chore: update vendor submodules to latest main

Comprehensive QEMU emulation strategy for ESP32-S3 CSI node firmware:
- Layer 1: Mock CSI generator with 10 test scenarios
- Layer 2: QEMU runner + CI workflow with NVS matrix
- Layer 3: Multi-node mesh simulation (TAP networking)
- Layer 4: GDB remote debugging (zero-cost, no JTAG)
- Layer 5: Code coverage (gcov/lcov)
- Layer 6: Fuzz testing (libFuzzer for CSI parser, NVS, WASM)
- Layer 7: NVS provisioning matrix (14 configs)
- Layer 8: Snapshot & replay (<100ms restore)
- Layer 9: Chaos testing (9 fault injection scenarios)

Co-Authored-By: claude-flow <ruv@ruv.net>

README.md

@@ -75,7 +75,7 @@ docker run -p 3000:3000 ruvnet/wifi-densepose:latest
 |----------|-------------|
 | [User Guide](docs/user-guide.md) | Step-by-step guide: installation, first run, API usage, hardware setup, training |
 | [Build Guide](docs/build-guide.md) | Building from source (Rust and Python) |
-| [Architecture Decisions](docs/adr/README.md) | 48 ADRs — why each technical choice was made, organized by domain (hardware, signal processing, ML, platform, infrastructure) |
+| [Architecture Decisions](docs/adr/README.md) | 49 ADRs — why each technical choice was made, organized by domain (hardware, signal processing, ML, platform, infrastructure) |
 | [Domain Models](docs/ddd/README.md) | 7 DDD models (RuvSense, Signal Processing, Training Pipeline, Hardware Platform, Sensing Server, WiFi-Mat, CHCI) — bounded contexts, aggregates, domain events, and ubiquitous language |
 | [Desktop App](rust-port/wifi-densepose-rs/crates/wifi-densepose-desktop/README.md) | **WIP** — Tauri v2 desktop app for node management, OTA updates, WASM deployment, and mesh visualization |
 
@@ -87,10 +87,14 @@ docker run -p 3000:3000 ruvnet/wifi-densepose:latest
   </a>
   <br>
   <em>Real-time pose skeleton from WiFi CSI signals — no cameras, no wearables</em>
-  <br>
+  <br><br>
   <a href="https://ruvnet.github.io/RuView/"><strong>▶ Live Observatory Demo</strong></a>
+  &nbsp;|&nbsp;
+  <a href="https://ruvnet.github.io/RuView/pose-fusion.html"><strong>▶ Dual-Modal Pose Fusion Demo</strong></a>
 
 > The [server](#-quick-start) is optional for visualization and aggregation — the ESP32 [runs independently](#esp32-s3-hardware-pipeline) for presence detection, vital signs, and fall alerts.
+>
+> **Live ESP32 pipeline**: Connect an ESP32-S3 node → run the [sensing server](#sensing-server) → open the [pose fusion demo](https://ruvnet.github.io/RuView/pose-fusion.html) for real-time dual-modal pose estimation (webcam + WiFi CSI). See [ADR-059](docs/adr/ADR-059-live-esp32-csi-pipeline.md).
 
 
 ## 🚀 Key Features

docs/adr/ADR-058-ruvector-wasm-browser-pose-example.md

@@ -0,0 +1,392 @@
+# ADR-058: Dual-Modal WASM Browser Pose Estimation — Live Video + WiFi CSI Fusion
+
+- **Status**: Proposed
+- **Date**: 2026-03-12
+- **Deciders**: ruv
+- **Tags**: wasm, browser, cnn, pose-estimation, ruvector, video, multimodal, fusion
+
+## Context
+
+WiFi-DensePose estimates human poses from WiFi CSI (Channel State Information).
+The `ruvector-cnn` crate provides a pure Rust CNN (MobileNet-V3) with WASM bindings.
+Both modalities exist independently — what's missing is **fusing live webcam video
+with WiFi CSI** in a single browser demo to achieve robust pose estimation that
+works even when one modality degrades (occlusion, signal noise, poor lighting).
+
+Existing assets:
+
+1. **`wifi-densepose-wasm`** — CSI signal processing compiled to WASM
+2. **`wifi-densepose-sensing-server`** — Axum server streaming live CSI via WebSocket
+3. **`ruvector-cnn`** — Pure Rust CNN with MobileNet-V3 backbones, SIMD, contrastive learning
+4. **`ruvector-cnn-wasm`** — wasm-bindgen bindings: `WasmCnnEmbedder`, `SimdOps`, `LayerOps`, contrastive losses
+5. **`vendor/ruvector/examples/wasm-vanilla/`** — Reference vanilla JS WASM example
+
+Research shows multi-modal fusion (camera + WiFi) significantly outperforms either alone:
+- Camera fails under occlusion, poor lighting, privacy constraints
+- WiFi CSI fails with signal noise, multipath, low spatial resolution
+- Fusion compensates: WiFi provides through-wall coverage, camera provides fine-grained detail
+
+## Decision
+
+Build a **dual-modal browser demo** at `examples/wasm-browser-pose/` that:
+
+1. Captures **live webcam video** via `getUserMedia` API
+2. Receives **live WiFi CSI** via WebSocket from the sensing server
+3. Processes **both streams** through separate CNN pipelines in `ruvector-cnn-wasm`
+4. **Fuses embeddings** with learned attention weights for combined pose estimation
+5. Renders **video overlay** with skeleton + WiFi confidence heatmap on Canvas
+6. Runs entirely in the browser — all inference client-side via WASM
+
+### Architecture
+
+```
+┌──────────────────────────────────────────────────────────────────┐
+│  Browser                                                         │
+│                                                                  │
+│  ┌────────────┐    ┌────────────────┐    ┌───────────────────┐   │
+│  │ getUserMedia│───▶│ Video Frame    │───▶│ CNN WASM          │   │
+│  │ (Webcam)   │    │ Capture        │    │ (Visual Embedder) │   │
+│  └────────────┘    │ 224×224 RGB    │    │ → 512-dim         │   │
+│                    └────────────────┘    └────────┬──────────┘   │
+│                                                   │              │
+│                                          visual_embedding        │
+│                                                   │              │
+│                                            ┌──────▼──────┐       │
+│  ┌────────────┐    ┌────────────────┐      │             │       │
+│  │ WebSocket  │───▶│ CSI WASM       │      │  Attention  │       │
+│  │ Client     │    │ (densepose-    │      │  Fusion     │       │
+│  │            │    │  wasm)         │      │  Module     │       │
+│  └────────────┘    └───────┬────────┘      │             │       │
+│                            │               └──────┬──────┘       │
+│                    ┌───────▼────────┐             │              │
+│                    │ CNN WASM       │      fused_embedding       │
+│                    │ (CSI Embedder) │             │              │
+│                    │ → 512-dim      │      ┌──────▼──────┐       │
+│                    └───────┬────────┘      │ Pose        │       │
+│                            │               │ Decoder     │       │
+│                     csi_embedding           │ → 17 kpts   │       │
+│                            │               └──────┬──────┘       │
+│                            └──────────────────────┘              │
+│                                                   │              │
+│                    ┌──────────────┐         ┌─────▼──────┐       │
+│                    │ Video Canvas │◀────────│ Overlay    │       │
+│                    │ + Skeleton   │         │ Renderer   │       │
+│                    │ + Heatmap    │         └────────────┘       │
+│                    └──────────────┘                               │
+│                                                                  │
+└──────────────────────────────────────────────────────────────────┘
+         ▲                                     ▲
+         │ getUserMedia                        │ WebSocket
+         │ (camera)                            │ (ws://host:3030/ws/csi)
+         │                                     │
+    ┌────┴────┐                        ┌───────┴─────────┐
+    │ Webcam  │                        │ Sensing Server   │
+    └─────────┘                        └─────────────────┘
+```
+
+### Dual Pipeline Design
+
+Two parallel CNN pipelines run on each frame tick (~30 FPS):
+
+| Pipeline | Input | Preprocessing | CNN Config | Output |
+|----------|-------|---------------|------------|--------|
+| **Visual** | Webcam frame (640×480) | Resize to 224×224 RGB, ImageNet normalize | MobileNet-V3 Small, 512-dim | Visual embedding |
+| **CSI** | CSI frame (ADR-018 binary) | Amplitude/phase/delta → 224×224 pseudo-RGB | MobileNet-V3 Small, 512-dim | CSI embedding |
+
+Both use the same `WasmCnnEmbedder` but with separate instances and weight sets.
+
+### Fusion Strategy
+
+**Learned attention-weighted fusion** combines the two 512-dim embeddings:
+
+```javascript
+// Attention fusion: learn which modality to trust per-dimension
+// α ∈ [0,1]^512 — attention weights (shipped as JSON, trained offline)
+// visual_emb, csi_emb ∈ R^512
+
+function fuseEmbeddings(visual_emb, csi_emb, attention_weights) {
+    const fused = new Float32Array(512);
+    for (let i = 0; i < 512; i++) {
+        const α = attention_weights[i];
+        fused[i] = α * visual_emb[i] + (1 - α) * csi_emb[i];
+    }
+    return fused;
+}
+```
+
+**Dynamic confidence gating** adjusts fusion based on signal quality:
+
+| Condition | Behavior |
+|-----------|----------|
+| Good video + good CSI | Balanced fusion (α ≈ 0.5) |
+| Poor lighting / occlusion | CSI-dominant (α → 0, WiFi takes over) |
+| CSI noise / no ESP32 | Video-dominant (α → 1, camera only) |
+| Video-only mode (no WiFi) | α = 1.0, pure visual CNN pose estimation |
+| CSI-only mode (no camera) | α = 0.0, pure WiFi pose estimation |
+
+Quality detection:
+- **Video quality**: Frame brightness variance (dark = low quality), motion blur score
+- **CSI quality**: Signal-to-noise ratio from `wifi-densepose-wasm`, coherence gate output
+
+### CSI-to-Image Encoding
+
+CSI data encoded as 3-channel pseudo-image for the CSI CNN pipeline:
+
+| Channel | Data | Normalization |
+|---------|------|---------------|
+| R | CSI amplitude (subcarrier × time window) | Min-max to [0, 255] |
+| G | CSI phase (unwrapped, subcarrier × time window) | Min-max to [0, 255] |
+| B | Temporal difference (frame-to-frame Δ amplitude) | Abs, min-max to [0, 255] |
+
+### Video Processing
+
+Webcam frames processed through standard ImageNet pipeline:
+
+```javascript
+// Capture frame from video element
+const frame = captureVideoFrame(videoElement, 224, 224); // Returns Uint8Array RGB
+
+// ImageNet normalization happens inside WasmCnnEmbedder.extract()
+const visual_embedding = visual_embedder.extract(frame, 224, 224);
+```
+
+### Pose Keypoint Mapping
+
+17 COCO-format keypoints decoded from the fused 512-dim embedding:
+
+```
+ 0: nose          1: left_eye       2: right_eye
+ 3: left_ear      4: right_ear      5: left_shoulder
+ 6: right_shoulder 7: left_elbow    8: right_elbow
+ 9: left_wrist   10: right_wrist   11: left_hip
+12: right_hip    13: left_knee     14: right_knee
+15: left_ankle   16: right_ankle
+```
+
+Each keypoint decoded as (x, y, confidence) = 51 values from the 512-dim embedding
+via a learned linear projection.
+
+### Operating Modes
+
+The demo supports three modes, selectable in the UI:
+
+| Mode | Video | CSI | Fusion | Use Case |
+|------|-------|-----|--------|----------|
+| **Dual (default)** | ✅ | ✅ | Attention-weighted | Best accuracy, full demo |
+| **Video Only** | ✅ | ❌ | α = 1.0 | No ESP32 available, quick demo |
+| **CSI Only** | ❌ | ✅ | α = 0.0 | Privacy mode, through-wall sensing |
+
+**Video Only mode works without any hardware** — just a webcam — making the demo
+instantly accessible for anyone wanting to try it.
+
+### File Layout
+
+```
+examples/wasm-browser-pose/
+├── index.html                  # Single-page app (vanilla JS, no bundler)
+├── js/
+│   ├── app.js                  # Main entry, mode selection, orchestration
+│   ├── video-capture.js        # getUserMedia, frame extraction, quality detection
+│   ├── csi-processor.js        # WebSocket CSI client, frame parsing, pseudo-image encoding
+│   ├── fusion.js               # Attention-weighted embedding fusion, confidence gating
+│   ├── pose-decoder.js         # Fused embedding → 17 keypoints
+│   └── canvas-renderer.js      # Video overlay, skeleton, CSI heatmap, confidence bars
+├── data/
+│   ├── visual-weights.json     # Visual CNN → embedding projection (placeholder until trained)
+│   ├── csi-weights.json        # CSI CNN → embedding projection (placeholder until trained)
+│   ├── fusion-weights.json     # Attention fusion α weights (512 values)
+│   └── pose-weights.json       # Fused embedding → keypoint projection
+├── css/
+│   └── style.css               # Dark theme UI styling
+├── pkg/                        # Built WASM packages (gitignored, built by script)
+│   ├── wifi_densepose_wasm/
+│   └── ruvector_cnn_wasm/
+├── build.sh                    # wasm-pack build script for both packages
+└── README.md                   # Setup and usage instructions
+```
+
+### Build Pipeline
+
+```bash
+#!/bin/bash
+# build.sh — builds both WASM packages into pkg/
+
+set -e
+
+# Build wifi-densepose-wasm (CSI processing)
+wasm-pack build ../../rust-port/wifi-densepose-rs/crates/wifi-densepose-wasm \
+  --target web --out-dir "$(pwd)/pkg/wifi_densepose_wasm" --no-typescript
+
+# Build ruvector-cnn-wasm (CNN inference for both video and CSI)
+wasm-pack build ../../vendor/ruvector/crates/ruvector-cnn-wasm \
+  --target web --out-dir "$(pwd)/pkg/ruvector_cnn_wasm" --no-typescript
+
+echo "Build complete. Serve with: python3 -m http.server 8080"
+```
+
+### UI Layout
+
+```
+┌─────────────────────────────────────────────────────────┐
+│  WiFi-DensePose — Live Dual-Modal Pose Estimation       │
+│  [Dual Mode ▼]  [⚙ Settings]          FPS: 28  ◉ Live  │
+├───────────────────────────┬─────────────────────────────┤
+│                           │                             │
+│   ┌───────────────────┐   │   ┌───────────────────┐     │
+│   │                   │   │   │                   │     │
+│   │  Video + Skeleton │   │   │  CSI Heatmap      │     │
+│   │  Overlay          │   │   │  (amplitude ×     │     │
+│   │  (main canvas)    │   │   │   subcarrier)     │     │
+│   │                   │   │   │                   │     │
+│   └───────────────────┘   │   └───────────────────┘     │
+│                           │                             │
+├───────────────────────────┴─────────────────────────────┤
+│  Fusion Confidence: ████████░░ 78%                      │
+│  Video: ██████████ 95%  │  CSI: ██████░░░░ 61%          │
+├─────────────────────────────────────────────────────────┤
+│  ┌─────────────────────────────────────────────────┐    │
+│  │  Embedding Space (2D projection)                 │    │
+│  │     ·  ·    ·                                    │    │
+│  │   · · ·  ·    · ·    (color = pose cluster)     │    │
+│  │      ·  · · ·                                    │    │
+│  └─────────────────────────────────────────────────┘    │
+├─────────────────────────────────────────────────────────┤
+│  Latency: Video 12ms │ CSI 8ms │ Fusion 1ms │ Total 21ms│
+│  [▶ Record]  [📷 Snapshot]  [Confidence: ████ 0.6]      │
+└─────────────────────────────────────────────────────────┘
+```
+
+### WASM Module Structure
+
+| Package | Source Crate | Provides | Size (est.) |
+|---------|-------------|----------|-------------|
+| `wifi_densepose_wasm` | `wifi-densepose-wasm` | CSI frame parsing, signal processing, feature extraction | ~200KB |
+| `ruvector_cnn_wasm` | `ruvector-cnn-wasm` | `WasmCnnEmbedder` (×2 instances), `SimdOps`, `LayerOps`, contrastive losses | ~150KB |
+
+Two `WasmCnnEmbedder` instances are created — one for video frames, one for CSI pseudo-images.
+They share the same WASM module but have independent state.
+
+### Browser API Requirements
+
+| API | Purpose | Required | Fallback |
+|-----|---------|----------|----------|
+| `getUserMedia` | Webcam capture | For video mode | CSI-only mode |
+| WebAssembly | CNN inference | Yes | None (hard requirement) |
+| WASM SIMD128 | Accelerated inference | No | Scalar fallback (~2× slower) |
+| WebSocket | CSI data stream | For CSI mode | Video-only mode |
+| Canvas 2D | Rendering | Yes | None |
+| `requestAnimationFrame` | Render loop | Yes | `setTimeout` fallback |
+| ES Modules | Code organization | Yes | None |
+
+Target: Chrome 89+, Firefox 89+, Safari 15+, Edge 89+
+
+### Performance Budget
+
+| Stage | Target Latency | Notes |
+|-------|---------------|-------|
+| Video frame capture + resize | <3ms | `drawImage` to offscreen canvas |
+| Video CNN embedding | <15ms | 224×224 RGB → 512-dim |
+| CSI receive + parse | <2ms | Binary WebSocket message |
+| CSI pseudo-image encoding | <3ms | Amplitude/phase/delta channels |
+| CSI CNN embedding | <15ms | 224×224 pseudo-RGB → 512-dim |
+| Attention fusion | <1ms | Element-wise weighted sum |
+| Pose decoding | <1ms | Linear projection |
+| Canvas overlay render | <3ms | Video + skeleton + heatmap |
+| **Total (dual mode)** | **<33ms** | **30 FPS capable** |
+| **Total (video only)** | **<22ms** | **45 FPS capable** |
+
+Note: Video and CSI CNN pipelines can run in parallel using Web Workers,
+reducing dual-mode latency to ~max(15, 15) + 5 = ~20ms (50 FPS).
+
+### Contrastive Learning Integration
+
+The demo optionally shows real-time contrastive learning in the browser:
+
+- **InfoNCE loss** (`WasmInfoNCELoss`): Compare video vs CSI embeddings for the same pose — trains cross-modal alignment
+- **Triplet loss** (`WasmTripletLoss`): Push apart different poses, pull together same pose across modalities
+- **SimdOps**: Accelerated dot products for real-time similarity computation
+- **Embedding space panel**: Live 2D projection shows video and CSI embeddings converging when viewing the same person
+
+### Relationship to Existing Crates
+
+| Existing Crate | Role in This Demo |
+|---------------|-------------------|
+| `ruvector-cnn-wasm` | CNN inference for **both** video frames and CSI pseudo-images |
+| `wifi-densepose-wasm` | CSI frame parsing and signal processing |
+| `wifi-densepose-sensing-server` | WebSocket CSI data source |
+| `wifi-densepose-core` | ADR-018 frame format definitions |
+| `ruvector-cnn` | Underlying MobileNet-V3, layers, contrastive learning |
+
+No new Rust crates are needed. The example is pure HTML/JS consuming existing WASM packages.
+
+## Consequences
+
+### Positive
+
+- **Instant demo**: Video-only mode works with just a webcam — no ESP32 needed
+- **Multi-modal showcase**: Demonstrates camera + WiFi fusion, the core innovation of the project
+- **Graceful degradation**: Works with video-only, CSI-only, or both
+- **Through-wall capability**: CSI mode shows pose estimation where cameras cannot reach
+- **Zero-install**: Anyone with a browser can try it
+- **Training data collection**: Can record paired (video, CSI) data for offline model training
+- **Reusable**: JS modules embed directly in the Tauri desktop app's webview
+
+### Negative
+
+- **Model weights**: Requires offline-trained weights for visual CNN, CSI CNN, fusion, and pose decoder (~200KB total JSON)
+- **WASM size**: Two WASM modules total ~350KB (acceptable)
+- **No GPU**: CPU-only WASM inference; adequate at 224×224 but limits resolution scaling
+- **Camera privacy**: Video mode requires camera permission (mitigated: CSI-only mode available)
+- **Two CNN instances**: Memory footprint doubles vs single-modal (~10MB total, acceptable for desktop browsers)
+
+### Risks
+
+- **Cross-modal alignment**: Video and CSI embeddings must be trained jointly for fusion to work;
+  without proper training, fusion may be worse than either modality alone
+- **Latency on mobile**: Dual CNN on mobile browsers may exceed 33ms; implement automatic quality reduction
+- **WebSocket drops**: Network jitter → CSI frame gaps; buffer last 3 frames, interpolate missing data
+
+## Implementation Plan
+
+1. **Phase 1 — Scaffold**: File layout, build.sh, index.html shell, mode selector UI
+2. **Phase 2 — Video pipeline**: getUserMedia → frame capture → CNN embedding → basic pose display
+3. **Phase 3 — CSI pipeline**: WebSocket client → CSI parsing → pseudo-image → CNN embedding
+4. **Phase 4 — Fusion**: Attention-weighted combination, confidence gating, mode switching
+5. **Phase 5 — Pose decoder**: Linear projection with placeholder weights → 17 keypoints
+6. **Phase 6 — Overlay renderer**: Video canvas with skeleton overlay, CSI heatmap panel
+7. **Phase 7 — Training**: Use `wifi-densepose-train` to generate real weights for both CNNs + fusion + decoder
+8. **Phase 8 — Contrastive demo**: Embedding space visualization, cross-modal similarity display
+9. **Phase 9 — Web Workers**: Move CNN inference to workers for parallel video + CSI processing
+10. **Phase 10 — Polish**: Recording, snapshots, adaptive quality, mobile optimization
+
+## Alternatives Considered
+
+### 1. CSI-Only (No Video)
+Rejected: Misses the opportunity to show multi-modal fusion and makes the demo less
+accessible (requires ESP32 hardware). Video-only mode as a fallback is strictly better.
+
+### 2. Server-Side Video Inference
+Rejected: Adds latency, requires webcam stream upload (privacy concern), and defeats
+the WASM-first architecture. All inference must be client-side.
+
+### 3. TensorFlow.js for Video, ruvector-cnn-wasm for CSI
+Rejected: Would require two different ML frameworks. Using `ruvector-cnn-wasm` for both
+keeps a single WASM module, unified embedding space, and simpler fusion.
+
+### 4. Pre-recorded Video Demo
+Rejected: Live webcam input is far more compelling for demonstrations.
+Pre-recorded mode can be added as a secondary option.
+
+### 5. React/Vue Framework
+Rejected: Adds build tooling. Vanilla JS + ES modules keeps the demo self-contained.
+
+## References
+
+- [ADR-018: Binary CSI Frame Format](ADR-018-binary-csi-frame-format.md)
+- [ADR-024: Contrastive CSI Embedding / AETHER](ADR-024-contrastive-csi-embedding.md)
+- [ADR-055: Integrated Sensing Server](ADR-055-integrated-sensing-server.md)
+- `vendor/ruvector/crates/ruvector-cnn/src/lib.rs` — CNN embedder implementation
+- `vendor/ruvector/crates/ruvector-cnn-wasm/src/lib.rs` — WASM bindings
+- `vendor/ruvector/examples/wasm-vanilla/index.html` — Reference vanilla JS WASM pattern
+- Person-in-WiFi: Fine-grained Person Perception using WiFi (ICCV 2019) — camera+WiFi fusion precedent
+- WiPose: Multi-Person WiFi Pose Estimation (TMC 2022) — cross-modal embedding approach

