feat: ADR-023 full DensePose training pipeline (Phases 1-8)
Implement complete WiFi CSI-to-DensePose neural network pipeline:
Phase 1 - Dataset loaders: .npy/.mat v5 parsers, MM-Fi + Wi-Pose
loaders, subcarrier resampling (114->56, 30->56), DataPipeline
Phase 2 - Graph transformer: COCO BodyGraph (17 kp, 16 edges),
AntennaGraph, multi-head CrossAttention, GCN message passing,
CsiToPoseTransformer full pipeline
Phase 4 - Training loop: 6-term composite loss (MSE, cross-entropy,
UV regression, temporal consistency, bone length, symmetry),
SGD+momentum, cosine+warmup scheduler, PCK/OKS metrics, checkpoints
Phase 5 - SONA adaptation: LoRA (rank-4, A*B delta), EWC++ Fisher
regularization, EnvironmentDetector (3-sigma drift), temporal
consistency loss
Phase 6 - Sparse inference: NeuronProfiler hot/cold partitioning,
SparseLinear (skip cold rows), INT8/FP16 quantization with <0.01
MSE, SparseModel engine, BenchmarkRunner
Phase 7 - RVF pipeline: 6 new segment types (Index, Overlay, Crypto,
WASM, Dashboard, AggregateWeights), HNSW index, OverlayGraph,
RvfModelBuilder, ProgressiveLoader (3-layer: A=instant, B=hot, C=full)
Phase 8 - Server integration: --model, --progressive CLI flags,
4 new REST endpoints, WebSocket pose_keypoints + model_status
229 tests passing (147 unit + 48 bin + 34 integration)
Benchmark: 9,520 frames/sec (105μs/frame), 476x real-time at 20 Hz
7,832 lines of pure Rust, zero external ML dependencies
Co-Authored-By: claude-flow <ruv@ruv.net>
fc409dfd6a