Reduce GPU text memory, page-switch and resize latency

Directory.Packages.props

@@ -10,12 +10,14 @@
     <PackageVersion Include="Avalonia.Themes.Fluent" Version="12.0.5" />
     <PackageVersion Include="DocumentFormat.OpenXml" Version="3.5.1" />
     <PackageVersion Include="Markdig" Version="0.37.0" />
+    <PackageVersion Include="Microsoft.Diagnostics.Tracing.TraceEvent" Version="3.1.15" />
     <PackageVersion Include="Microsoft.NET.Test.Sdk" Version="17.11.1" />
     <PackageVersion Include="netDxf.netstandard" Version="2.4.0" />
     <PackageVersion Include="OpenFontSharp" Version="1.0.0" />
     <PackageVersion Include="ProDiagnostics" Version="12.0.0" />
     <PackageVersion Include="Silk.NET.Input" Version="2.23.0" />
     <PackageVersion Include="Silk.NET.Input.Common" Version="2.23.0" />
+    <PackageVersion Include="Silk.NET.Input.Glfw" Version="2.23.0" />
     <PackageVersion Include="Silk.NET.GLFW" Version="2.23.0" />
     <PackageVersion Include="Silk.NET.WebGPU" Version="2.23.0" />
     <PackageVersion Include="Silk.NET.WebGPU.Native.WGPU" Version="2.23.0" />

docs/SAMPLE_PAGE_MEMORY_PROFILE.md

@@ -0,0 +1,196 @@
+# Release sample-page memory profile
+
+## Outcome
+
+The complete desktop gallery was measured page-by-page in fresh Release processes before
+and after the managed-memory work. Across all 47 pages, average retained managed memory
+fell from 80.51 MiB to 30.00 MiB (**62.7% lower**), average managed allocation fell from
+154.10 KiB/frame to 43.29 KiB/frame (**71.9% lower**), and average compositor compilation
+fell from 1.223 ms to 0.362 ms (**70.4% lower**). The sum of the 47 isolated retained-heap
+measurements fell from 3.70 GiB to 1.38 GiB. This sum is useful for comparing the complete
+gallery but is not the memory of one process because every page was deliberately isolated.
+
+The worst retained managed heap changed from Text Shaping Lab at 513.91 MiB to Text &
+Documents at 177.46 MiB. Average macOS physical footprint, which also includes the runtime,
+native WebGPU/Metal objects, driver caches, and mapped resources, fell from 403.90 MiB to
+370.82 MiB (**8.2% lower**). Compile frames over the 16.67 ms budget fell from 428 to 3;
+426 of the baseline frames were from Text & Documents.
+
+| Representative page | Managed before | Managed after | Change | Allocation/frame before | Allocation/frame after | Compile before | Compile after |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| Text Shaping Lab | 513.91 MiB | 70.04 MiB | **-86.4%** | 22.8 KiB | 16.9 KiB | 0.082 ms | 0.052 ms |
+| Typography & Scripts | 345.18 MiB | 82.17 MiB | **-76.2%** | 10.2 KiB | 8.4 KiB | 0.036 ms | 0.055 ms |
+| Text & Documents | 472.28 MiB | 177.46 MiB | **-62.4%** | 4,128.4 KiB | 197.8 KiB | 41.728 ms | 2.738 ms |
+| Font Glyph Browser | 60.85 MiB | 31.22 MiB | **-48.7%** | 104.9 KiB | 53.0 KiB | 2.491 ms | 0.899 ms |
+
+## Measurement method
+
+`tools/profile-sample-memory.sh` discovers the gallery pages, builds the desktop application
+and analyzer in Release, and launches a new traced child process for each page. Each process
+runs 180 scrolling warm-up frames, performs a blocking compacting full collection, measures
+600 scrolling frames, performs another compacting collection, writes its metrics, and exits.
+This avoids navigation order, previously loaded fonts, and retained driver resources
+contaminating another page's result.
+
+The benchmark and analyzer record:
+
+- exact process-wide managed allocation from `GC.GetTotalAllocatedBytes`, retained managed
+  bytes, GC heap/committed/fragmented bytes, generations, collection counts, pause duration,
+  pinned objects, and finalization state;
+- working set and virtual memory plus macOS `proc_pid_rusage` resident, wired, current physical
+  footprint, and lifetime maximum physical footprint;
+- explicit glyph, color-glyph, and path texture residency, staging capacity, outline buffers,
+  atlas entries/generations/evictions, upload writes, raster batches, and compute passes;
+- compositor, upload, render, host-update, layout, animation, acquire, present, frame-budget,
+  and scene-cache measurements;
+- EventPipe CPU samples and randomized allocation samples separated into startup, warm-up,
+  and measurement phases by `ProGPU-SampleBenchmark` EventSource markers.