docs/adr/ADR-059-live-esp32-csi-pipeline.md

@@ -0,0 +1,83 @@
+# ADR-059: Live ESP32 CSI Pipeline Integration
+
+## Status
+
+Accepted
+
+## Date
+
+2026-03-12
+
+## Context
+
+ADR-058 established a dual-modal browser demo combining webcam video and WiFi CSI for pose estimation. However, it used simulated CSI data. To demonstrate real-world capability, we need an end-to-end pipeline from physical ESP32 hardware through to the browser visualization.
+
+The ESP32-S3 firmware (`firmware/esp32-csi-node/`) already supports CSI collection and UDP streaming (ADR-018). The sensing server (`wifi-densepose-sensing-server`) already supports UDP ingestion and WebSocket bridging. The missing piece was connecting these components and enabling the browser demo to consume live data.
+
+## Decision
+
+Implement a complete live CSI pipeline:
+
+```
+ESP32-S3 (CSI capture) → UDP:5005 → sensing-server (Rust/Axum) → WS:8765 → browser demo
+```
+
+### Components
+
+1. **ESP32 Firmware** — Rebuilt with native Windows ESP-IDF v5.4.0 toolchain (no Docker). Configured for target network and PC IP via `sdkconfig`. Helper scripts added:
+   - `build_firmware.ps1` — Sets up IDF environment, cleans, builds, and flashes
+   - `read_serial.ps1` — Serial monitor with DTR/RTS reset capability
+
+2. **Sensing Server** — `wifi-densepose-sensing-server` started with:
+   - `--source esp32` — Expect real ESP32 UDP frames
+   - `--bind-addr 0.0.0.0` — Accept connections from any interface
+   - `--ui-path <path>` — Serve the demo UI via HTTP
+
+3. **Browser Demo** — `main.js` updated to auto-connect to `ws://localhost:8765/ws/sensing` on page load. Falls back to simulated CSI if the WebSocket is unavailable (GitHub Pages).
+
+### Network Configuration
+
+The ESP32 sends UDP packets to a configured target IP. If the PC's IP doesn't match the firmware's compiled target, a secondary IP alias can be added:
+
+```powershell
+# PowerShell (Admin)
+New-NetIPAddress -IPAddress 192.168.1.100 -PrefixLength 24 -InterfaceAlias "Wi-Fi"
+```
+
+### Data Flow
+
+| Stage | Protocol | Format | Rate |
+|-------|----------|--------|------|
+| ESP32 → Server | UDP | ADR-018 binary frame (magic `0xC5110001`, I/Q pairs) | ~100 Hz |
+| Server → Browser | WebSocket | ADR-018 binary frame (forwarded) | ~10 Hz (tick-ms=100) |
+| Browser decode | JavaScript | Float32 amplitude/phase arrays | Per frame |
+
+### Build Environment (Windows)
+
+ESP-IDF v5.4.0 on Windows requires:
+- IDF_PATH pointing to the ESP-IDF framework
+- IDF_TOOLS_PATH pointing to toolchain binaries
+- MSYS/MinGW environment variables removed (ESP-IDF rejects them)
+- Python venv from ESP-IDF tools for `idf.py` execution
+
+The `build_firmware.ps1` script handles all of this automatically.
+
+## Consequences
+
+### Positive
+- First end-to-end demonstration of real WiFi CSI → pose estimation in a browser
+- No Docker required for firmware builds on Windows
+- Demo gracefully degrades to simulated CSI when no server is available
+- Same demo works on GitHub Pages (simulated) and locally (live ESP32)
+
+### Negative
+- ESP32 target IP is compiled into firmware; changing it requires a rebuild or NVS override
+- Windows firewall may block UDP:5005; user must allow it
+- Mixed content restrictions prevent HTTPS pages from connecting to ws:// (local only)
+
+## Related
+
+- [ADR-018](ADR-018-esp32-dev-implementation.md) — ESP32 CSI frame format and UDP streaming
+- [ADR-058](ADR-058-ruvector-wasm-browser-pose-example.md) — Dual-modal WASM browser pose demo
+- [ADR-039](ADR-039-edge-intelligence-framework.md) — Edge intelligence on ESP32
+- Issue [#245](https://github.com/ruvnet/RuView/issues/245) — Tracking issue

