::Device,
+        queue: &wgpu::Queue,
         wires: &[WireModel],
         depth_map: &rustc_hash::FxHashMap<u64, [f32; 2]>,
         const_bgl: Option<&wgpu::BindGroupLayout>,
@@ -1134,13 +1162,13 @@ impl WireGpu {
             if instances.is_empty() {
                 return vec![];
             }
-            let bg = build_const_bind_group(device, const_bgl, &consts);
+            let bg = build_const_bind_group(device, queue, const_bgl, &consts);
             return instances
                 .chunks(MAX_INSTANCES)
                 .enumerate()
                 .map(|(i, chunk)| {
                     let label = format!("wire.batch.ibuf.{i}");
-                    let instance_buffer = instance_buffer_mapped(device, &label, chunk);
+                    let instance_buffer = instance_buffer(device, queue, &label, chunk);
                     Self {
                         instance_buffer,
                         first_instance: 0,
@@ -1167,7 +1195,7 @@ impl WireGpu {
             .enumerate()
             .map(|(i, chunk)| {
                 let label = format!("wire.batch.compat.ibuf.{i}");
-                let instance_buffer = instance_buffer_mapped(device, &label, chunk);
+                let instance_buffer = instance_buffer(device, queue, &label, chunk);
                 Self {
                     instance_buffer,
                     first_instance: 0,

src/scene/view/render.rs

@@ -546,6 +546,7 @@ impl shader::Primitive for Primitive {
             if wipeout_changed || fill_changed {
                 inner.upload_wipeouts(
                     device,
+                    queue,
                     if fill_mode {
                         &vp.wipeout_hatches[..]
                     } else {
@@ -608,6 +609,7 @@ impl shader::Primitive for Primitive {
             {
                 inner.upload_face3d(
                     device,
+                    queue,
                     &vp.face3d_wires[..],
                     &vp_wires[..],
                     !face3d_fill_active,
@@ -785,6 +787,7 @@ impl shader::Primitive for Primitive {
                                 inner.wire_arena_fallback = std::sync::Arc::new(
                                     crate::scene::pipeline::WireGpu::from_run_refs(
                                         device,
+                                        queue,
                                         &regular,
                                         &draw_depths,
                                         false,
@@ -821,6 +824,7 @@ impl shader::Primitive for Primitive {
                                 inner.wire_arena_fallback = std::sync::Arc::new(
                                     crate::scene::pipeline::WireGpu::from_run_refs(
                                         device,
+                                        queue,
                                         &mesh,
                                         &draw_depths,
                                         true,
@@ -917,22 +921,40 @@ impl shader::Primitive for Primitive {
                     let built = match cached {
                         Some(entry) => entry,
                         None => {
+                            // Free the superseded content BEFORE allocating its
+                            // replacement. Every geometry edit mints a new content
+                            // id, so a miss here means this slot's current buffers
+                            // are already dead weight — and on a large drawing one
+                            // set is most of the card (measured: 1.3 GB of a 2 GB
+                            // GT 1030). Holding the old set while the new one is
+                            // allocated is what exhausts VRAM after the first edit.
+                            //
+                            // Dropping this slot's handles first is what makes the
+                            // retain below able to see the old entry as garbage; a
+                            // pane still drawing that content keeps its own clone,
+                            // so shared geometry is never pulled out from under it.
+                            inner.gpu_wires = std::sync::Arc::new(Vec::new());
+                            inner.gpu_block_wires = std::sync::Arc::new(Vec::new());
+                            // Entries no slot still holds: only the cache references
+                            // them. An entry drawn by any pane keeps a strong count
+                            // ≥ 2, so this never drops live geometry.
+                            let held_before = pipeline.wire_buffer_cache.len();
+                            pipeline.wire_buffer_cache.retain(|_, (w, b, _)| {
+                                std::sync::Arc::strong_count(w) > 1
+                                    || std::sync::Arc::strong_count(b) > 1
+                            });
+                            if _perf {
+                                crate::perf_record!(
+                                    "[perf] wire-cache evicted={} held={}",
+                                    held_before - pipeline.wire_buffer_cache.len(),
+                                    pipeline.wire_buffer_cache.len(),
+                                );
+                            }
                             let entry =
-                                inner.build_wire_buffers(device, &vp_wires[..], &draw_depths);
+                                inner.build_wire_buffers(device, queue, &vp_wires[..], &draw_depths);
                             pipeline
                                 .wire_buffer_cache
                                 .insert(vp.wire_content_id, entry.clone());
-                            // Evict entries no slot still holds (only the cache
-                            // references them). An entry drawn by any pane keeps a
-                            // strong count ≥ 2, so this never drops live geometry.
-                            if pipeline.wire_buffer_cache.len() > 16 {
-                                pipeline
-                                    .wire_buffer_cache
-                                    .retain(|_, (w, b, _)| {
-                                        std::sync::Arc::strong_count(w) > 1
-                                            || std::sync::Arc::strong_count(b) > 1
-                                    });
-                            }
                             entry
                         }
                     };
@@ -1000,6 +1022,7 @@ impl shader::Primitive for Primitive {
                 };
                 inner.upload_selected_wires(
                     device,
+                    queue,
                     &vp_wires[..],
                     &vp.selected_handles,
                     &vp.hover_handles,
@@ -1012,6 +1035,7 @@ impl shader::Primitive for Primitive {
                 // base text buffer.
                 inner.upload_text_highlight(
                     device,
+                    queue,
                     &vp_wires[..],
                     &vp.selected_handles,
                     &vp.hover_handles,
@@ -1066,7 +1090,7 @@ impl shader::Primitive for Primitive {
             // Live overlay (command preview / interim / grip drag) — small and
             // refreshed every frame it's present, so a drag never re-uploads
             // the resident base wire buffer.
-            inner.upload_preview_wires(device, &vp.preview_wires[..], &draw_depths);
+            inner.upload_preview_wires(device, queue, &vp.preview_wires[..], &draw_depths);
             inner.upload_preview_text(device, queue, &vp.preview_text_verts[..]);
             // Cull / scissor / LOD project AABBs relative-to-eye (matching the
             // GPU's RTE path) so the math stays precise at UTM-scale coords.
@@ -1087,13 +1111,14 @@ impl shader::Primitive for Primitive {
             if inner.silhouette_key != silhouette_key {
                 inner.upload_silhouettes(
                     device,
+                    queue,
                     if silhouette_enabled { &vp.meshes[..] } else { &[] },
                     vp.wire_content_id,
                     vp.view_dir,
                 );
                 inner.silhouette_key = silhouette_key;
             }
-            inner.upload_clip_boundary(device, &vp.clip_boundary_ndc);
+            inner.upload_clip_boundary(device, queue, &vp.clip_boundary_ndc);
             let hatch_lod_key = (
                 Arc::as_ptr(&vp.hatches) as usize,
                 vp.camera_generation,