+
+Randomized allocation events are converted to probability-weighted estimates using the .NET
+runtime's documented `size / (1 - q^size)` estimator with `p = 1 / 102400` and `q = 1 - p`.
+The exact allocation/frame counter is the comparison metric; sampled types and stacks are
+used for attribution because a small number of samples has wide statistical error. The
+profiler retains JSON and logs for every page and can optionally retain `.nettrace` files.
+
+Run it with:
+
+```bash
+PROGPU_MEMORY_KEEP_TRACES=0 tools/profile-sample-memory.sh artifacts/sample-memory-profile
+```
+
+Use `PROGPU_MEMORY_PAGE_FILTER` for a regular-expression page subset and
+`PROGPU_MEMORY_WARMUP_FRAMES` / `PROGPU_MEMORY_MEASURE_FRAMES` to change the workload.
+
+## Attribution and implementation
+
+The baseline GC dumps for the three largest pages showed that `OpenFontSharp` whole-font
+CFF and OpenType layout object graphs retained glyph objects and decoded arrays for fonts
+whose pages had requested only a small set of glyphs. Allocation stacks on Text & Documents
+also showed a second shaping/layout pass during rendering and per-realization rich-character
+objects. CPU samples attributed its severe frame stalls to quadratic free-rectangle pruning
+during a path-atlas recovery containing more than one thousand live entries.
+
+The implementation therefore:
+
+1. Reads GSUB/GPOS feature records and uses the existing typed raw-table shaping pipeline,
+   removing the eager whole-font layout typeface.
+2. Adds an original, clean-room CFF 1 INDEX/DICT and Type 2 evaluator. It retains compact
+   offsets and evaluates only a requested glyph, including local/global subroutines and CID
+   FDSelect/FDArray data. Deprecated `seac` and unsupported malformed programs fall back to
+   the existing parser lazily, so compatibility is preserved without paying its whole-font
+   cost on normal fonts.
+3. Reuses the already-shaped cluster advances for rich-text decoration widths instead of
+   constructing another `TextLayout`; caches immutable common shaping options; and recycles
+   positioned rich characters through a presenter-local pool capped at 4,096 objects.
+4. Keeps MaxRects recovery deterministic but compacts unaffected free regions linearly and
+   compares only newly split regions with the pruned set. Worst-case complexity remains
+   `O(F^2)`, while the measured bounded-intersection case is `O(F)` and uses one pooled
+   temporary array.
+5. Preserves the earlier page-switch solution: newly visible glyphs are demand-driven and
+   coalesced into bounded outline/uniform writes, one compute pass, and one submission per
+   batch. Coverage atlases survive navigation and grow geometrically on demand, avoiding
+   both hidden-page preloading and repeated uploads.
+
+## Cross-engine design record
+
+This is a clean-room implementation. No external implementation source was copied, ported,
+translated, or structurally reproduced. Primary sources were used to select behavior and
+architecture; the implementation follows ProGPU's own typed retained-scene, invalidation,
+atlas-generation, and GPU ownership contracts.