docs/adr/ADR-060-provision-channel-mac-filter.md

@@ -0,0 +1,59 @@
+# ADR-060: Provision Channel Override and MAC Address Filtering
+
+- **Status:** Accepted
+- **Date:** 2026-03-12
+- **Issues:** [#247](https://github.com/ruvnet/RuView/issues/247), [#229](https://github.com/ruvnet/RuView/issues/229)
+
+## Context
+
+Two related provisioning gaps were reported by users:
+
+1. **Channel mismatch (Issue #247):** The CSI collector initializes on the
+   Kconfig default channel (typically 6), even when the ESP32 connects to an AP
+   on a different channel (e.g. 11). On managed networks where the user cannot
+   change the router channel, this makes nodes undiscoverable. The
+   `provision.py` script has no `--channel` argument.
+
+2. **Missing MAC filter (Issue #229):** The v0.2.0 release notes documented a
+   `--filter-mac` argument for `provision.py`, but it was never implemented.
+   The firmware's CSI callback accepts frames from all sources, causing signal
+   mixing in multi-AP environments.
+
+## Decision
+
+### Channel configuration
+
+- Add `--channel` argument to `provision.py` that writes a `csi_channel` key
+  (u8) to NVS.
+- In `nvs_config.c`, read the `csi_channel` key and override
+  `channel_list[0]` when present.
+- In `csi_collector_init()`, after WiFi connects, auto-detect the AP channel
+  via `esp_wifi_sta_get_ap_info()` and use it as the default CSI channel when
+  no NVS override is set. This ensures the CSI collector always matches the
+  connected AP's channel without requiring manual provisioning.
+
+### MAC address filtering
+
+- Add `--filter-mac` argument to `provision.py` that writes a `filter_mac`
+  key (6-byte blob) to NVS.
+- In `nvs_config.h`, add a `filter_mac[6]` field and `filter_mac_set` flag.
+- In `nvs_config.c`, read the `filter_mac` blob from NVS.
+- In the CSI callback (`wifi_csi_callback`), if `filter_mac_set` is true,
+  compare the source MAC from the received frame against the configured MAC
+  and drop non-matching frames.
+
+### Provisioning flow
+
+```
+python provision.py --port COM7 --channel 11
+python provision.py --port COM7 --filter-mac "AA:BB:CC:DD:EE:FF"
+python provision.py --port COM7 --channel 11 --filter-mac "AA:BB:CC:DD:EE:FF"
+```
+
+## Consequences
+
+- Users on managed networks can force the CSI channel to match their AP
+- Multi-AP environments can filter CSI to a single source
+- Auto-channel detection eliminates the most common misconfiguration
+- Backward compatible: existing provisioned nodes without these keys behave
+  as before (use Kconfig default channel, accept all MACs)

