perf: account for the device memory the renderer holds

Stage 0 of the roadmap asks for a memory baseline before any more
optimisation, and explicitly not from RSS: the allocator and the driver
both retain high-water memory that RSS cannot tell apart from live
resources.

`gpu-live` counts from the sizes this renderer allocated, in the
categories large enough to decide where the next fix goes:

    slots  total  shadow  targets  text_atlas  wire_arena  block_geometry

It is partial by construction and says so, rather than presenting itself
as a total. Hatches, images, meshes and Face3D are not counted yet.

Three of the numbers needed a size that was not reachable before: the
block-geometry chunk records its bytes at build because in storage mode
the buffer belongs to the bind group, the arena exposes its allocation
rather than its occupied prefix — headroom is what a budget has to answer
for — and the text atlas reports its `R8` texel count.

Emitted only when a number moves. `prepare` runs thousands of times in a
session and these change a handful of times, so a per-frame line would
bury the log it is meant to explain.

This answers the question the previous fix left open: whether the arena's
52 MiB per slot, the 16 MiB shadow map every slot allocates whether or not
shadows are on, or the render targets dominate once a paper layout has
four or five viewports on screen.

Co-Authored-By: Claude Opus 5 <[email protected]>

src/scene/pipeline/mod.rs

@@ -4727,6 +4727,80 @@ fn aabb_below_pixel(
     (max_px - min_px).max(max_py - min_py) < threshold_px
 }
 
+/// Device bytes a renderer is holding, counted from the sizes it allocated
+/// rather than from process RSS — the allocator and the driver both keep
+/// high-water memory that RSS cannot distinguish from live resources.
+///
+/// Partial by construction: it covers the categories large enough to decide
+/// where the next fix goes, and says so rather than pretending to be a total.
+#[derive(Default, Clone, Copy, PartialEq)]
+pub(crate) struct GpuLiveBytes {
+    pub slots: usize,
+    pub shadow: u64,
+    pub render_targets: u64,
+    pub text_atlas: u64,
+    pub wire_arena: u64,
+    pub block_geometry: u64,
+}
+
+impl GpuLiveBytes {
+    pub(crate) fn total(&self) -> u64 {
+        self.shadow + self.render_targets + self.text_atlas + self.wire_arena + self.block_geometry
+    }
+}
+
+impl Pipeline {
+    /// This slot's share. `alloc_size` is the real allocation, which is rounded
+    /// up from the requested size and grows only.
+    fn gpu_live_bytes(&self) -> GpuLiveBytes {
+        // 4x MSAA colour (4 B/sample) + 4x MSAA depth-stencil (4 B/sample) +
+        // one single-sample resolve target.
+        const BYTES_PER_PIXEL: u64 = (MSAA_SAMPLES as u64) * 4 + (MSAA_SAMPLES as u64) * 4 + 4;
+        let pixels = u64::from(self.alloc_size.width) * u64::from(self.alloc_size.height);
+        GpuLiveBytes {
+            slots: 1,
+            // Allocated by every slot whether or not shadows are enabled.
+            shadow: u64::from(SHADOW_MAP_SIZE) * u64::from(SHADOW_MAP_SIZE) * 4,
+            render_targets: pixels * BYTES_PER_PIXEL,
+            text_atlas: self
+                .text_atlas_gpu
+                .as_ref()
+                .map(text_gpu::TextAtlasGpu::gpu_bytes)
+                .unwrap_or(0),
+            wire_arena: self
+                .wire_arena
+                .as_ref()
+                .map(wire_arena::PersistentWireArena::gpu_bytes)
+                .unwrap_or(0)
+                + self
+                    .wire_arena_mesh
+                    .as_ref()
+                    .map(wire_arena::PersistentWireArena::gpu_bytes)
+                    .unwrap_or(0),
+            block_geometry: 0,
+        }
+    }
+}
+
+impl MultiPipeline {
+    /// Sum across slots, plus the caches held once for all of them.
+    pub(crate) fn gpu_live_bytes(&self) -> GpuLiveBytes {
+        let mut total = GpuLiveBytes {
+            block_geometry: wire_gpu::block_geometry_bytes(&self.block_geometry),
+            ..Default::default()
+        };
+        for inner in &self.inners {
+            let slot = inner.gpu_live_bytes();
+            total.slots += slot.slots;
+            total.shadow += slot.shadow;
+            total.render_targets += slot.render_targets;
+            total.text_atlas += slot.text_atlas;
+            total.wire_arena += slot.wire_arena;
+        }
+        total
+    }
+}
+
 /// Bind one block-wire batch and draw it.
 ///
 /// The two pipeline modes lay the vertex slots out differently — packed puts

src/scene/pipeline/text_gpu.rs

@@ -221,6 +221,11 @@ pub struct TextAtlasGpu {
 }
 
 impl TextAtlasGpu {
+    /// Device bytes this atlas holds. `R8`, so one byte per texel.
+    pub fn gpu_bytes(&self) -> u64 {
+        u64::from(self._texture.width()) * u64::from(self._texture.height())
+    }
+
     /// Bind-group layout for group 1: `{ R8 atlas texture, linear sampler }`.
     pub fn bind_group_layout(device: &wgpu::Device) -> wgpu::BindGroupLayout {
         device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {

src/scene/pipeline/wire_arena.rs

@@ -1369,6 +1369,18 @@ impl PersistentWireArena {
         }
     }
 
+    /// Device bytes this arena holds. The instance buffer is reserved with
+    /// headroom, so this is the allocation, not the occupied prefix — which is
+    /// what a memory budget has to answer for.
+    pub fn gpu_bytes(&self) -> u64 {
+        match &self.inner {
+            PersistentWireArenaKind::Indexed(arena) => {
+                arena.inst_buf.size() + arena.const_buf.size()
+            }
+            PersistentWireArenaKind::Packed(arena) => arena.inst_buf.size(),
+        }
+    }
+
     pub fn wire_gpus_visible(
         &self,
         view_rot: glam::Mat4,

src/scene/pipeline/wire_gpu.rs

@@ -703,6 +703,15 @@ pub struct BlockWireTarget<'a> {
     pub mode: WirePipelineMode,
 }
 
+/// Device bytes held by a whole block-geometry cache.
+pub fn block_geometry_bytes(cache: &BlockGeometryCache) -> u64 {
+    cache
+        .values()
+        .flat_map(|chunks| chunks.iter())
+        .map(|chunk| chunk.gpu_bytes)
+        .sum()
+}
+
 /// `(source_id, mesh_edge, colour bits, base translation bits)`.
 pub type BlockGeometryKey = (u64, bool, [u32; 4], [u64; 3]);
 
@@ -722,6 +731,9 @@ pub struct BlockGeometryChunk {
     pub vertex_buffer: Option<wgpu::Buffer>,
     pub vertex_count: u32,
     pub bind_group: std::sync::Arc<wgpu::BindGroup>,
+    /// Device bytes this chunk holds. Recorded at build because in storage mode
+    /// the buffer is owned by `bind_group` and cannot be measured from here.
+    pub gpu_bytes: u64,
 }
 
 impl BlockWireGpu {
@@ -856,6 +868,7 @@ impl BlockWireGpu {
                         BlockGeometryChunk {
                             vertex_buffer: None,
                             vertex_count: chunk.len() as u32 * 6,
+                            gpu_bytes: std::mem::size_of_val(chunk) as u64,
                             bind_group: std::sync::Arc::new(device.create_bind_group(
                                 &wgpu::BindGroupDescriptor {
                                     label: Some("block_wire.const.bg"),
@@ -897,6 +910,7 @@ impl BlockWireGpu {
                 vertices
                     .chunks(max_verts)
                     .map(|chunk| BlockGeometryChunk {
+                        gpu_bytes: std::mem::size_of_val(chunk) as u64,
                         vertex_buffer: Some(super::gpu_upload::upload_buffer(
                             device,
                             queue,

src/scene/view/render.rs

@@ -1217,6 +1217,7 @@ impl shader::Primitive for Primitive {
                 self.viewports.len(),
             );
         }
+        report_gpu_live(pipeline);
     }
 
     fn render(
@@ -1568,6 +1569,42 @@ fn solar_direction(
 
 // ── Render-style helpers (impl Scene) ────────────────────────────────────
 
+/// Report what the renderer is holding on the device, when it changes.
+///
+/// Only on change: `prepare` runs thousands of times a session and these
+/// numbers move a handful of times. RSS is deliberately not used — the
+/// allocator and the driver both retain high-water memory that RSS cannot tell
+/// apart from live resources.
+fn report_gpu_live(pipeline: &crate::scene::pipeline::MultiPipeline) {
+    use std::cell::Cell;
+    if !crate::perf::enabled() {
+        return;
+    }
+    thread_local! {
+        static LAST: Cell<Option<crate::scene::pipeline::GpuLiveBytes>> = const { Cell::new(None) };
+    }
+    let now = pipeline.gpu_live_bytes();
+    LAST.with(|last| {
+        if last.get() == Some(now) {
+            return;
+        }
+        last.set(Some(now));
+        const MIB: f64 = 1024.0 * 1024.0;
+        let mib = |bytes: u64| bytes as f64 / MIB;
+        crate::perf_record!(
+            "[perf] gpu-live slots={} total={:.1}MiB shadow={:.1} targets={:.1} \
+text_atlas={:.1} wire_arena={:.1} block_geometry={:.1}",
+            now.slots,
+            mib(now.total()),
+            mib(now.shadow),
+            mib(now.render_targets),
+            mib(now.text_atlas),
+            mib(now.wire_arena),
+            mib(now.block_geometry),
+        );
+    });
+}
+
 impl Scene {
     fn model_tile_vport(&self, index: usize) -> Option<&acadrust::tables::VPort> {
         let rect = self.model_tiles.borrow().get(index)?.rect;