+
+| Source | Concept considered | ProGPU decision |
+|---|---|---|
+| [Skia strike cache](https://skia.googlesource.com/skia/+/main/src/core/SkStrikeCache.h), [SkParagraph](https://skia.googlesource.com/skia/+/refs/heads/main/modules/skparagraph/include/Paragraph.h), [GrDrawOpAtlas](https://skia.googlesource.com/skia/+/main/src/gpu/ganesh/GrDrawOpAtlas.h) | Bounded caches, retained layout, lazy atlas residency and eviction | Adopt bounded residency and retained CPU layout; adapt it to analytic R8 WebGPU coverage and ProGPU generation-based invalidation. |
+| [DirectWrite glyph-run analysis](https://learn.microsoft.com/en-us/windows/win32/api/dwrite/nn-dwriteglyphrunanalysis), [Direct2D performance guidance](https://learn.microsoft.com/en-us/windows/win32/direct2d/improving-direct2d-performance), [Win2D CanvasTextLayout](https://microsoft.github.io/Win2D/WinUI3/html/T_Microsoft_Graphics_Canvas_Text_CanvasTextLayout.htm) | Reusable glyph runs/layout and alpha coverage | Keep shaping/layout as reusable CPU results and upload compact coverage only when visible. |
+| [WebRender overview](https://searchfox.org/firefox-main/source/gfx/docs/RenderingOverview.rst), [blob images](https://searchfox.org/firefox-main/source/gfx/wr/webrender/doc/blob.md) | Visible-resource preparation and delayed texture-cache upload | Batch only current-scene misses before drawing; retain valid atlas entries across page switches. |
+| [Vello glyph design](https://github.com/linebender/vello/issues/204), [Parley](https://docs.rs/parley/latest/parley/) | GPU vector work paired with reusable CPU shaping/layout | Preserve CPU Unicode/OpenType shaping and GPU raster/composition; reject incomplete GPU-only shaping. |
+| [HarfBuzz shaping-plan caching](https://harfbuzz.github.io/shaping-plans-and-caching.html) | Cache stable face/script/language/feature decisions | Reuse immutable common shaping options and raw table plans; avoid rebuilding parsed font graphs. |
+| [FreeType glyph retrieval](https://freetype.org/freetype2/docs/reference/ft2-glyph_retrieval.html) | Load one requested glyph into a replaceable slot | Adopt the observable demand-loading behavior, implemented independently from the [CFF table specification](https://learn.microsoft.com/en-us/typography/opentype/spec/cff) and [Adobe Type 2 Charstring specification](https://adobe-type-tools.github.io/font-tech-notes/pdfs/5177.Type2.pdf). |
+| [.NET randomized allocation sampling](https://github.com/dotnet/runtime/blob/main/docs/design/features/RandomizedAllocationSampling.md), [dotnet-trace](https://learn.microsoft.com/en-us/dotnet/core/diagnostics/dotnet-trace), [`GC.GetGCMemoryInfo`](https://learn.microsoft.com/en-us/dotnet/api/system.gc.getgcmemoryinfo?view=net-10.0) | Phase markers, randomized allocation attribution, and GC/process counters | Combine exact counters with sampled types/stacks; never treat a low-sample allocation estimate as an exact total. |
+
+Rejected alternatives were eager parsing of every glyph and layout table, prewarming every
+hidden sample page, an unbounded exact-position atlas key, lowering raster quality, or moving
+Unicode/OpenType shaping to an incomplete GPU implementation. Those options either increase
+retained memory, recreate page-switch upload spikes, break the bounded cache contract, or
+reduce correctness and quality.
+
+## Complete 47-page comparison
+
+All values below are from the same 180-frame warm-up and 600-frame measured Release workload.
+Allocation is exact process-wide managed allocation per measured frame. Small differences
+below roughly 0.05 ms are generally timer/driver noise; the frame-budget count and full-suite
+average are the regression gates.
+
+| Page | Managed before | Managed after | Change | Alloc KiB/frame before | Alloc KiB/frame after | Compile ms before | Compile ms after |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| 3D Mesh Viewer | 52.30 | 22.50 | -57.0% | 19.1 | 13.5 | 0.056 | 0.041 |