docs/adr/ADR-061-qemu-esp32s3-firmware-testing.md

@@ -0,0 +1,864 @@
+# ADR-061: QEMU ESP32-S3 Emulation for Firmware Testing & Development
+
+| Field       | Value                                          |
+|-------------|------------------------------------------------|
+| **Status**  | Proposed                                       |
+| **Date**    | 2026-03-13                                     |
+| **Authors** | RuView Team                                    |
+| **Relates** | ADR-018 (binary frame), ADR-039 (edge intel), ADR-040 (WASM), ADR-057 (build guard), ADR-060 (channel/MAC filter) |
+
+## Context
+
+The ESP32-S3 CSI node firmware (`firmware/esp32-csi-node/`) has grown to 16 source files spanning:
+
+| Module | File | Testable in QEMU? |
+|--------|------|--------------------|
+| NVS config load | `nvs_config.c` | Yes — NVS partition in flash image |
+| Edge processing (DSP) | `edge_processing.c` | Yes — all math, no HW dependency |
+| ADR-018 frame serialization | `csi_collector.c:csi_serialize_frame()` | Yes — pure buffer ops |
+| UDP stream sender | `stream_sender.c` | Yes — QEMU has lwIP via SLIRP |
+| WASM runtime | `wasm_runtime.c` | Yes — CPU only |
+| OTA update | `ota_update.c` | Partial — needs HTTP mock |
+| Power management | `power_mgmt.c` | Partial — no real light-sleep |
+| Display (OLED) | `display_*.c` | No — I2C hardware |
+| WiFi CSI callback | `csi_collector.c:wifi_csi_callback()` | **No** — requires RF PHY |
+| Channel hopping | `csi_collector.c:hop_timer_cb()` | **No** — requires `esp_wifi_set_channel()` |
+
+Currently, **every code change requires flashing to physical hardware** on COM7. This creates a bottleneck:
+- Build + flash cycle: ~20 seconds
+- Serial monitor: manual inspection
+- No automated CI (no ESP32-S3 in GitHub Actions runners)
+- Contributors without hardware cannot test firmware changes
+
+Espressif maintains an official QEMU fork (`github.com/espressif/qemu`) with ESP32-S3 machine support, including dual-core Xtensa LX7, flash mapping, UART, GPIO, timers, and FreeRTOS.
+
+## Decision
+
+Introduce a **comprehensive QEMU testing platform** for the ESP32-S3 CSI node firmware with nine capability layers:
+
+1. **Mock CSI generator** — compile-time synthetic CSI frame injection
+2. **QEMU runner** — automated build, run, and validation
+3. **Multi-node mesh simulation** — TDM and aggregation testing across QEMU instances
+4. **GDB remote debugging** — zero-cost breakpoint debugging without JTAG
+5. **Code coverage** — gcov/lcov integration for path analysis
+6. **Fuzz testing** — malformed input resilience for CSI parser, NVS, WASM
+7. **NVS provisioning matrix** — exhaustive config combination testing
+8. **Snapshot & replay** — sub-100ms state restore for fast iteration
+9. **Chaos testing** — fault injection for resilience validation
+
+---
+
+## Layer 1: Mock CSI Generator
+
+### Architecture
+
+```
+┌─────────────────────────────────────────────────────┐
+│                  ESP32-S3 Firmware                    │
+│                                                       │
+│  ┌─────────────┐    ┌──────────────────────────────┐ │
+│  │  Real WiFi   │    │  Mock CSI Generator          │ │
+│  │  CSI Callback │ OR │  (timer → synthetic frames)  │ │
+│  │  (HW only)   │    │  (QEMU + unit tests)         │ │
+│  └──────┬───────┘    └──────────┬───────────────────┘ │
+│         │                       │                     │
+│         └───────────┬───────────┘                     │
+│                     ▼                                 │
+│  ┌──────────────────────────────────────────────────┐ │
+│  │  edge_enqueue_csi() → SPSC ring → DSP Core 1    │ │
+│  │  ├── Biquad bandpass (breathing / heart rate)    │ │
+│  │  ├── Phase unwrapping + Welford stats            │ │
+│  │  ├── Top-K subcarrier selection                  │ │
+│  │  ├── Presence detection (adaptive threshold)     │ │
+│  │  ├── Fall detection (phase acceleration)         │ │
+│  │  └── Multi-person vitals clustering              │ │
+│  └──────────────────┬───────────────────────────────┘ │
+│                     ▼                                 │
+│  ┌──────────────────────────────────────────────────┐ │
+│  │  csi_serialize_frame() → ADR-018 binary format   │ │
+│  │  stream_sender_send() → UDP to aggregator        │ │
+│  │  edge vitals packet   → 0xC5110002 (32 bytes)    │ │
+│  └──────────────────────────────────────────────────┘ │
+└─────────────────────────────────────────────────────┘
+```
+
+### Mock CSI Generator Design
+
+When `CONFIG_CSI_MOCK_ENABLED=y` (Kconfig option), the build replaces `esp_wifi_set_csi_config()` / `esp_wifi_set_csi_rx_cb()` with a periodic timer that injects synthetic CSI frames:
+
+```c
+// mock_csi.c — synthetic CSI frame generator
+
+#define MOCK_CSI_INTERVAL_MS   50   // 20 Hz (matches real CSI rate)
+#define MOCK_N_SUBCARRIERS     52   // HT20 mode
+#define MOCK_IQ_LEN            (MOCK_N_SUBCARRIERS * 2)  // I + Q bytes
+
+typedef struct {
+    uint8_t  scenario;        // 0=empty, 1=person_static, 2=person_walking, 3=fall
+    uint32_t frame_count;
+    float    person_x;        // Simulated position [0..1]
+    float    person_speed;    // Movement speed per frame
+    uint8_t  breathing_phase; // Simulated breathing cycle
+} mock_state_t;
+
+// Generates realistic CSI I/Q data:
+// - Empty room: Gaussian noise + stable phase (low variance)
+// - Static person: Phase shift proportional to distance, breathing modulation
+// - Walking person: Progressive phase drift + Doppler-like amplitude change
+// - Fall event: Sudden phase acceleration spike
+void mock_generate_csi_frame(mock_state_t *state, wifi_csi_info_t *out_info);
+```
+
+### Signal Model
+
+The synthetic CSI generator models subcarrier amplitude and phase as:
+
+```
+A_k(t) = A_base + A_person * exp(-d_k²/σ²) + noise
+φ_k(t) = φ_base + (2π * d / λ) + breathing_mod(t) + noise
+
+where:
+  k         = subcarrier index
+  d_k       = simulated distance effect on subcarrier k
+  A_person  = amplitude perturbation from human body (scenario-dependent)
+  d         = simulated person-to-antenna distance
+  λ         = wavelength at subcarrier frequency
+  breathing_mod(t) = sin(2π * f_breath * t) * amplitude_breath
+  noise     = Gaussian, σ tuned to match real ESP32-S3 CSI noise floor (~-90 dBm)
+```
+
+This model exercises:
+- Presence detection (amplitude variance exceeds threshold)
+- Breathing rate extraction (periodic phase modulation at 0.1-0.5 Hz)
+- Fall detection (sudden phase acceleration exceeding `fall_thresh`)
+- Multi-person separation (distinct subcarrier groups with different breathing frequencies)
+
+### Scenarios
+
+| ID | Scenario | Duration | Expected Output |
+|----|----------|----------|-----------------|
+| 0 | Empty room | 10s | `presence=0`, `motion_energy < thresh` |
+| 1 | Static person | 10s | `presence=1`, `breathing_rate ∈ [10,25]`, `fall=0` |
+| 2 | Walking person | 10s | `presence=1`, `motion_energy > 0.5`, `fall=0` |
+| 3 | Fall event | 5s | `fall=1` flag set, `motion_energy` spike |
+| 4 | Multi-person | 15s | `n_persons=2`, independent breathing rates |
+| 5 | Channel sweep | 5s | Frames on channels 1, 6, 11 in sequence |
+| 6 | MAC filter test | 5s | Frames with wrong MAC are dropped (counter check) |
+| 7 | Ring buffer overflow | 3s | 1000 frames in 100ms burst, graceful drop |
+| 8 | Boundary RSSI | 5s | RSSI sweeps -127 to 0, no crash |
+| 9 | Zero-length frame | 2s | `iq_len=0` frames, serialize returns 0 |
+
+---
+
+## Layer 2: QEMU Runner & CI
+
+### QEMU Runner Script
+
+```bash
+#!/bin/bash
+# scripts/qemu-esp32s3-test.sh
+
+set -euo pipefail
+
+FIRMWARE_DIR="firmware/esp32-csi-node"
+BUILD_DIR="$FIRMWARE_DIR/build"
+QEMU_BIN="${QEMU_PATH:-qemu-system-xtensa}"
+FLASH_IMAGE="$BUILD_DIR/qemu_flash.bin"
+LOG_FILE="$BUILD_DIR/qemu_output.log"
+TIMEOUT_SEC="${QEMU_TIMEOUT:-60}"
+
+echo "=== QEMU ESP32-S3 Firmware Test ==="
+
+# 1. Build with mock CSI enabled
+echo "[1/4] Building firmware (mock CSI mode)..."
+idf.py -C "$FIRMWARE_DIR" \
+  -D SDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.qemu" \
+  build
+
+# 2. Merge binaries into single flash image
+echo "[2/4] Creating merged flash image..."
+esptool.py --chip esp32s3 merge_bin -o "$FLASH_IMAGE" \
+  --flash_mode dio --flash_freq 80m --flash_size 8MB \
+  0x0     "$BUILD_DIR/bootloader/bootloader.bin" \
+  0x8000  "$BUILD_DIR/partition_table/partition-table.bin" \
+  0xf000  "$BUILD_DIR/ota_data_initial.bin" \
+  0x20000 "$BUILD_DIR/esp32-csi-node.bin"
+
+# 3. Optionally inject pre-provisioned NVS partition
+if [ -f "$BUILD_DIR/nvs_test.bin" ]; then
+  echo "[2b] Injecting pre-provisioned NVS partition..."
+  dd if="$BUILD_DIR/nvs_test.bin" of="$FLASH_IMAGE" \
+    bs=1 seek=$((0x9000)) conv=notrunc
+fi
+
+# 4. Run in QEMU with timeout, capture UART output
+echo "[3/4] Running QEMU (timeout: ${TIMEOUT_SEC}s)..."
+timeout "$TIMEOUT_SEC" "$QEMU_BIN" \
+  -machine esp32s3 \
+  -nographic \
+  -drive file="$FLASH_IMAGE",if=mtd,format=raw \
+  -serial mon:stdio \
+  -no-reboot \
+  2>&1 | tee "$LOG_FILE" || true
+
+# 5. Validate expected output
+echo "[4/4] Validating output..."
+python3 scripts/validate_qemu_output.py "$LOG_FILE"
+```
+
+### QEMU sdkconfig overlay (`sdkconfig.qemu`)
+
+```
+# Enable mock CSI generator (disables real WiFi CSI)
+CONFIG_CSI_MOCK_ENABLED=y
+
+# Skip WiFi STA connection (no AP in QEMU)
+CONFIG_CSI_MOCK_SKIP_WIFI_CONNECT=y
+
+# Run all scenarios sequentially
+CONFIG_CSI_MOCK_SCENARIO=255
+
+# Use loopback for UDP (QEMU SLIRP provides 10.0.2.x network)
+CONFIG_CSI_TARGET_IP="10.0.2.2"
+
+# Shorter test durations
+CONFIG_CSI_MOCK_SCENARIO_DURATION_MS=5000
+
+# Enable verbose logging for validation
+CONFIG_LOG_DEFAULT_LEVEL_INFO=y
+CONFIG_CSI_MOCK_LOG_FRAMES=y
+```
+
+### Output Validation Script
+
+`scripts/validate_qemu_output.py` parses the UART log and checks:
+
+| Check | Pass Criteria | Severity |
+|-------|---------------|----------|
+| Boot | `app_main()` called, no panic/assert | FATAL |
+| NVS load | `nvs_config:` log line present | FATAL |
+| Mock CSI init | `mock_csi: Starting mock CSI generator` | FATAL |
+| Frame generation | `mock_csi: Generated N frames` where N > 0 | ERROR |
+| Edge pipeline | `edge_processing: DSP task started on Core 1` | ERROR |
+| Vitals output | At least one `vitals:` log line with valid BPM | ERROR |
+| Presence detection | `presence=1` appears during person scenarios | WARN |
+| Fall detection | `fall=1` appears during fall scenario | WARN |
+| MAC filter | `csi_collector: MAC filter dropped N frames` where N > 0 | WARN |
+| ADR-018 serialize | `csi_collector: Serialized N frames` where N > 0 | ERROR |
+| No crash | No `Guru Meditation Error`, no `assert failed`, no `abort()` | FATAL |
+| Clean exit | Firmware reaches end of scenario sequence | ERROR |
+| Heap OK | No `HEAP_ERROR` or `out of memory` | FATAL |
+| Stack OK | No `Stack overflow` detected | FATAL |
+
+Exit codes: `0` = all pass, `1` = WARN only, `2` = ERROR, `3` = FATAL
+
+### CI Workflow
+
+```yaml
+# .github/workflows/firmware-qemu.yml
+name: Firmware QEMU Tests
+on:
+  push:
+    paths: ['firmware/**']
+  pull_request:
+    paths: ['firmware/**']
+
+jobs:
+  qemu-test:
+    runs-on: ubuntu-latest
+    container:
+      image: espressif/idf:v5.4
+    strategy:
+      matrix:
+        scenario: [default, nvs-full, nvs-edge-tier0, nvs-tdm-3node]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Espressif QEMU
+        run: |
+          apt-get update && apt-get install -y libslirp-dev libglib2.0-dev ninja-build
+          git clone --depth 1 https://github.com/espressif/qemu.git /tmp/qemu
+          cd /tmp/qemu
+          ./configure --target-list=xtensa-softmmu --enable-slirp
+          make -j$(nproc)
+          cp build/qemu-system-xtensa /usr/local/bin/
+        env:
+          QEMU_PATH: /usr/local/bin/qemu-system-xtensa
+
+      - name: Prepare NVS for scenario
+        run: |
+          case "${{ matrix.scenario }}" in
+            nvs-full)
+              python firmware/esp32-csi-node/provision.py --dry-run \
+                --port dummy --ssid "TestWiFi" --password "test1234" \
+                --target-ip "10.0.2.2" --target-port 5005 \
+                --channel 6 --filter-mac AA:BB:CC:DD:EE:FF \
+                --node-id 1 --edge-tier 2
+              cp nvs_provision.bin firmware/esp32-csi-node/build/nvs_test.bin
+              ;;
+            nvs-edge-tier0)
+              python firmware/esp32-csi-node/provision.py --dry-run \
+                --port dummy --edge-tier 0 --node-id 5
+              cp nvs_provision.bin firmware/esp32-csi-node/build/nvs_test.bin
+              ;;
+            nvs-tdm-3node)
+              python firmware/esp32-csi-node/provision.py --dry-run \
+                --port dummy --tdm-slot 1 --tdm-total 3 --node-id 1
+              cp nvs_provision.bin firmware/esp32-csi-node/build/nvs_test.bin
+              ;;
+          esac
+
+      - name: Build firmware (mock CSI mode)
+        run: |
+          cd firmware/esp32-csi-node
+          idf.py -D SDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.qemu" set-target esp32s3
+          idf.py build
+
+      - name: Run QEMU tests
+        run: bash scripts/qemu-esp32s3-test.sh
+        env:
+          QEMU_PATH: /usr/local/bin/qemu-system-xtensa
+          QEMU_TIMEOUT: 90
+
+      - name: Upload QEMU log
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: qemu-output-${{ matrix.scenario }}
+          path: firmware/esp32-csi-node/build/qemu_output.log
+```
+
+---
+
+## Layer 3: Multi-Node Mesh Simulation
+
+Run multiple QEMU instances with TAP networking to test TDM slot coordination and multi-node aggregation.
+
+### Architecture
+
+```
+┌──────────┐   ┌──────────┐   ┌──────────┐
+│ QEMU #0  │   │ QEMU #1  │   │ QEMU #2  │
+│ slot=0   │   │ slot=1   │   │ slot=2   │
+│ node_id=0│   │ node_id=1│   │ node_id=2│
+└────┬─────┘   └────┬─────┘   └────┬─────┘
+     │              │              │
+     └──────────┬───┴──────────────┘
+                ▼
+        ┌───────────────┐
+        │ TAP bridge    │
+        │ (10.0.0.0/24) │
+        └───────┬───────┘
+                ▼
+        ┌───────────────┐
+        │ Rust aggregator│
+        │ (UDP :5005)   │
+        └───────────────┘
+```
+
+### Multi-Node Runner
+
+```bash
+#!/bin/bash
+# scripts/qemu-mesh-test.sh — run 3 QEMU nodes + Rust aggregator
+
+set -euo pipefail
+
+N_NODES=${1:-3}
+AGGREGATOR_PORT=5005
+BRIDGE="qemu-br0"
+
+# Create bridge
+ip link add "$BRIDGE" type bridge
+ip addr add 10.0.0.1/24 dev "$BRIDGE"
+ip link set "$BRIDGE" up
+
+# Build flash images with per-node NVS
+for i in $(seq 0 $((N_NODES - 1))); do
+  python firmware/esp32-csi-node/provision.py --dry-run \
+    --port dummy --node-id "$i" --tdm-slot "$i" --tdm-total "$N_NODES" \
+    --target-ip 10.0.0.1 --target-port "$AGGREGATOR_PORT"
+  cp nvs_provision.bin "build/nvs_node${i}.bin"
+
+  # Inject NVS into per-node flash image
+  cp build/qemu_flash.bin "build/qemu_flash_node${i}.bin"
+  dd if="build/nvs_node${i}.bin" of="build/qemu_flash_node${i}.bin" \
+    bs=1 seek=$((0x9000)) conv=notrunc
+done
+
+# Start Rust aggregator in background
+cargo run -p wifi-densepose-hardware --bin aggregator -- \
+  --listen 0.0.0.0:${AGGREGATOR_PORT} \
+  --expect-nodes "$N_NODES" \
+  --output build/mesh_test_results.json &
+AGGREGATOR_PID=$!
+
+# Launch QEMU nodes
+for i in $(seq 0 $((N_NODES - 1))); do
+  TAP="tap${i}"
+  ip tuntap add "$TAP" mode tap
+  ip link set "$TAP" master "$BRIDGE"
+  ip link set "$TAP" up
+
+  qemu-system-xtensa \
+    -machine esp32s3 \
+    -nographic \
+    -drive file="build/qemu_flash_node${i}.bin",if=mtd,format=raw \
+    -serial file:"build/qemu_node${i}.log" \
+    -nic tap,ifname="$TAP",script=no,downscript=no \
+    -no-reboot &
+  echo "Started QEMU node $i (PID: $!)"
+done
+
+# Wait for test duration
+sleep 30
+
+# Validate results
+kill $AGGREGATOR_PID 2>/dev/null || true
+python3 scripts/validate_mesh_test.py build/mesh_test_results.json --nodes "$N_NODES"
+```
+
+### Mesh Validation Checks
+
+| Check | Pass Criteria |
+|-------|---------------|
+| All nodes booted | N distinct `node_id` values in received frames |
+| TDM ordering | Slot 0 frames arrive before slot 1 within each TDM cycle |
+| No slot collision | No two frames from different nodes with overlapping timestamps within TDM window |
+| Frame count balance | Each node contributes ±10% of total frames |
+| ADR-018 compliance | All frames have valid magic `0xC5110001` and correct node IDs |
+| Vitals per node | Each node produces independent vitals packets |
+
+---
+
+## Layer 4: GDB Remote Debugging
+
+QEMU provides a built-in GDB stub for zero-cost debugging without JTAG hardware.
+
+### Usage
+
+```bash
+# Launch QEMU with GDB stub (paused at boot)
+qemu-system-xtensa \
+  -machine esp32s3 \
+  -nographic \
+  -drive file=build/qemu_flash.bin,if=mtd,format=raw \
+  -serial mon:stdio \
+  -s -S   # -s = GDB on :1234, -S = pause at start
+
+# In another terminal: attach GDB
+xtensa-esp-elf-gdb build/esp32-csi-node.elf \
+  -ex "target remote :1234" \
+  -ex "b edge_processing.c:dsp_task" \
+  -ex "b csi_collector.c:wifi_csi_callback" \
+  -ex "b mock_csi.c:mock_generate_csi_frame" \
+  -ex "watch g_nvs_config.csi_channel" \
+  -ex "continue"
+```
+
+### Key Breakpoint Locations
+
+| Breakpoint | Purpose |
+|-----------|---------|
+| `edge_processing.c:dsp_task` | DSP consumer loop entry |
+| `edge_processing.c:presence_detect` | Threshold comparison |
+| `edge_processing.c:fall_detect` | Phase acceleration check |
+| `csi_collector.c:wifi_csi_callback` | Frame ingestion (or mock injection point) |
+| `csi_collector.c:csi_serialize_frame` | ADR-018 serialization |
+| `nvs_config.c:nvs_config_load` | NVS parse logic |
+| `wasm_runtime.c:wasm_on_csi` | WASM module dispatch |
+| `mock_csi.c:mock_generate_csi_frame` | Synthetic frame generation |
+
+### VS Code Integration
+
+```json
+// .vscode/launch.json
+{
+  "version": "0.2.0",
+  "configurations": [{
+    "name": "QEMU ESP32-S3 Debug",
+    "type": "cppdbg",
+    "request": "launch",
+    "program": "${workspaceFolder}/firmware/esp32-csi-node/build/esp32-csi-node.elf",
+    "miDebuggerPath": "xtensa-esp-elf-gdb",
+    "miDebuggerServerAddress": "localhost:1234",
+    "setupCommands": [
+      { "text": "set remote hardware-breakpoint-limit 2" },
+      { "text": "set remote hardware-watchpoint-limit 2" }
+    ]
+  }]
+}
+```
+
+---
+
+## Layer 5: Code Coverage (gcov/lcov)
+
+### Build with Coverage
+
+```
+# sdkconfig.coverage (overlay)
+CONFIG_COMPILER_OPTIMIZATION_NONE=y
+CONFIG_GCOV_ENABLE=y
+CONFIG_APPTRACE_GCOV_ENABLE=y
+```
+
+### Coverage Collection
+
+```bash
+# After QEMU run, extract gcov data from flash dump
+esptool.py --chip esp32s3 read_flash 0x300000 0x100000 gcov_data.bin
+
+# Or use ESP-IDF's app_trace + gcov integration:
+# QEMU + GDB → "monitor gcov dump" → .gcda files
+
+# Generate HTML report
+lcov --capture --directory build --output-file coverage.info
+lcov --remove coverage.info '*/esp-idf/*' '*/test/*' --output-file coverage_filtered.info
+genhtml coverage_filtered.info --output-directory build/coverage_report
+```
+
+### Coverage Targets
+
+| Module | Target | Critical Paths |
+|--------|--------|---------------|
+| `edge_processing.c` | ≥80% | `dsp_task`, `biquad_filter`, `fall_detect`, `multi_person_cluster` |
+| `csi_collector.c` | ≥90% | `csi_serialize_frame`, `wifi_csi_callback`, MAC filter branch |
+| `nvs_config.c` | ≥95% | Every NVS key read path, default fallback paths |
+| `mock_csi.c` | ≥95% | All scenarios, all signal model branches |
+| `stream_sender.c` | ≥80% | Init, send, error paths |
+| `wasm_runtime.c` | ≥70% | Module load, dispatch, signature verify |
+
+---
+
+## Layer 6: Fuzz Testing
+
+### Fuzz Targets
+
+| Target | Input | Mutation Strategy | Looking For |
+|--------|-------|-------------------|-------------|
+| `csi_serialize_frame()` | Random `wifi_csi_info_t` | Extreme `len` (0, 65535), NULL `buf`, negative RSSI, channel 255 | Buffer overflow, NULL deref |
+| `nvs_config_load()` | Crafted NVS partition binary | Truncated strings, out-of-range u8/u16, missing keys, corrupt headers | Kconfig fallback, no crash |
+| `edge_enqueue_csi()` | Rapid-fire 10,000 frames | Vary `iq_len` (0 to `EDGE_MAX_IQ_BYTES+1`), randomize RSSI | Ring overflow, no data corruption |
+| `rvf_parser.c` | Malformed RVF network packets | Bad magic, truncated headers, oversized payloads | Parse rejection, no crash |
+| `wasm_upload.c` | Corrupt WASM blobs | Invalid magic, oversized modules, bad Ed25519 signatures, truncated | Rejection without crash, no code execution |
+| `csi_serialize_frame()` + `edge_enqueue_csi()` | Chained: generate → serialize → enqueue | End-to-end with random data | Pipeline integrity |
+
+### Implementation Approach
+
+```c
+// test/fuzz_csi_serialize.c — runs on host (not ESP32)
+// Compiled with: clang -fsanitize=fuzzer,address
+
+#include "csi_collector.h"
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+    if (size < sizeof(wifi_csi_info_t)) return 0;
+
+    wifi_csi_info_t info;
+    memcpy(&info, data, sizeof(info));
+
+    // Point buf at remaining fuzz data
+    size_t remaining = size - sizeof(info);
+    uint8_t iq_buf[2048];
+    if (remaining > sizeof(iq_buf)) remaining = sizeof(iq_buf);
+    memcpy(iq_buf, data + sizeof(info), remaining);
+    info.buf = iq_buf;
+    info.len = (int)remaining;
+
+    uint8_t out[4096];
+    csi_serialize_frame(&info, out, sizeof(out));
+    return 0;
+}
+```
+
+### Fuzz CI Job
+
+```yaml
+  fuzz-test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Build fuzz targets
+        run: |
+          cd firmware/esp32-csi-node/test
+          clang -fsanitize=fuzzer,address -I../main \
+            fuzz_csi_serialize.c ../main/csi_collector.c \
+            -o fuzz_serialize
+      - name: Run fuzz (5 min per target)
+        run: |
+          cd firmware/esp32-csi-node/test
+          timeout 300 ./fuzz_serialize corpus/ || true
+      - name: Upload crashes
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: fuzz-crashes
+          path: firmware/esp32-csi-node/test/crash-*
+```
+
+---
+
+## Layer 7: NVS Provisioning Matrix
+
+### Config Combinations
+
+| Config | NVS Values | Validates |
+|--------|-----------|-----------|
+| `default` | (empty NVS) | Kconfig fallback paths |
+| `wifi-only` | ssid, password | Basic provisioning |
+| `full-adr060` | channel=6, filter_mac=AA:BB:CC:DD:EE:FF | Channel override + MAC filter |
+| `edge-tier0` | edge_tier=0 | Raw CSI passthrough (no DSP) |
+| `edge-tier1` | edge_tier=1, pres_thresh=100, fall_thresh=2000 | Stats-only mode |
+| `edge-tier2-custom` | edge_tier=2, vital_win=128, vital_int=500, subk_count=16 | Full vitals with custom params |
+| `tdm-3node` | tdm_slot=1, tdm_nodes=3, node_id=1 | TDM mesh timing |
+| `wasm-signed` | wasm_max=4, wasm_verify=1, wasm_pubkey=<32 bytes> | WASM with Ed25519 verification |
+| `wasm-unsigned` | wasm_max=2, wasm_verify=0 | WASM without signature check |
+| `5ghz-channel` | channel=36, filter_mac=... | 5 GHz CSI collection |
+| `boundary-max` | target_port=65535, node_id=255, top_k=32, vital_win=256 | Max-range values |
+| `boundary-min` | target_port=1, node_id=0, top_k=1, vital_win=32 | Min-range values |
+| `power-save` | power_duty=10, edge_tier=0 | Low-power mode |
+| `corrupt-nvs` | (manually crafted partial/corrupt partition) | Graceful fallback to defaults |
+
+### Automated Matrix Generation
+
+```python
+# scripts/generate_nvs_matrix.py
+# Generates all 14 NVS partition binaries for CI matrix
+
+CONFIGS = [
+    {"name": "default", "args": []},
+    {"name": "wifi-only", "args": ["--ssid", "Test", "--password", "test1234"]},
+    {"name": "full-adr060", "args": ["--channel", "6", "--filter-mac", "AA:BB:CC:DD:EE:FF",
+                                      "--ssid", "Test", "--password", "test"]},
+    {"name": "edge-tier0", "args": ["--edge-tier", "0"]},
+    # ... all 14 configs
+]
+```
+
+---
+
+## Layer 8: Snapshot & Replay
+
+### QEMU Snapshot Commands
+
+```bash
+# Save snapshot after boot + NVS load (skip 3s boot time)
+(qemu) savevm post_boot
+
+# Save after WiFi connect + first CSI frame
+(qemu) savevm post_connect
+
+# Save after edge pipeline calibration complete (~60s)
+(qemu) savevm post_calibration
+
+# Restore any snapshot (< 100ms)
+(qemu) loadvm post_connect
+```
+
+### Automated Snapshot Pipeline
+
+```bash
+# scripts/qemu-snapshot-test.sh
+
+# Phase 1: Create base snapshots (one-time, cached in CI)
+qemu-system-xtensa ... -monitor unix:qemu.sock,server,nowait &
+sleep 5
+echo "savevm post_boot" | socat - UNIX-CONNECT:qemu.sock
+sleep 10
+echo "savevm post_first_frame" | socat - UNIX-CONNECT:qemu.sock
+
+# Phase 2: Run quick tests from snapshots (< 1s each)
+for test in test_presence test_fall test_multi_person; do
+  echo "loadvm post_first_frame" | socat - UNIX-CONNECT:qemu.sock
+  echo "cont" | socat - UNIX-CONNECT:qemu.sock
+  sleep 2  # Run test scenario
+  # Validate output
+done
+```
+
+### Performance Impact
+
+| Operation | Without Snapshots | With Snapshots |
+|-----------|-------------------|----------------|
+| Full boot + NVS + WiFi mock | ~5 seconds | ~5 seconds (first run) |
+| Run single scenario | ~5s boot + ~5s test = 10s | ~0.1s restore + ~5s test = 5.1s |
+| Run all 10 scenarios | ~100 seconds | ~51 seconds (49% faster) |
+| Run 14 NVS configs × 10 scenarios | ~23 minutes | ~12 minutes (48% faster) |
+
+---
+
+## Layer 9: Chaos Testing
+
+### Fault Injection Table
+
+| Fault | Injection Method | Expected Behavior | Severity |
+|-------|-----------------|-------------------|----------|
+| WiFi disconnect | Timer kills mock WiFi connection after N frames | Reconnect attempt, CSI pauses and resumes | HIGH |
+| Ring buffer overflow | Burst 1000 frames in 100ms | Frame drop counter increments, no crash, no data corruption | HIGH |
+| NVS corruption | Flash image with partial-write NVS partition | Falls back to Kconfig defaults, logs warning | MEDIUM |
+| Stack overflow | Deep recursion in WASM module callback | Watchdog fires, task restarts, no hang | HIGH |
+| Heap exhaustion | `malloc` returns NULL after N allocations | Graceful degradation, logs OOM, continues operation | HIGH |
+| Timer starvation | Block DSP task for 500ms | Frames dropped from ring, no deadlock, recovers | MEDIUM |
+| UDP send failure | SLIRP network down | `stream_sender_send` returns -1, error counter increments | LOW |
+| Corrupt CSI frame | Inject frame with invalid magic in I/Q data | Edge pipeline rejects, increments error counter | LOW |
+| NVS write during read | Concurrent NVS open for write while config loads | No corruption, NVS handle isolation | MEDIUM |
+
+### Chaos Runner
+
+```bash
+# scripts/qemu-chaos-test.sh
+
+# Run with fault injection enabled
+qemu-system-xtensa ... \
+  -monitor unix:qemu.sock,server,nowait &
+
+# Inject faults via GDB or monitor commands
+for fault in wifi_kill heap_exhaust ring_flood; do
+  echo "[CHAOS] Injecting: $fault"
+  python3 scripts/inject_fault.py --socket qemu.sock --fault "$fault"
+  sleep 5
+  python3 scripts/check_health.py --log "$LOG_FILE" --after-fault "$fault"
+done
+```
+
+---
+
+## Implementation Plan
+
+| Phase | Layer | Deliverables | Effort | Priority |
+|-------|-------|-------------|--------|----------|
+| **P1** | L1 + L2 | `mock_csi.c`, `mock_csi.h`, `Kconfig.projbuild`, `sdkconfig.qemu`, `qemu-esp32s3-test.sh`, `validate_qemu_output.py`, `firmware-qemu.yml` | 2 days | Critical |
+| **P2** | L4 + L5 | GDB launch config, `sdkconfig.coverage`, lcov integration, coverage CI job | 1 day | High |
+| **P3** | L7 | `generate_nvs_matrix.py`, 14 NVS configs, CI matrix expansion | 1 day | High |
+| **P4** | L6 | `fuzz_csi_serialize.c`, `fuzz_nvs_config.c`, `fuzz_edge_enqueue.c`, fuzz CI job | 2 days | High |
+| **P5** | L3 | `qemu-mesh-test.sh`, TAP bridge setup, `validate_mesh_test.py`, Rust aggregator integration | 3 days | High |
+| **P6** | L8 | Snapshot pipeline, cached base images in CI | 0.5 day | Medium |
+| **P7** | L9 | `inject_fault.py`, `check_health.py`, `qemu-chaos-test.sh`, 9 fault scenarios | 2 days | Medium |
+| **P8** | Performance | Instruction counting, DSP cycle profiling, optimization report | 1 day | Low |
+
+**Total**: ~12.5 days across 8 phases
+
+---
+
+## File Layout
+
+```
+firmware/esp32-csi-node/
+├── main/
+│   ├── mock_csi.c              # NEW — synthetic CSI frame generator
+│   ├── mock_csi.h              # NEW — mock API + scenario definitions
+│   ├── Kconfig.projbuild       # MODIFIED — CONFIG_CSI_MOCK_* options
+│   ├── CMakeLists.txt          # MODIFIED — conditional mock_csi.c inclusion
+│   └── ... (existing files unchanged)
+├── test/
+│   ├── fuzz_csi_serialize.c    # NEW — libFuzzer target for serialization
+│   ├── fuzz_nvs_config.c       # NEW — libFuzzer target for NVS parsing
+│   ├── fuzz_edge_enqueue.c     # NEW — libFuzzer target for ring buffer
+│   └── corpus/                 # NEW — seed inputs for fuzz targets
+├── sdkconfig.qemu             # NEW — QEMU-specific sdkconfig overlay
+├── sdkconfig.coverage         # NEW — gcov-enabled sdkconfig overlay
+└── ...
+
+scripts/
+├── qemu-esp32s3-test.sh       # NEW — single-node QEMU runner
+├── qemu-mesh-test.sh          # NEW — multi-node mesh runner
+├── qemu-chaos-test.sh         # NEW — chaos/fault injection runner
+├── validate_qemu_output.py    # NEW — UART log validation
+├── validate_mesh_test.py      # NEW — mesh test validation
+├── generate_nvs_matrix.py     # NEW — NVS config matrix generator
+├── inject_fault.py            # NEW — QEMU fault injection
+└── check_health.py            # NEW — post-fault health checker
+
+.vscode/
+└── launch.json                # MODIFIED — add QEMU GDB debug config
+
+.github/workflows/
+└── firmware-qemu.yml          # NEW — CI workflow with matrix
+```
+
+---
+
+## Consequences
+
+### Benefits
+
+1. **No hardware required** — contributors validate firmware changes with QEMU alone
+2. **Automated CI** — every PR touching `firmware/` runs 14 NVS configs × 10 scenarios in parallel
+3. **10× faster iteration** — snapshot restore in <100ms vs 20s flash cycle
+4. **Security hardening** — fuzz testing catches buffer overflows, NULL derefs, and parser bugs before they reach hardware
+5. **Mesh validation** — multi-node TDM tested without 3 physical ESP32s
+6. **Coverage visibility** — lcov reports show untested edge processing paths
+7. **Resilience proof** — chaos tests verify firmware recovers from WiFi drops, OOM, and ring overflow
+8. **GDB debugging** — set breakpoints on DSP pipeline without JTAG adapter
+9. **Regression detection** — boot failures, NVS parsing errors, and FreeRTOS deadlocks caught in CI
+
+### Limitations
+
+1. **No real WiFi/CSI** — QEMU cannot emulate the ESP32-S3 WiFi radio or CSI extraction hardware
+2. **Synthetic CSI fidelity** — mock frames approximate real CSI patterns but don't capture real-world multipath, interference, or antenna characteristics
+3. **Timing differences** — QEMU timing is not cycle-accurate; FreeRTOS tick rates may differ from hardware
+4. **No peripheral testing** — I2C display, real GPIO, and light-sleep power management cannot be tested
+5. **QEMU build requirement** — Espressif's QEMU fork must be built from source (not in Ubuntu packages)
+6. **Coverage overhead** — gcov-enabled builds are ~2× slower in QEMU
+
+### What QEMU Testing Covers vs Requires Hardware
+
+| Test Domain | QEMU | Hardware |
+|-------------|------|----------|
+| Boot + NVS config (14 configs) | Full | Full |
+| Edge DSP pipeline (biquad, Welford, top-K) | Full | Full |
+| ADR-018 frame serialization | Full | Full |
+| Vitals packet generation (0xC5110002) | Full | Full |
+| WASM module loading + execution | Full | Full |
+| Multi-node TDM mesh (3+ nodes) | Full (TAP) | Full |
+| Fuzz testing (CSI parser, NVS) | Full | N/A |
+| Code coverage analysis | Full | Partial |
+| GDB breakpoint debugging | Full | Full (JTAG) |
+| Chaos/fault injection | Full | Manual |
+| OTA update flow | Partial (HTTP mock) | Full |
+| Real WiFi connection | No | Full |
+| Real CSI data quality | No | Full |
+| Channel hopping on RF | No | Full |
+| MAC filter on real frames | No | Full |
+| Power management (light-sleep) | No | Full |
+| Display rendering (OLED) | No | Full |
+| UDP over real network | No | Full |
+
+---
+
+## Alternatives Considered
+
+### 1. Host-native unit tests (no QEMU)
+Extract pure C functions (`csi_serialize_frame`, edge DSP math) and compile/test on host with CMock/Unity. Simpler but doesn't test FreeRTOS integration, NVS, or boot sequence.
+
+**Verdict**: Complementary — do both. Host unit tests for math, QEMU for integration. Fuzz targets (Layer 6) already use host-native compilation.
+
+### 2. Hardware-in-the-loop CI (real ESP32 on runner)
+Use a self-hosted GitHub Actions runner with a physical ESP32-S3 attached.
+
+**Verdict**: Valuable but expensive and fragile. QEMU covers ~85% of test cases (up from 70% with all 9 layers). Add HIL later for real CSI validation only.
+
+### 3. Docker-based ESP-IDF build only (no runtime test)
+Just verify the firmware compiles in CI without running it.
+
+**Verdict**: Already possible but insufficient — compilation doesn't catch runtime bugs (stack overflow, NVS parsing errors, FreeRTOS deadlocks).
+
+### 4. Renode emulator
+Alternative to QEMU with better peripheral modeling for some platforms.
+
+**Verdict**: Renode has ESP32 support but ESP32-S3 support is less mature than Espressif's own QEMU fork. Revisit if Renode adds full S3 support.
+
+---
+
+## References
+
+- [Espressif QEMU fork](https://github.com/espressif/qemu) — official ESP32/S3/C3/H2 support
+- [ESP-IDF QEMU guide](https://docs.espressif.com/projects/esp-idf/en/latest/esp32s3/api-guides/tools/qemu.html)
+- [libFuzzer documentation](https://llvm.org/docs/LibFuzzer.html) — LLVM-based coverage-guided fuzzing
+- [lcov](https://github.com/linux-test-project/lcov) — Linux test coverage visualization
+- ADR-018: Binary CSI frame format (magic `0xC5110001`)
+- ADR-039: Edge intelligence pipeline (biquad, vitals, fall detection)
+- ADR-040: WASM programmable sensing runtime
+- ADR-057: Build-time CSI guard (`CONFIG_ESP_WIFI_CSI_ENABLED`)
+- ADR-060: Channel override and MAC address filter

