<summary><strong>🧠 Models & Training</strong> — DensePose pipeline, RVF containers, SONA adaptation</summary>

The neural pipeline uses a graph transformer with cross-attention to map CSI feature matrices to 17 COCO body keypoints and DensePose UV coordinates. Models are packaged as single-file `.rvf` containers with progressive loading (Layer A instant, Layer B warm, Layer C full). SONA (Self-Optimizing Neural Architecture) enables continuous on-device adaptation via micro-LoRA + EWC++ without catastrophic forgetting.

| [Training & Fine-Tuning](#training--fine-tuning) | MM-Fi/Wi-Pose pre-training, 6-term composite loss, cosine-scheduled SGD, SONA LoRA | [ADR-023](docs/adr/ADR-023-trained-densepose-model-ruvector-pipeline.md) |

| [RuVector Crates](#ruvector-crates) | 11 vendored Rust crates: HNSW, attention, GNN, temporal compression, min-cut, solver | [Source](vendor/ruvector/) |

The RuVector Format (RVF) packages an entire trained model — weights, HNSW indexes, quantization codebooks, SONA adaptation deltas, and WASM inference runtime — into a single self-contained binary file. No external dependencies are needed at deployment time.

| **Quantization** | Per-segment temperature-tiered: f32 (full), f16 (half), u8 (int8), binary — with SIMD-accelerated distance computation |

| **Size** | Typical: 13 KB (sensing-only) to 50+ MB (full DensePose) |

See [ADR-023](docs/adr/ADR-023-trained-densepose-model-ruvector-pipeline.md).

The training pipeline implements 8 phases in pure Rust (7,832 lines, zero external ML dependencies). It trains a graph transformer with cross-attention to map CSI feature matrices to 17 COCO body keypoints and DensePose UV coordinates — following the approach from the CMU "DensePose From WiFi" paper (arXiv:2301.00250).

| Tier | Method | Purpose |

|------|--------|---------|

| **1. Pre-train** | MM-Fi + Wi-Pose public datasets | Cross-environment generalization (multi-subject, multi-room) |

| **2. Fine-tune** | ESP32 CSI + camera pseudo-labels | Environment-specific multipath adaptation |

| **3. SONA adapt** | Micro-LoRA (rank-4) + EWC++ | Continuous on-device learning without catastrophic forgetting |

| Phase | Module | What It Does |

|-------|--------|-------------|

| 1 | `dataset.rs` (850 lines) | MM-Fi `.npy` + Wi-Pose `.mat` loaders, subcarrier resampling (114→56, 30→56), windowing, normalization |

| 2 | `graph_transformer.rs` (855 lines) | COCO BodyGraph (17 kp, 16 edges), AntennaGraph, multi-head CrossAttention, GCN message passing |

| 3 | `trainer.rs` (881 lines) | 6-term composite loss (MSE, cross-entropy, UV regression, temporal consistency, bone length, symmetry), SGD+momentum, cosine+warmup scheduler, PCK/OKS metrics |

| 4 | `sona.rs` (639 lines) | LoRA adapters (A×B delta), EWC++ Fisher regularization, EnvironmentDetector (3-sigma drift detection) |

| 5 | `sparse_inference.rs` (753 lines) | NeuronProfiler hot/cold partitioning, SparseLinear (skip cold rows), INT8/FP16 quantization with <0.01 MSE |

| 6 | `rvf_pipeline.rs` (1,027 lines) | Progressive 3-layer loader, HNSW index, OverlayGraph, `RvfModelBuilder` |

| 7 | `rvf_container.rs` (914 lines) | Binary container format, 6+ segment types, CRC32 integrity |

| 8 | `main.rs` integration | `--train`, `--model`, `--progressive` CLI flags, REST endpoints |

**Best-epoch snapshotting**: the trainer saves the best validation loss weights and restores them before checkpoint/export — prevents exporting overfit final-epoch parameters.

See [ADR-023](docs/adr/ADR-023-trained-densepose-model-ruvector-pipeline.md).

<summary><a id="ruvector-crates"></a><strong>🔩 RuVector Crates</strong> — 11 vendored signal intelligence crates</summary>

| Crate | Purpose |

|-------|---------|

| `ruvector-core` | VectorDB, HNSW index, SIMD distance, quantization |

| `ruvector-attention` | Scaled dot-product, MoE, sparse attention |

| `ruvector-gnn` | Graph neural network, graph attention, EWC training |

| `ruvector-nervous-system` | PredictiveLayer, OscillatoryRouter, Hopfield |

| `ruvector-coherence` | Spectral coherence, HNSW health, Fiedler value |

| `ruvector-temporal-tensor` | Tiered temporal compression (8/7/5/3-bit) |

| `ruvector-mincut` | Subpolynomial dynamic min-cut |

| `ruvector-attn-mincut` | Attention-gated min-cut |

| `ruvector-solver` | Sparse Neumann solver O(sqrt(n)) |

| `ruvector-graph-transformer` | Proof-gated graph transformer |

| `ruvector-sparse-inference` | PowerInfer-style sparse execution |

See `vendor/ruvector/` for full source.