Add 3DGS SIMD forward renderer plan

Author: AdaWorldAPI

.claude/plans/3DGS-SIMD-forward-renderer-plan.md

@@ -0,0 +1,132 @@
+# 3DGS SIMD Forward Renderer Plan — ndarray
+
+## Goal
+
+Productionize `ndarray::hpc::splat3d` as the CPU-SIMD forward renderer for 3D Gaussian Splatting.
+
+The renderer should be usable by higher layers as a deterministic kernel package:
+
+```text
+camera + splat block + render budget
+        ->
+projected splat footprints + optional framebuffer / visibility report
+```
+
+## Non-goals
+
+- Do not implement 3D Tiles parsing here.
+- Do not implement ArcGIS or Cesium service APIs here.
+- Do not own tile graph traversal policy here.
+- Do not require WGPU or GPU availability.
+
+## Core modules
+
+Target namespace:
+
+```text
+src/hpc/splat3d/
+  mod.rs
+  types.rs
+  camera.rs
+  covariance.rs
+  projection.rs
+  ewa.rs
+  raster.rs
+  visibility.rs
+  simd.rs
+  report.rs
+```
+
+## Required public DTOs
+
+```rust
+pub struct Splat3dCamera {
+    pub view: [[f32; 4]; 4],
+    pub proj: [[f32; 4]; 4],
+    pub viewport_width: u32,
+    pub viewport_height: u32,
+    pub near: f32,
+    pub far: f32,
+}
+
+pub struct Splat3dBlockView<'a> {
+    pub pos_x: &'a [f32],
+    pub pos_y: &'a [f32],
+    pub pos_z: &'a [f32],
+    pub scale_x: &'a [f32],
+    pub scale_y: &'a [f32],
+    pub scale_z: &'a [f32],
+    pub quat_w: &'a [f32],
+    pub quat_x: &'a [f32],
+    pub quat_y: &'a [f32],
+    pub quat_z: &'a [f32],
+    pub opacity: &'a [f32],
+    pub rgba: Option<&'a [[u8; 4]]>,
+}
+
+pub struct ProjectedSplat {
+    pub screen_x: f32,
+    pub screen_y: f32,
+    pub depth: f32,
+    pub radius_px: f32,
+    pub covariance_2d: [f32; 3],
+    pub opacity: f32,
+    pub valid: bool,
+}
+```
+
+## Hot path
+
+1. Load SoA splat columns.
+2. SIMD transform world coordinates to view space.
+3. Reject behind-camera, outside-near/far, NaN, invalid scale, invalid quaternion.
+4. Construct 3D SPD covariance from scale/quaternion.
+5. Push covariance through EWA sandwich.
+6. Compute projected screen footprint.
+7. Emit projected splat data or render into framebuffer.
+8. Return counters and failure reasons.
+
+## SIMD tiers
+
+Use existing ndarray SIMD dispatch style:
+
+- scalar baseline
+- AVX2/FMA
+- AVX-512
+- NEON
+- runtime-dispatch optional path
+
+Every SIMD kernel must have a scalar reference implementation and deterministic tests.
+
+## Error handling
+
+Do not panic on malformed splats in batch paths. Return counters:
+
+```rust
+pub struct Splat3dRenderReport {
+    pub input_count: usize,
+    pub projected_count: usize,
+    pub rejected_behind_camera: usize,
+    pub rejected_invalid_covariance: usize,
+    pub rejected_too_small: usize,
+    pub rejected_outside_view: usize,
+    pub max_radius_px: f32,
+    pub min_depth: f32,
+    pub max_depth: f32,
+}
+```
+
+## Acceptance criteria
+
+- `cargo test -p ndarray --features std,linalg,splat3d`
+- Scalar and SIMD paths match within explicit tolerances.
+- Invalid splats are counted, not fatal.
+- EWA covariance remains PSD or is rejected with reason.
+- Renderer can process a columnar block without heap allocation in the inner loop.
+- Benchmarks include 1k, 10k, 100k, and 1M splat projection-only workloads.
+
+## Follow-up hooks
+
+- Connect to `src/hpc/pillar/ewa_sandwich_3d.rs` for certification.
+- Expose projection-only mode for `lance-graph` tile preflight.
+- Expose framebuffer mode for CPU preview / headless validation.