firmware/esp32-csi-node/build_firmware.ps1

@@ -0,0 +1,31 @@
+# Remove MSYS environment variables that trigger ESP-IDF's MinGW rejection
+Remove-Item env:MSYSTEM -ErrorAction SilentlyContinue
+Remove-Item env:MSYSTEM_CARCH -ErrorAction SilentlyContinue
+Remove-Item env:MSYSTEM_CHOST -ErrorAction SilentlyContinue
+Remove-Item env:MSYSTEM_PREFIX -ErrorAction SilentlyContinue
+Remove-Item env:MINGW_CHOST -ErrorAction SilentlyContinue
+Remove-Item env:MINGW_PACKAGE_PREFIX -ErrorAction SilentlyContinue
+Remove-Item env:MINGW_PREFIX -ErrorAction SilentlyContinue
+
+$env:IDF_PATH = "C:\Users\ruv\esp\v5.4\esp-idf"
+$env:IDF_TOOLS_PATH = "C:\Espressif\tools"
+$env:IDF_PYTHON_ENV_PATH = "C:\Espressif\tools\python\v5.4\venv"
+$env:PATH = "C:\Espressif\tools\xtensa-esp-elf\esp-14.2.0_20241119\xtensa-esp-elf\bin;C:\Espressif\tools\cmake\3.30.2\cmake-3.30.2-windows-x86_64\bin;C:\Espressif\tools\ninja\1.12.1;C:\Espressif\tools\ccache\4.10.2\ccache-4.10.2-windows-x86_64;C:\Espressif\tools\idf-exe\1.0.3;C:\Espressif\tools\python\v5.4\venv\Scripts;$env:PATH"
+
+Set-Location "C:\Users\ruv\Projects\wifi-densepose\firmware\esp32-csi-node"
+
+$python = "$env:IDF_PYTHON_ENV_PATH\Scripts\python.exe"
+$idf = "$env:IDF_PATH\tools\idf.py"
+
+Write-Host "=== Cleaning stale build cache ==="
+& $python $idf fullclean
+
+Write-Host "=== Building firmware (SSID=ruv.net, target=192.168.1.20:5005) ==="
+& $python $idf build
+
+if ($LASTEXITCODE -eq 0) {
+    Write-Host "=== Build succeeded! Flashing to COM7 ==="
+    & $python $idf -p COM7 flash
+} else {
+    Write-Host "=== Build failed with exit code $LASTEXITCODE ==="
+}

firmware/esp32-csi-node/main/csi_collector.c

@@ -12,6 +12,7 @@
  */
 
 #include "csi_collector.h"
+#include "nvs_config.h"
 #include "stream_sender.h"
 #include "edge_processing.h"
 
