Add 3DGS HHTL CPU cascade plan

Author: AdaWorldAPI

.claude/plans/3DGS-HHTL-CPU-cascade-plan.md

@@ -0,0 +1,139 @@
+# 3DGS HHTL CPU Cascade Plan — ndarray
+
+## Goal
+
+Adapt the existing HHTL / HEEL-HIP-TWIG-LEAF cascade ideas into reusable CPU-side kernels for 3DGS tile and splat preselection.
+
+This plan is not the full geospatial traversal policy. It defines fast kernels that higher layers can call.
+
+## Cascade shape
+
+```text
+HEEL  coarse tile/block rejection
+HIP   screen-space and density scoring
+TWIG  covariance/opacity/quantization refinement
+LEAF  exact projection/render payload
+```
+
+## Kernel responsibilities
+
+### HEEL
+
+Fast broad-phase rejection:
+
+- bounding sphere/box vs frustum
+- tile availability bitmask scan
+- distance band classification
+- optional Morton range locality scan
+
+### HIP
+
+Mid-tier priority estimation:
+
+- screen-space error estimate
+- projected tile/block radius
+- splat density estimate
+- opacity budget estimate
+- camera motion/fovea relaxation factor
+
+### TWIG
+
+Block-level refinement:
+
+- covariance stability estimate
+- quantization error estimate
+- depth bucket estimate
+- weak-dependence correction hook
+- certificate-ready stats
+
+### LEAF
+
+Exact hot path:
+
+- call `splat3d` projection kernel
+- optionally call raster kernel
+- emit exact counters and certificate values
+
+## DTOs
+
+```rust
+pub struct Hhtl3dgsRequest<'a> {
+    pub camera: Splat3dCamera,
+    pub block_bounds: &'a [BlockBounds],
+    pub block_stats: &'a [SplatBlockStats],
+    pub budget: Hhtl3dgsBudget,
+}
+
+pub struct Hhtl3dgsBudget {
+    pub max_error_px: f32,
+    pub min_confidence: f32,
+    pub max_projected_radius_px: f32,
+    pub max_blocks: usize,
+    pub allow_skip_lod: bool,
+}
+
+pub struct Hhtl3dgsDecision {
+    pub block_index: usize,
+    pub tier_reached: HhtlTier,
+    pub priority: f32,
+    pub estimated_error_px: f32,
+    pub action: HhtlAction,
+}
+
+pub enum HhtlAction {
+    Reject,
+    KeepCoarse,
+    Refine,
+    ProjectExact,
+    RenderExact,
+}
+```
+
+## SIMD strategy
+
+- Batch bounding-volume tests.
+- Batch distance-to-camera.
+- Batch projected radius approximations.
+- Use bitset outputs for rejected/accepted blocks.
+- Keep scalar reference path for every tier.
+
+## Integration with existing ndarray capabilities
+
+Use or extend existing modules where possible:
+
+- `hpc::cascade`
+- `hpc::clam`
+- `hpc::cam_pq`
+- `distance`, `byte_scan`, `spatial_hash`
+- `simd_dispatch`
+- `splat3d`
+- `pillar`
+
+## Runtime principles
+
+- No heap allocation in inner loops.
+- No hidden global state except existing SIMD dispatch tables.
+- No policy decisions that belong to `lance-graph`.
+- Return measurable reasons for rejection/refinement.
+
+## Acceptance criteria
+
+- Scalar and SIMD broad-phase decisions match.
+- Benchmarks for 1k, 10k, 100k block candidates.
+- HHTL decision output can be consumed without depending on renderer internals.
+- Exact LEAF mode can call `splat3d` projection and merge reports.
+- Test cases include camera motion and foveated relaxation inputs, even if policy remains in `lance-graph`.
+
+## Cross-repo usage
+
+`lance-graph` calls this cascade after it has selected candidate tiles from 3D Tiles / Lance metadata.
+
+```text
+lance-graph tile candidates
+        ->
+ndarray HHTL cascade
+        ->
+ranked block decisions
+        ->
+lance-graph traversal/render scheduler
+```