+| Advanced Controls | 50.61 | 20.79 | -58.9% | 12.1 | 7.3 | 0.042 | 0.025 |
+| Basic Input | 50.03 | 20.23 | -59.6% | 10.9 | 6.9 | 0.038 | 0.020 |
+| Color Picker | 50.50 | 20.55 | -59.3% | 7.8 | 7.0 | 0.010 | 0.013 |
+| Compositor API | 50.85 | 20.84 | -59.0% | 14.6 | 9.1 | 0.744 | 0.694 |
+| Compute FX | 50.88 | 20.84 | -59.1% | 41.4 | 35.8 | 0.019 | 0.026 |
+| Data Virtualization | 54.10 | 24.06 | -55.5% | 37.8 | 32.0 | 1.146 | 0.919 |
+| Dock Panel | 49.39 | 19.50 | -60.5% | 6.8 | 6.5 | 0.008 | 0.023 |
+| Drawing Context | 49.45 | 19.38 | -60.8% | 11.0 | 5.3 | 0.038 | 0.027 |
+| DXF CAD Viewer | 51.97 | 21.00 | -59.6% | 16.3 | 10.8 | 0.028 | 0.027 |
+| File Storage | 49.61 | 19.75 | -60.2% | 11.9 | 6.5 | 0.042 | 0.032 |
+| Font Glyph Browser | 60.85 | 31.22 | -48.7% | 104.9 | 53.0 | 2.491 | 0.899 |
+| Framework Effects | 51.05 | 18.78 | -63.2% | 10.2 | 9.8 | 0.351 | 0.324 |
+| GDI Shim Showcase | 51.06 | 21.11 | -58.7% | 10.8 | 9.2 | 0.034 | 0.043 |
+| Glyph Run Showcase | 50.40 | 20.60 | -59.1% | 10.3 | 6.8 | 0.299 | 0.439 |
+| GPU Charting | 97.22 | 68.70 | -29.3% | 18.2 | 22.2 | 0.021 | 0.058 |
+| Grid Splitter | 49.48 | 19.56 | -60.5% | 7.2 | 6.0 | 0.012 | 0.030 |
+| Image & Buttons | 50.87 | 21.02 | -58.7% | 40.8 | 33.6 | 0.023 | 0.024 |
+| Image Effects | 50.70 | 20.67 | -59.2% | 37.6 | 35.8 | 0.013 | 0.016 |
+| Inter Typeface | 99.93 | 53.02 | -46.9% | 36.3 | 35.1 | 0.500 | 0.898 |
+| Interactive Input | 60.09 | 21.79 | -63.7% | 7.9 | 7.6 | 0.012 | 0.039 |
+| Keyboard & Focus | 52.60 | 23.03 | -56.2% | 15.5 | 10.5 | 0.037 | 0.036 |
+| Layout Panels | 50.51 | 20.03 | -60.4% | 12.4 | 8.1 | 0.026 | 0.026 |
+| LOL/s Benchmark | 65.83 | 36.98 | -43.8% | 1,994.6 | 1,073.1 | 2.999 | 2.831 |
+| Markdown Playground | 89.72 | 34.78 | -61.2% | 26.6 | 12.2 | 1.131 | 1.351 |
+| Motion & Animations | 50.40 | 18.13 | -64.0% | 20.4 | 13.8 | 0.603 | 0.516 |
+| MotionMark Showcase | 59.40 | 30.58 | -48.5% | 81.8 | 44.3 | 1.123 | 0.503 |
+| Password Box | 50.42 | 20.63 | -59.1% | 8.1 | 6.1 | 0.016 | 0.021 |
+| Path Operations | 49.96 | 19.99 | -60.0% | 8.1 | 6.5 | 0.021 | 0.043 |
+| Picture Caching | 51.18 | 21.29 | -58.4% | 21.7 | 13.9 | 0.802 | 0.592 |
+| Radio Button | 50.43 | 20.63 | -59.1% | 12.0 | 7.7 | 0.036 | 0.029 |
+| Rating Control | 50.28 | 20.50 | -59.2% | 8.0 | 6.4 | 0.020 | 0.029 |
+| Rich Document Editor | 50.32 | 20.38 | -59.5% | 12.8 | 7.7 | 0.032 | 0.023 |
+| ShaderToy Playground | 51.48 | 21.47 | -58.3% | 148.0 | 32.6 | 0.707 | 0.780 |
+| SkiaSharp Shim | 51.80 | 21.77 | -58.0% | 5.9 | 5.4 | 0.009 | 0.028 |
+| SplitView Layout | 50.14 | 20.32 | -59.5% | 13.0 | 7.5 | 0.033 | 0.037 |
+| Styles Showcase | 49.41 | 19.50 | -60.5% | 7.3 | 5.7 | 0.028 | 0.016 |
+| Text & Documents | 472.28 | 177.46 | -62.4% | 4,128.4 | 197.8 | 41.728 | 2.738 |
+| Text Shaping Lab | 513.91 | 70.04 | -86.4% | 22.8 | 16.9 | 0.082 | 0.052 |
+| Theme Showcase | 52.77 | 22.96 | -56.5% | 15.1 | 16.8 | 0.043 | 0.041 |
+| Touch & Gestures | 50.47 | 21.00 | -58.4% | 12.7 | 7.1 | 0.042 | 0.036 |
+| Typography & Scripts | 345.18 | 82.17 | -76.2% | 10.2 | 8.4 | 0.036 | 0.055 |
+| Vector Shapes | 50.67 | 20.73 | -59.1% | 7.5 | 6.2 | 0.013 | 0.018 |
+| Virtualization Controls | 55.09 | 26.66 | -51.6% | 22.9 | 20.6 | 0.016 | 0.014 |
+| Visual Designer | 81.26 | 51.92 | -36.1% | 129.5 | 124.0 | 1.906 | 2.537 |
+| WPF Shim Showcase | 57.22 | 21.26 | -62.9% | 13.9 | 8.3 | 0.031 | 0.020 |
+| Wrap Panel | 49.78 | 19.82 | -60.2% | 9.2 | 7.5 | 0.014 | 0.023 |
+
+GPU Charting and Theme Showcase had small allocation increases in their isolated optimized
+runs (about 4 KiB/frame and 1.7 KiB/frame respectively), while their retained managed heaps
+still fell by 29.3% and 56.5%. Their absolute compile costs stayed below 0.06 ms. Visual
+Designer rose by 0.63 ms but remained at 2.54 ms with no over-budget frames. These small
+single-run movements are retained in the table rather than hidden by aggregate reporting.