@@ -21,6 +22,9 @@
 #include "esp_timer.h"
 #include "sdkconfig.h"
 
+/* ADR-060: Access the global NVS config for MAC filter and channel override. */
+extern nvs_config_t g_nvs_config;
+
 /* ADR-057: Build-time guard — fail early if CSI is not enabled in sdkconfig.
  * Without this, the firmware compiles but crashes at runtime with:
  *   "E (xxxx) wifi:CSI not enabled in menuconfig!"
@@ -151,6 +155,14 @@ size_t csi_serialize_frame(const wifi_csi_info_t *info, uint8_t *buf, size_t buf
 static void wifi_csi_callback(void *ctx, wifi_csi_info_t *info)
 {
     (void)ctx;
+
+    /* ADR-060: MAC address filtering — drop frames from non-matching sources. */
+    if (g_nvs_config.filter_mac_set) {
+        if (memcmp(info->mac, g_nvs_config.filter_mac, 6) != 0) {
+            return;  /* Source MAC doesn't match filter — skip frame. */
+        }
+    }
+
     s_cb_count++;
 
     if (s_cb_count <= 3 || (s_cb_count % 100) == 0) {
@@ -203,6 +215,29 @@ static void wifi_promiscuous_cb(void *buf, wifi_promiscuous_pkt_type_t type)
 
 void csi_collector_init(void)
 {
+    /* ADR-060: Determine the CSI channel.
+     * Priority: 1) NVS override (--channel), 2) connected AP channel, 3) Kconfig default. */
+    uint8_t csi_channel = (uint8_t)CONFIG_CSI_WIFI_CHANNEL;
+
+    if (g_nvs_config.csi_channel > 0) {
+        /* Explicit NVS override via provision.py --channel */
+        csi_channel = g_nvs_config.csi_channel;
+        ESP_LOGI(TAG, "Using NVS channel override: %u", (unsigned)csi_channel);
+    } else {
+        /* Auto-detect from connected AP */
+        wifi_ap_record_t ap_info;
+        if (esp_wifi_sta_get_ap_info(&ap_info) == ESP_OK && ap_info.primary > 0) {
+            csi_channel = ap_info.primary;
+            ESP_LOGI(TAG, "Auto-detected AP channel: %u", (unsigned)csi_channel);
+        } else {
+            ESP_LOGW(TAG, "Could not detect AP channel, using Kconfig default: %u",
+                     (unsigned)csi_channel);
+        }
+    }
+
+    /* Update the hop table's first channel to match. */
+    s_hop_channels[0] = csi_channel;
+
     /* Enable promiscuous mode — required for reliable CSI callbacks.
      * Without this, CSI only fires on frames destined to this station,
      * which may be very infrequent on a quiet network. */
@@ -230,8 +265,15 @@ void csi_collector_init(void)
     ESP_ERROR_CHECK(esp_wifi_set_csi_rx_cb(wifi_csi_callback, NULL));
     ESP_ERROR_CHECK(esp_wifi_set_csi(true));
 
-    ESP_LOGI(TAG, "CSI collection initialized (node_id=%d, channel=%d)",
-             CONFIG_CSI_NODE_ID, CONFIG_CSI_WIFI_CHANNEL);
+    if (g_nvs_config.filter_mac_set) {
+        ESP_LOGI(TAG, "MAC filter active: %02x:%02x:%02x:%02x:%02x:%02x",
+                 g_nvs_config.filter_mac[0], g_nvs_config.filter_mac[1],
+                 g_nvs_config.filter_mac[2], g_nvs_config.filter_mac[3],
+                 g_nvs_config.filter_mac[4], g_nvs_config.filter_mac[5]);
+    }
+
+    ESP_LOGI(TAG, "CSI collection initialized (node_id=%d, channel=%u)",
+             CONFIG_CSI_NODE_ID, (unsigned)csi_channel);
 }
 
 /* ---- ADR-029: Channel hopping ---- */

firmware/esp32-csi-node/main/nvs_config.c

@@ -91,6 +91,11 @@ void nvs_config_load(nvs_config_t *cfg)
     cfg->wasm_verify = 0;  /* Kconfig disabled signature verification. */
 #endif
 
+    /* ADR-060: Channel override and MAC filter defaults. */
+    cfg->csi_channel = 0;  /* 0 = auto-detect from connected AP. */
+    cfg->filter_mac_set = 0;
+    memset(cfg->filter_mac, 0, 6);
+
     /* Try to override from NVS */
     nvs_handle_t handle;
     esp_err_t err = nvs_open("csi_cfg", NVS_READONLY, &handle);
@@ -277,6 +282,26 @@ void nvs_config_load(nvs_config_t *cfg)
         ESP_LOGW(TAG, "wasm_verify=1 but no wasm_pubkey in NVS — uploads will be rejected");
     }
 
+    /* ADR-060: CSI channel override. */
+    uint8_t csi_ch_val;
+    if (nvs_get_u8(handle, "csi_channel", &csi_ch_val) == ESP_OK) {
+        if ((csi_ch_val >= 1 && csi_ch_val <= 14) || (csi_ch_val >= 36 && csi_ch_val <= 177)) {
+            cfg->csi_channel = csi_ch_val;
+            ESP_LOGI(TAG, "NVS override: csi_channel=%u", (unsigned)cfg->csi_channel);
+        } else {
+            ESP_LOGW(TAG, "NVS csi_channel=%u invalid, ignored", (unsigned)csi_ch_val);
+        }
+    }
+
+    /* ADR-060: MAC address filter (6-byte blob). */
+    size_t mac_len = 6;
+    if (nvs_get_blob(handle, "filter_mac", cfg->filter_mac, &mac_len) == ESP_OK && mac_len == 6) {
+        cfg->filter_mac_set = 1;
+        ESP_LOGI(TAG, "NVS override: filter_mac=%02x:%02x:%02x:%02x:%02x:%02x",
+                 cfg->filter_mac[0], cfg->filter_mac[1], cfg->filter_mac[2],
+                 cfg->filter_mac[3], cfg->filter_mac[4], cfg->filter_mac[5]);
+    }
+
     /* Validate tdm_slot_index < tdm_node_count */
     if (cfg->tdm_slot_index >= cfg->tdm_node_count) {
         ESP_LOGW(TAG, "tdm_slot_index=%u >= tdm_node_count=%u, clamping to 0",

firmware/esp32-csi-node/main/nvs_config.h

@@ -50,6 +50,11 @@ typedef struct {
     uint8_t  wasm_verify;                    /**< Require Ed25519 signature for uploads. */
     uint8_t  wasm_pubkey[32];               /**< Ed25519 public key for WASM signature. */
     uint8_t  wasm_pubkey_valid;             /**< 1 if pubkey was loaded from NVS. */
+
+    /* ADR-060: Channel override and MAC address filtering */
+    uint8_t  csi_channel;                    /**< Explicit CSI channel override (0 = auto-detect). */
+    uint8_t  filter_mac[6];                  /**< MAC address to filter CSI frames. */
+    uint8_t  filter_mac_set;                 /**< 1 if filter_mac was loaded from NVS. */
 } nvs_config_t;
 
 /**

firmware/esp32-csi-node/provision.py

@@ -64,6 +64,13 @@ def build_nvs_csv(args):
         writer.writerow(["vital_int", "data", "u16", str(args.vital_int)])
     if args.subk_count is not None:
         writer.writerow(["subk_count", "data", "u8", str(args.subk_count)])
+    # ADR-060: Channel override and MAC filter
+    if args.channel is not None:
+        writer.writerow(["csi_channel", "data", "u8", str(args.channel)])
+    if args.filter_mac is not None:
+        mac_bytes = bytes(int(b, 16) for b in args.filter_mac.split(":"))
+        # NVS blob: write as hex-encoded string for CSV compatibility
+        writer.writerow(["filter_mac", "data", "hex2bin", mac_bytes.hex()])
     return buf.getvalue()
 
 
@@ -165,6 +172,10 @@ def main():
     parser.add_argument("--vital-win", type=int, help="Phase history window in frames (default: 300)")
     parser.add_argument("--vital-int", type=int, help="Vitals packet interval in ms (default: 1000)")
     parser.add_argument("--subk-count", type=int, help="Top-K subcarrier count (default: 32)")
+    # ADR-060: Channel override and MAC filter
+    parser.add_argument("--channel", type=int, help="CSI channel (1-14 for 2.4GHz, 36-177 for 5GHz). "
+                        "Overrides auto-detection from connected AP.")
+    parser.add_argument("--filter-mac", type=str, help="MAC address to filter CSI frames (AA:BB:CC:DD:EE:FF)")
     parser.add_argument("--dry-run", action="store_true", help="Generate NVS binary but don't flash")
 
     args = parser.parse_args()
@@ -176,6 +187,7 @@ def main():
         args.edge_tier is not None, args.pres_thresh is not None,
         args.fall_thresh is not None, args.vital_win is not None,
         args.vital_int is not None, args.subk_count is not None,
+        args.channel is not None, args.filter_mac is not None,
     ])
     if not has_value:
         parser.error("At least one config value must be specified")
@@ -186,6 +198,22 @@ def main():
     if args.tdm_slot is not None and args.tdm_slot >= args.tdm_total:
         parser.error(f"--tdm-slot ({args.tdm_slot}) must be less than --tdm-total ({args.tdm_total})")
 
+    # ADR-060: Validate channel and MAC filter
+    if args.channel is not None:
+        if not ((1 <= args.channel <= 14) or (36 <= args.channel <= 177)):
+            parser.error(f"--channel must be 1-14 (2.4GHz) or 36-177 (5GHz), got {args.channel}")
+    if args.filter_mac is not None:
+        parts = args.filter_mac.split(":")
+        if len(parts) != 6:
+            parser.error(f"--filter-mac must be in AA:BB:CC:DD:EE:FF format, got '{args.filter_mac}'")
+        try:
+            for p in parts:
+                val = int(p, 16)
+                if val < 0 or val > 255:
+                    raise ValueError
+        except ValueError:
+            parser.error(f"--filter-mac contains invalid hex bytes: '{args.filter_mac}'")
+
     print("Building NVS configuration:")
     if args.ssid:
         print(f"  WiFi SSID:     {args.ssid}")
@@ -212,6 +240,10 @@ def main():
         print(f"  Vital Interval:{args.vital_int} ms")
     if args.subk_count is not None:
         print(f"  Top-K Subcarr: {args.subk_count}")
+    if args.channel is not None:
+        print(f"  CSI Channel:   {args.channel}")
+    if args.filter_mac is not None:
+        print(f"  Filter MAC:    {args.filter_mac}")
 
     csv_content = build_nvs_csv(args)
 

firmware/esp32-csi-node/read_serial.ps1

@@ -0,0 +1,14 @@
+$p = New-Object System.IO.Ports.SerialPort('COM7', 115200)
+$p.ReadTimeout = 5000
+$p.Open()
+Start-Sleep -Milliseconds 200
+
+for ($i = 0; $i -lt 60; $i++) {
+    try {
+        $line = $p.ReadLine()
+        Write-Host $line
+    } catch {
+        break
+    }
+}
+$p.Close()

ui/index.html

@@ -29,6 +29,7 @@
             <button class="nav-tab" data-tab="applications">Applications</button>
             <button class="nav-tab" data-tab="sensing">Sensing</button>
             <button class="nav-tab" data-tab="training">Training</button>
+            <a href="pose-fusion.html" class="nav-tab" style="text-decoration:none">Pose Fusion</a>
             <a href="observatory.html" class="nav-tab" style="text-decoration:none">Observatory</a>
         </nav>
 

