packed_spatial_index 0.26.0

Packed static spatial index (Hilbert R-tree) for 2D/3D AABBs — SIMD range, kNN, raycast, and spatial-join queries, with zero-copy and streaming serialization.
Documentation
# Internals: how the engine goes fast

Deep-dives into the techniques behind the query kernels — for contributors and the
curious. These explain *how* a result is produced quickly. For *what* each query
does and when to use it, see the [guide](../guide.md). For the numbers, see
[performance.md](../performance.md).

- **[simd.md]simd.md** — runtime-dispatched SIMD range search, visit and raycast
  (AVX-512 `VPCOMPRESSQ`, the AVX2 left-pack that emulates the missing compress,
  the SSE2 fallback). The "path to the black belt" of the box-test kernels.
- **[knn.md]knn.md** — nearest-neighbor traversal as two-queue *distance
  browsing* (Hjaltason & Samet): why kNN keeps a node queue and an item queue
  separate, the k=1 fast path, and how custom metrics and f32 refinement reuse the
  one collect kernel.
- **[prefetch.md]prefetch.md** — hiding the cold cache miss on every tree node by
  prefetching the next node's box while the current one is hit-tested. A free
  latency hint on range search and all-hits raycast.
- **[payload-prefix-scans.md]payload-prefix-scans.md** — why reading the first
  few bytes of every matching payload used to cost one range request per match
  over object storage, why neither coalescing setting alone fixes it, and the
  contiguous `PFIX` section that does. The body size at which each tool starts
  paying, measured.