Changelog

v0.1.0 — Initial public release

What's in this release

Streaming ingest

• Bounded ring buffer with configurable capacity. Memory never grows beyond maxPoints, no matter the stream rate or duration.
• Off-thread ingest via a dedicated ingest worker. Typed-array payloads are Transferred (zero-copy) across the thread boundary.
• Optional worker-side frustum prefilter (workerCulling) to discard off-screen points at ingest time.

Rendering

• WebGPU compute path with importance-weighted GPU sampling. A WGSL compute shader runs frustum culling, temporal window filtering, and per-point importance sampling in one pass. drawIndirect means zero CPU draw preparation.
• Automatic WebGL fallback for browsers without WebGPU support.
• LOD with auto-detection by camera distance.
• Frustum culling on both paths.
• Adaptive refresh rate.

Importance engine

• Unified importance score drives both buffer eviction (K=16 lookahead) and per-frame render sampling (GPU PCG hash).
• importanceField: any numeric attribute channel.
• maxStalenessMs: exponential recency decay with configurable half-life.
• importanceSamplingEnabled: GPU stochastic sampling on WebGPU.
• fovStrength: foveated boost for points near screen center.
• accumulationMode: full-detail rendering when camera is static.

Static file loading

• <PointCloud> for PLY (binary LE/BE/ASCII), XYZ/CSV/TXT, and LAS 1.0–1.4. Off-thread parsing, progressive rendering.
• <PointCloudDropzone> for drag-and-drop workflows.
• LAZ (compressed LAS) via opt-in pointflow/laz subpath. laz-perf WASM inlined — no CDN dependency.
• Unified source API: src accepts string, URL, Request, File, or Blob.
• onLoadControls for imperative cancellation.

COPC streaming

• <CopcPointCloud> from pointflow/copc. HTTP range requests, LRU tile cache, frustum+LOD selection.
• Prefetch strategies: frustum-priority, depth-first, nearest, bandwidth-saver.
• OPFS persistent tile cache. Decoded tiles survive page reload with zero network cost.
• Configurable GPU atlas tiers for memory tuning.

Transport adapters

• createWebSocketAdapter — JSON PointChunk messages.
• createSSEAdapter — Server-Sent Events.
• createRosbridgeAdapter — rosbridge v2 / sensor_msgs/PointCloud2.
• createQuantizedAdapter — 6 bytes/point binary encoding.
• mergeChunkStreams / withSourceTag — multi-source merging.

Advanced features

• CPU point picking. onPointPick callback with configurable radius and strategy.
• Temporal time window. GPU compute filter on WebGPU, zero CPU overhead.
• Dynamic buffer allocation. Buffers grow from initialCapacity toward maxPoints.
• pointflow/config module for shared defaults across components and hooks.
• Reproducible benchmark surface with fixed profiles, rolling p95, and JSON export.

Performance (i7-13700HX, RTX 4060 Laptop, Chrome 147, Windows 11)

Metric	Result
FPS (WebGPU, balanced, 50k points)	163–166
Rolling p95 frame time	< 50 ms
Ingest throughput	20,000 pts/s
Test suite	393 passing, 0 type errors

Package structure

pointflow          — main package (streaming + static loading)
pointflow/laz      — opt-in LAZ decoder (laz-perf WASM)
pointflow/copc     — opt-in COPC renderer
pointflow/config   — shared config module

Breaking changes

None — this is the first public release.

---

For the full development history, see the GitHub repository.