ui/pose-fusion.html

@@ -0,0 +1,201 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1.0">
+  <title>RuView — Dual-Modal Pose Estimation</title>
+  <link rel="stylesheet" href="pose-fusion/css/style.css?v=13">
+</head>
+<body>
+
+  <!-- Header -->
+  <header class="header">
+    <div class="header-left">
+      <div class="logo"><span class="pi">&pi;</span> RuView</div>
+      <div class="header-title">Dual-Modal Pose Estimation — Live Video + WiFi CSI Fusion</div>
+    </div>
+    <div class="header-right">
+      <select id="mode-select" class="mode-select">
+        <option value="dual">Dual Mode (Video + CSI)</option>
+        <option value="video">Video Only</option>
+        <option value="csi">CSI Only (WiFi)</option>
+      </select>
+      <div class="status-badge">
+        <span id="status-dot" class="status-dot offline"></span>
+        <span id="status-label">READY</span>
+      </div>
+      <span id="fps-display" class="fps-badge">-- FPS</span>
+      <a href="index.html" class="back-link">&larr; Dashboard</a>
+      <a href="observatory.html" class="back-link">Observatory &rarr;</a>
+    </div>
+  </header>
+
+  <!-- Main Grid -->
+  <div class="main-grid">
+
+    <!-- Video + Skeleton Panel -->
+    <div class="video-panel">
+      <video id="webcam" autoplay playsinline muted></video>
+      <canvas id="skeleton-canvas"></canvas>
+      <div class="video-overlay-label" id="mode-label">DUAL FUSION</div>
+
+      <div id="camera-prompt" class="camera-prompt">
+        <div class="camera-prompt-label" id="prompt-mode-label">DUAL FUSION</div>
+        <p>Enable your webcam for live video pose estimation.<br>
+           Or switch to <strong>CSI Only</strong> mode for WiFi-based sensing.</p>
+        <button id="start-camera-btn">Enable Camera</button>
+      </div>
+    </div>
+
+    <!-- Side Panels -->
+    <div class="side-panels">
+
+      <!-- Fusion Confidence -->
+      <div class="panel">
+        <div class="panel-title">&#9670; Fusion Confidence</div>
+        <div class="fusion-bars">
+          <div class="bar-row">
+            <span class="bar-label">Video</span>
+            <div class="bar-track"><div class="bar-fill video" id="video-bar" style="width:0%"></div></div>
+            <span class="bar-value" id="video-bar-val">0%</span>
+          </div>
+          <div class="bar-row">
+            <span class="bar-label">CSI</span>
+            <div class="bar-track"><div class="bar-fill csi" id="csi-bar" style="width:0%"></div></div>
+            <span class="bar-value" id="csi-bar-val">0%</span>
+          </div>
+          <div class="bar-row">
+            <span class="bar-label">Fused</span>
+            <div class="bar-track"><div class="bar-fill fused" id="fused-bar" style="width:0%"></div></div>
+            <span class="bar-value" id="fused-bar-val">0%</span>
+          </div>
+        </div>
+        <div style="margin-top:8px; font-size:10px; color:var(--text-label)">
+          Cross-modal: <span id="cross-modal-sim" style="color:var(--green-glow)">0.000</span>
+        </div>
+      </div>
+
+      <!-- CSI Heatmap -->
+      <div class="panel">
+        <div class="panel-title">&#9670; CSI Amplitude Heatmap</div>
+        <div class="csi-canvas-wrapper">
+          <canvas id="csi-canvas" width="320" height="100"></canvas>
+        </div>
+      </div>
+
+      <!-- RSSI Signal Strength -->
+      <div class="panel">
+        <div class="panel-title">&#9670; RSSI Signal Strength</div>
+        <div class="rssi-row">
+          <div class="rssi-gauge">
+            <div class="rssi-bar-track">
+              <div class="rssi-bar-fill" id="rssi-bar" style="width:0%"></div>
+            </div>
+            <div class="rssi-values">
+              <span class="rssi-dbm" id="rssi-value">-- dBm</span>
+              <span class="rssi-quality" id="rssi-quality">--</span>
+            </div>
+          </div>
+          <canvas id="rssi-sparkline" width="160" height="32"></canvas>
+        </div>
+      </div>
+
+      <!-- Embedding Space -->
+      <div class="panel">
+        <div class="panel-title">&#9670; Embedding Space (2D Projection)</div>
+        <div class="embedding-canvas-wrapper">
+          <canvas id="embedding-canvas" width="320" height="100"></canvas>
+        </div>
+      </div>
+
+      <!-- RuVector Attention Pipeline -->
+      <div class="panel">
+        <div class="panel-title">&#9670; RuVector WASM Attention Pipeline</div>
+        <div class="rv-pipeline">
+          <div class="rv-stage" id="rv-flash">Flash</div>
+          <div class="rv-arrow">&rarr;</div>
+          <div class="rv-stage" id="rv-mha">MHA</div>
+          <div class="rv-arrow">&rarr;</div>
+          <div class="rv-stage" id="rv-hyp">Hyper</div>
+          <div class="rv-arrow">&rarr;</div>
+          <div class="rv-stage" id="rv-lin">Linear</div>
+          <div class="rv-arrow">&rarr;</div>
+          <div class="rv-stage" id="rv-moe">MoE</div>
+          <div class="rv-arrow">&rarr;</div>
+          <div class="rv-stage" id="rv-lg">L+G</div>
+        </div>
+        <div class="rv-stats">
+          <span>Energy: <span id="rv-energy" style="color:var(--green-glow)">--</span></span>
+          <span>Refinement: <span id="rv-refine" style="color:var(--cyan)">--</span></span>
+          <span>Pose Impact: <span id="rv-impact" style="color:var(--amber)">--</span></span>
+        </div>
+      </div>
+
+      <!-- Latency -->
+      <div class="panel">
+        <div class="panel-title">&#9670; Pipeline Latency</div>
+        <div class="latency-grid">
+          <div class="latency-item">
+            <div class="latency-value" id="lat-video">--</div>
+            <div class="latency-label">Video CNN</div>
+          </div>
+          <div class="latency-item">
+            <div class="latency-value" id="lat-csi">--</div>
+            <div class="latency-label">CSI CNN</div>
+          </div>
+          <div class="latency-item">
+            <div class="latency-value" id="lat-fusion">--</div>
+            <div class="latency-label">Fusion</div>
+          </div>
+          <div class="latency-item">
+            <div class="latency-value" id="lat-total">--</div>
+            <div class="latency-label">Total</div>
+          </div>
+        </div>
+      </div>
+
+      <!-- Controls -->
+      <div class="panel">
+        <div class="panel-title">&#9670; Controls</div>
+        <div class="controls-row">
+          <button class="btn" id="pause-btn">⏸ Pause</button>
+        </div>
+
+        <div class="slider-row">
+          <label>Confidence</label>
+          <input type="range" id="confidence-slider" min="0" max="1" step="0.05" value="0.3">
+          <span class="slider-val" id="confidence-value">0.30</span>
+        </div>
+
+        <div style="margin-top:10px">
+          <div class="panel-title" style="margin-bottom:6px">&#9670; Live CSI Source</div>
+          <div style="display:flex;gap:6px">
+            <input type="text" id="ws-url" placeholder="ws://localhost:3030/ws/csi"
+              style="flex:1;background:rgba(255,255,255,0.05);border:1px solid var(--bg-panel-border);
+                     color:var(--text-primary);padding:5px 8px;border-radius:4px;font-size:11px;
+                     font-family:'JetBrains Mono',monospace">
+            <button class="btn" id="connect-ws-btn">Connect</button>
+          </div>
+        </div>
+      </div>
+
+    </div><!-- /side-panels -->
+
+    <!-- Bottom Bar -->
+    <div class="bottom-bar">
+      <div>
+        RuView &middot; Dual-Modal Pose Estimation &middot;
+        Architecture: Conv2D &rarr; RuVector 6-Stage Attention (Flash+MHA+Hyperbolic+Linear+MoE+L/G) &rarr; Fusion &rarr; 26-Keypoint Pose
+      </div>
+      <div>
+        <a href="https://github.com/ruvnet/RuView">GitHub</a> &middot;
+        CNN: <span id="cnn-backend">ruvector-cnn (loading…)</span> &middot;
+        <a href="observatory.html">Observatory</a>
+      </div>
+    </div>
+
+  </div><!-- /main-grid -->
+
+  <script type="module" src="pose-fusion/js/main.js?v=13"></script>
+</body>
+</html>

ui/pose-fusion/build.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+# Build WASM packages for the dual-modal pose estimation demo.
+# Requires: wasm-pack (cargo install wasm-pack)
+#
+# Usage: ./build.sh
+#
+# Output: pkg/ruvector_cnn_wasm/ — WASM CNN embedder for browser
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+VENDOR_DIR="$SCRIPT_DIR/../../vendor/ruvector"
+OUT_DIR="$SCRIPT_DIR/pkg/ruvector_cnn_wasm"
+
+echo "Building ruvector-cnn-wasm..."
+wasm-pack build "$VENDOR_DIR/crates/ruvector-cnn-wasm" \
+  --target web \
+  --out-dir "$OUT_DIR" \
+  --no-typescript
+
+# Remove .gitignore so we can commit the build output for GitHub Pages
+rm -f "$OUT_DIR/.gitignore"
+
+echo ""
+echo "Build complete!"
+echo "  WASM: $(du -sh "$OUT_DIR/ruvector_cnn_wasm_bg.wasm" | cut -f1)"
+echo "  JS:   $(du -sh "$OUT_DIR/ruvector_cnn_wasm.js" | cut -f1)"
+echo ""
+echo "Serve the demo: cd $SCRIPT_DIR/.. && python3 -m http.server 8080"
+echo "Open: http://localhost:8080/pose-fusion.html"