+
+## Verification
+
+- `ProGPU.Tests` Release: **2,274 passed**, 0 failed.
+- `ProGPU.Tests.Headless` Release: **196 passed**, 0 failed.
+- Path-atlas focused tests: **29 passed**, including recovery, generation, and small-atlas
+  stress coverage.
+- CFF differential tests compare independently decoded `A`, `日`, `か`, and `ナ` geometry
+  against the established fallback to 0.001 font units; on-demand tests confirm the fallback
+  object graph is not created.
+- All 47 pages completed the fresh-process Release desktop profile.
+- Headless screenshots for Text Shaping Lab, Typography & Scripts, and Text & Documents were
+  visually inspected with unchanged text coverage and quality.
+- Browser Release publish AOT-compiled 69 assemblies; the WebGPU gallery opened under a real
+  browser with zero console errors or warnings beyond expected system-font fallback messages.
+- Quality-sensitive 8x8 analytic coverage, physical-DPI raster size, quarter-physical-pixel
+  snapping, direction-aware half-open winding, color glyphs, fallback/variation identity,
+  and same-frame atlas recovery were not reduced or bypassed.

src/ProGPU.Backend/GpuCoverageUpload.cs

@@ -0,0 +1,96 @@
+using Silk.NET.WebGPU;
+
+namespace ProGPU.Backend;
+
+/// <summary>
+/// Records copies from GPU-computed, tightly packed coverage bytes into a filterable R8 atlas.
+/// Rasterization remains GPU-only while avoiding the four-channel storage-texture footprint.
+/// </summary>
+public static unsafe class GpuCoverageUpload
+{
+    public const uint CopyRowAlignment = 256;
+
+    public static uint GetBytesPerRow(uint width)
+    {
+        if (width == 0)
+        {
+            return 0;
+        }
+
+        return checked((width + CopyRowAlignment - 1) & ~(CopyRowAlignment - 1));
+    }
+
+    public static uint GetRequiredBytes(uint width, uint height) =>
+        checked(GetBytesPerRow(width) * height);
+
+    public static void RecordCopy(
+        WgpuContext context,
+        CommandEncoder* encoder,
+        GpuBuffer source,
+        uint sourceOffset,
+        uint bytesPerRow,
+        GpuTexture destination,
+        uint destinationX,
+        uint destinationY,
+        uint width,
+        uint height)
+    {
+        ArgumentNullException.ThrowIfNull(context);
+        ArgumentNullException.ThrowIfNull(source);
+        ArgumentNullException.ThrowIfNull(destination);
+        if (encoder == null)
+        {
+            throw new ArgumentNullException(nameof(encoder));
+        }
+        if (destination.Format != TextureFormat.R8Unorm)
+        {
+            throw new ArgumentException("Coverage copies require an R8Unorm destination.", nameof(destination));
+        }
+        if (!source.Usage.HasFlag(BufferUsage.CopySrc))
+        {
+            throw new ArgumentException("Coverage staging buffers require CopySrc usage.", nameof(source));
+        }
+        if (width == 0 || height == 0)
+        {
+            return;
+        }
+        if (bytesPerRow < width || bytesPerRow % CopyRowAlignment != 0)
+        {
+            throw new ArgumentOutOfRangeException(nameof(bytesPerRow));
+        }
+
+        var copySource = new ImageCopyBuffer
+        {
+            Buffer = source.BufferPtr,
+            Layout = new TextureDataLayout
+            {
+                Offset = sourceOffset,
+                BytesPerRow = bytesPerRow,
+                RowsPerImage = height
+            }
+        };
+        var copyDestination = new ImageCopyTexture
+        {
+            Texture = destination.TexturePtr,
+            MipLevel = 0,
+            Origin = new Origin3D
+            {
+                X = destinationX,
+                Y = destinationY,
+                Z = 0
+            },
+            Aspect = TextureAspect.All
+        };
+        var copySize = new Extent3D
+        {
+            Width = width,
+            Height = height,
+            DepthOrArrayLayers = 1
+        };
+        context.Api.CommandEncoderCopyBufferToTexture(
+            encoder,
+            &copySource,
+            &copyDestination,
+            &copySize);
+    }
+}