ui/pose-fusion/css/style.css

@@ -0,0 +1,536 @@
+/* RuView — Dual-Modal Pose Fusion Demo
+   Dark theme matching Observatory */
+
+@import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;600;700&family=JetBrains+Mono:wght@400;600&display=swap');
+
+:root {
+  --bg-deep:     #080c14;
+  --bg-panel:    rgba(8, 16, 28, 0.92);
+  --bg-panel-border: rgba(0, 210, 120, 0.25);
+  --green-glow:  #00d878;
+  --green-bright:#3eff8a;
+  --green-dim:   #0a6b3a;
+  --amber:       #ffb020;
+  --amber-dim:   #a06800;
+  --blue-signal: #2090ff;
+  --blue-dim:    #0a3060;
+  --red-alert:   #ff3040;
+  --cyan:        #00e5ff;
+  --text-primary:   #e8ece0;
+  --text-secondary: rgba(232,236,224, 0.55);
+  --text-label:     rgba(232,236,224, 0.35);
+  --radius: 8px;
+}
+
+* { margin: 0; padding: 0; box-sizing: border-box; }
+
+body {
+  background: var(--bg-deep);
+  font-family: 'Inter', -apple-system, sans-serif;
+  color: var(--text-primary);
+  -webkit-font-smoothing: antialiased;
+  overflow-x: hidden;
+  min-height: 100vh;
+}
+
+/* === Header === */
+.header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 16px 24px;
+  border-bottom: 1px solid var(--bg-panel-border);
+  background: var(--bg-panel);
+  backdrop-filter: blur(12px);
+}
+
+.header-left {
+  display: flex;
+  align-items: center;
+  gap: 16px;
+}
+
+.logo {
+  font-weight: 700;
+  font-size: 24px;
+  color: var(--green-glow);
+}
+
+.logo .pi { font-style: normal; }
+
+.header-title {
+  font-size: 14px;
+  color: var(--text-secondary);
+  font-weight: 300;
+}
+
+.header-right {
+  display: flex;
+  align-items: center;
+  gap: 16px;
+}
+
+.mode-select {
+  background: rgba(0,210,120,0.1);
+  border: 1px solid var(--bg-panel-border);
+  color: var(--text-primary);
+  padding: 6px 12px;
+  border-radius: var(--radius);
+  font-family: inherit;
+  font-size: 13px;
+  cursor: pointer;
+}
+
+.mode-select option { background: #0c1420; }
+
+.status-badge {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 12px;
+  padding: 4px 10px;
+  border-radius: 12px;
+  background: rgba(0,210,120,0.1);
+  border: 1px solid var(--bg-panel-border);
+}
+
+.status-dot {
+  width: 8px; height: 8px;
+  border-radius: 50%;
+  background: var(--green-glow);
+  box-shadow: 0 0 8px var(--green-glow);
+  animation: pulse-dot 2s ease infinite;
+}
+
+.status-dot.offline { background: #555; box-shadow: none; animation: none; }
+.status-dot.warning { background: var(--amber); box-shadow: 0 0 8px var(--amber); }
+
+@keyframes pulse-dot {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0.5; }
+}
+
+.fps-badge {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 12px;
+  color: var(--green-glow);
+}
+
+.back-link {
+  color: var(--text-secondary);
+  text-decoration: none;
+  font-size: 13px;
+  transition: color 0.2s;
+}
+.back-link:hover { color: var(--green-glow); }
+
+/* === Main Layout === */
+.main-grid {
+  display: grid;
+  grid-template-columns: 1fr 360px;
+  grid-template-rows: 1fr auto;
+  gap: 16px;
+  padding: 16px 24px;
+  height: calc(100vh - 72px);
+  overflow: hidden;
+}
+
+.video-panel {
+  grid-row: 1;
+}
+
+.side-panels {
+  grid-row: 1;
+}
+
+/* === Video Panel === */
+.video-panel {
+  position: relative;
+  background: #000;
+  border-radius: var(--radius);
+  border: 1px solid var(--bg-panel-border);
+  overflow: hidden;
+  min-height: 0;
+}
+
+.video-panel video {
+  width: 100%;
+  height: 100%;
+  object-fit: cover;
+  transform: scaleX(-1);
+}
+
+.video-panel canvas {
+  position: absolute;
+  top: 0; left: 0;
+  width: 100%;
+  height: 100%;
+  transform: scaleX(-1);
+}
+
+.video-overlay-label {
+  position: absolute;
+  top: 12px; left: 12px;
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  padding: 4px 8px;
+  background: rgba(0,0,0,0.7);
+  border-radius: 4px;
+  color: var(--green-glow);
+  z-index: 5;
+  transform: scaleX(-1);
+}
+
+.camera-prompt {
+  position: absolute;
+  top: 0; left: 0; right: 0; bottom: 0;
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  text-align: center;
+  color: var(--text-secondary);
+  padding: 24px;
+  z-index: 6;
+  background: radial-gradient(ellipse at center, rgba(0,210,120,0.08) 0%, rgba(8,12,20,0.95) 70%);
+}
+
+.camera-prompt button {
+  margin-top: 16px;
+  padding: 10px 24px;
+  background: var(--green-glow);
+  color: #000;
+  border: none;
+  border-radius: var(--radius);
+  font-family: inherit;
+  font-weight: 600;
+  font-size: 14px;
+  cursor: pointer;
+  transition: background 0.2s;
+}
+
+.camera-prompt button:hover { background: var(--green-bright); }
+
+.camera-prompt-label {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 14px;
+  font-weight: 600;
+  letter-spacing: 2px;
+  color: var(--green-glow);
+  text-shadow: 0 0 12px rgba(0,216,120,0.4);
+  margin-bottom: 12px;
+}
+
+/* === Side Panels === */
+.side-panels {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+  overflow-y: auto;
+  min-height: 0;
+  max-height: 100%;
+  scrollbar-width: thin;
+  scrollbar-color: var(--green-dim) transparent;
+}
+
+.panel {
+  background: var(--bg-panel);
+  border: 1px solid var(--bg-panel-border);
+  border-radius: var(--radius);
+  padding: 10px 14px;
+  flex-shrink: 0;
+}
+
+.panel-title {
+  font-size: 11px;
+  text-transform: uppercase;
+  letter-spacing: 1.2px;
+  color: var(--text-label);
+  margin-bottom: 10px;
+  display: flex;
+  align-items: center;
+  gap: 6px;
+}
+
+/* === CSI Heatmap === */
+.csi-canvas-wrapper {
+  position: relative;
+  border-radius: 4px;
+  overflow: hidden;
+  background: #000;
+}
+
+.csi-canvas-wrapper canvas {
+  width: 100%;
+  display: block;
+}
+
+/* === Fusion Bars === */
+.fusion-bars {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+
+.bar-row {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+}
+
+.bar-label {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  color: var(--text-secondary);
+  width: 55px;
+  text-align: right;
+}
+
+.bar-track {
+  flex: 1;
+  height: 6px;
+  background: rgba(255,255,255,0.06);
+  border-radius: 3px;
+  overflow: hidden;
+}
+
+.bar-fill {
+  height: 100%;
+  border-radius: 3px;
+  transition: width 0.3s ease;
+}
+
+.bar-fill.video { background: var(--cyan); }
+.bar-fill.csi { background: var(--amber); }
+.bar-fill.fused { background: var(--green-glow); box-shadow: 0 0 8px var(--green-glow); }
+
+.bar-value {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  color: var(--text-primary);
+  width: 36px;
+}
+
+/* === Embedding Space === */
+.embedding-canvas-wrapper {
+  position: relative;
+  background: #000;
+  border-radius: 4px;
+  overflow: hidden;
+}
+.embedding-canvas-wrapper canvas {
+  width: 100%;
+  display: block;
+}
+
+/* === RuVector Pipeline === */
+.rv-pipeline {
+  display: flex;
+  align-items: center;
+  gap: 2px;
+  margin-bottom: 8px;
+  flex-wrap: wrap;
+}
+
+.rv-stage {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 10px;
+  padding: 3px 6px;
+  border-radius: 3px;
+  background: rgba(0,210,120,0.12);
+  border: 1px solid rgba(0,210,120,0.3);
+  color: var(--green-glow);
+  transition: all 0.3s;
+}
+
+.rv-stage.active {
+  background: rgba(0,210,120,0.25);
+  box-shadow: 0 0 6px rgba(0,210,120,0.3);
+}
+
+.rv-arrow {
+  font-size: 10px;
+  color: var(--text-label);
+}
+
+.rv-stats {
+  display: flex;
+  gap: 12px;
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 10px;
+  color: var(--text-secondary);
+}
+
+/* === Latency Panel === */
+.latency-grid {
+  display: grid;
+  grid-template-columns: repeat(4, 1fr);
+  gap: 6px;
+}
+
+.latency-item {
+  text-align: center;
+  padding: 6px 0;
+}
+
+.latency-value {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 16px;
+  font-weight: 600;
+  color: var(--green-glow);
+}
+
+.latency-label {
+  font-size: 10px;
+  color: var(--text-label);
+  margin-top: 2px;
+}
+
+/* === Controls === */
+.controls-row {
+  display: flex;
+  gap: 8px;
+  flex-wrap: wrap;
+}
+
+.btn {
+  padding: 6px 14px;
+  border: 1px solid var(--bg-panel-border);
+  background: rgba(0,210,120,0.08);
+  color: var(--text-primary);
+  border-radius: var(--radius);
+  font-family: inherit;
+  font-size: 12px;
+  cursor: pointer;
+  transition: all 0.2s;
+}
+.btn:hover { background: rgba(0,210,120,0.2); }
+.btn.active { background: var(--green-glow); color: #000; font-weight: 600; }
+
+.slider-row {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  margin-top: 8px;
+}
+
+.slider-row label {
+  font-size: 11px;
+  color: var(--text-secondary);
+  white-space: nowrap;
+}
+
+.slider-row input[type=range] {
+  flex: 1;
+  accent-color: var(--green-glow);
+}
+
+.slider-row .slider-val {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  width: 32px;
+  color: var(--green-glow);
+}
+
+/* === Bottom Bar === */
+.bottom-bar {
+  grid-column: 1 / -1;
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 10px 16px;
+  background: var(--bg-panel);
+  border: 1px solid var(--bg-panel-border);
+  border-radius: var(--radius);
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  color: var(--text-secondary);
+}
+
+.bottom-bar a {
+  color: var(--green-glow);
+  text-decoration: none;
+}
+
+/* === RSSI Signal Strength === */
+.rssi-row {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+}
+
+.rssi-gauge { flex: 1; }
+
+.rssi-bar-track {
+  height: 8px;
+  background: rgba(255,255,255,0.06);
+  border-radius: 4px;
+  overflow: hidden;
+  position: relative;
+}
+
+.rssi-bar-fill {
+  height: 100%;
+  border-radius: 4px;
+  background: linear-gradient(90deg, var(--red-alert), var(--amber), var(--green-glow));
+  transition: width 0.4s ease;
+  position: relative;
+  box-shadow: 0 0 6px rgba(0,210,120,0.3);
+}
+
+.rssi-bar-fill::after {
+  content: '';
+  position: absolute;
+  top: 0; left: 0; right: 0; bottom: 0;
+  background: linear-gradient(90deg, transparent 0%, rgba(255,255,255,0.2) 50%, transparent 100%);
+  animation: rssi-shimmer 2s ease-in-out infinite;
+}
+
+@keyframes rssi-shimmer {
+  0% { transform: translateX(-100%); }
+  100% { transform: translateX(100%); }
+}
+
+.rssi-values {
+  display: flex;
+  justify-content: space-between;
+  margin-top: 4px;
+}
+
+.rssi-dbm {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 14px;
+  font-weight: 600;
+  color: var(--green-glow);
+}
+
+.rssi-quality {
+  font-family: 'JetBrains Mono', monospace;
+  font-size: 11px;
+  color: var(--text-secondary);
+  text-transform: uppercase;
+}
+
+#rssi-sparkline {
+  flex-shrink: 0;
+  border-radius: 4px;
+  background: rgba(0,0,0,0.3);
+}
+
+/* === Skeleton colors === */
+.skeleton-joint { fill: var(--green-glow); }
+.skeleton-limb { stroke: var(--green-bright); }
+.skeleton-joint-csi { fill: var(--amber); }
+.skeleton-limb-csi { stroke: var(--amber); }
+
+/* === Responsive === */
+@media (max-width: 900px) {
+  .main-grid {
+    grid-template-columns: 1fr;
+    height: auto;
+    overflow: auto;
+  }
+  .video-panel { aspect-ratio: 16/9; max-height: 50vh; }
+  .side-panels { max-height: none; overflow: visible; }
+}

ui/pose-fusion/js/canvas-renderer.js

@@ -0,0 +1,307 @@
+/**
+ * CanvasRenderer — Renders skeleton overlay on video, CSI heatmap,
+ * embedding space visualization, and fusion confidence bars.
+ */
+
+import { SKELETON_CONNECTIONS } from './pose-decoder.js';
+
+export class CanvasRenderer {
+  constructor() {
+    this.colors = {
+      joint:      '#00d878',
+      jointGlow:  'rgba(0, 216, 120, 0.4)',
+      limb:       '#3eff8a',
+      limbGlow:   'rgba(62, 255, 138, 0.15)',
+      csiJoint:   '#ffb020',
+      csiLimb:    '#ffc850',
+      fused:      '#00e5ff',
+      confidence: 'rgba(255,255,255,0.3)',
+      videoEmb:   '#00e5ff',
+      csiEmb:     '#ffb020',
+      fusedEmb:   '#00d878',
+    };
+  }
+
+  /**
+   * Draw skeleton overlay on the video canvas
+   * @param {CanvasRenderingContext2D} ctx
+   * @param {Array<{x,y,confidence}>} keypoints - Normalized [0,1] coordinates
+   * @param {number} width - Canvas width
+   * @param {number} height - Canvas height
+   * @param {object} opts
+   */
+  drawSkeleton(ctx, keypoints, width, height, opts = {}) {
+    const minConf = opts.minConfidence || 0.3;
+    const color = opts.color || 'green';
+    const jointColor = color === 'amber' ? this.colors.csiJoint : this.colors.joint;
+    const limbColor = color === 'amber' ? this.colors.csiLimb : this.colors.limb;
+    const glowColor = color === 'amber' ? 'rgba(255,176,32,0.4)' : this.colors.jointGlow;
+
+    // Extended keypoint styling
+    const fingerColor = '#ff6ef0';    // Magenta for finger tips
+    const fingerGlow = 'rgba(255,110,240,0.4)';
+    const fingerLimb = 'rgba(255,110,240,0.5)';
+    const toeColor = '#6ef0ff';       // Cyan for toes
+    const neckColor = '#ffffff';       // White for neck
+
+    ctx.clearRect(0, 0, width, height);
+
+    if (!keypoints || keypoints.length === 0) return;
+
+    // Draw limbs first (behind joints)
+    ctx.lineCap = 'round';
+
+    for (const [i, j] of SKELETON_CONNECTIONS) {
+      const kpA = keypoints[i];
+      const kpB = keypoints[j];
+      if (!kpA || !kpB || kpA.confidence < minConf || kpB.confidence < minConf) continue;
+
+      const ax = kpA.x * width, ay = kpA.y * height;
+      const bx = kpB.x * width, by = kpB.y * height;
+      const avgConf = (kpA.confidence + kpB.confidence) / 2;
+
+      // Is this a hand/finger connection? (indices 17-22)
+      const isFingerLink = i >= 17 && i <= 22 || j >= 17 && j <= 22;
+      const isToeLink = i >= 23 && i <= 24 || j >= 23 && j <= 24;
+
+      // Glow
+      ctx.strokeStyle = isFingerLink ? fingerLimb : this.colors.limbGlow;
+      ctx.lineWidth = isFingerLink ? 4 : 8;
+      ctx.globalAlpha = avgConf * (isFingerLink ? 0.3 : 0.4);
+      ctx.beginPath();
+      ctx.moveTo(ax, ay);
+      ctx.lineTo(bx, by);
+      ctx.stroke();
+
+      // Main line
+      ctx.strokeStyle = isFingerLink ? fingerColor : isToeLink ? toeColor : limbColor;
+      ctx.lineWidth = isFingerLink || isToeLink ? 1.5 : 2.5;
+      ctx.globalAlpha = avgConf;
+      ctx.beginPath();
+      ctx.moveTo(ax, ay);
+      ctx.lineTo(bx, by);
+      ctx.stroke();
+    }
+
+    // Draw joints
+    ctx.globalAlpha = 1;
+    for (let idx = 0; idx < keypoints.length; idx++) {
+      const kp = keypoints[idx];
+      if (!kp || kp.confidence < minConf) continue;
+
+      const x = kp.x * width;
+      const y = kp.y * height;
+      const isFinger = idx >= 17 && idx <= 22;
+      const isToe = idx >= 23 && idx <= 24;
+      const isNeck = idx === 25;
+      const r = isFinger ? 2 + kp.confidence * 2 : isToe ? 2 : 3 + kp.confidence * 3;
+      const jColor = isFinger ? fingerColor : isToe ? toeColor : isNeck ? neckColor : jointColor;
+      const gColor = isFinger ? fingerGlow : glowColor;
+
+      // Glow
+      ctx.beginPath();
+      ctx.arc(x, y, r + (isFinger ? 3 : 4), 0, Math.PI * 2);
+      ctx.fillStyle = gColor;
+      ctx.globalAlpha = kp.confidence * (isFinger ? 0.5 : 0.6);
+      ctx.fill();
+
+      // Joint dot
+      ctx.beginPath();
+      ctx.arc(x, y, r, 0, Math.PI * 2);
+      ctx.fillStyle = jColor;
+      ctx.globalAlpha = kp.confidence;
+      ctx.fill();
+
+      // White center (body joints only)
+      if (!isFinger && !isToe) {
+        ctx.beginPath();
+        ctx.arc(x, y, r * 0.4, 0, Math.PI * 2);
+        ctx.fillStyle = '#fff';
+        ctx.globalAlpha = kp.confidence * 0.8;
+        ctx.fill();
+      }
+    }
+
+    ctx.globalAlpha = 1;
+
+    // Confidence label + keypoint count
+    if (opts.label) {
+      const visCount = keypoints.filter(kp => kp && kp.confidence >= minConf).length;
+      ctx.font = '11px "JetBrains Mono", monospace';
+      ctx.fillStyle = jointColor;
+      ctx.globalAlpha = 0.8;
+      ctx.fillText(`${opts.label} · ${visCount} joints`, 8, height - 8);
+      ctx.globalAlpha = 1;
+    }
+  }
+
+  /**
+   * Draw CSI amplitude heatmap
+   * @param {CanvasRenderingContext2D} ctx
+   * @param {{ data: Float32Array, width: number, height: number }} heatmap
+   * @param {number} canvasW
+   * @param {number} canvasH
+   */
+  drawCsiHeatmap(ctx, heatmap, canvasW, canvasH) {
+    ctx.clearRect(0, 0, canvasW, canvasH);
+
+    if (!heatmap || !heatmap.data || heatmap.height < 2) {
+      ctx.fillStyle = '#0a0e18';
+      ctx.fillRect(0, 0, canvasW, canvasH);
+      ctx.font = '11px "JetBrains Mono", monospace';
+      ctx.fillStyle = 'rgba(255,255,255,0.3)';
+      ctx.fillText('Waiting for CSI data...', 8, canvasH / 2);
+      return;
+    }
+
+    const { data, width: dw, height: dh } = heatmap;
+    const cellW = canvasW / dw;
+    const cellH = canvasH / dh;
+
+    for (let y = 0; y < dh; y++) {
+      for (let x = 0; x < dw; x++) {
+        const val = Math.min(1, Math.max(0, data[y * dw + x]));
+        ctx.fillStyle = this._heatmapColor(val);
+        ctx.fillRect(x * cellW, y * cellH, cellW + 0.5, cellH + 0.5);
+      }
+    }
+
+    // Axis labels
+    ctx.font = '9px "JetBrains Mono", monospace';
+    ctx.fillStyle = 'rgba(255,255,255,0.4)';
+    ctx.fillText('Subcarrier →', 4, canvasH - 4);
+    ctx.save();
+    ctx.translate(canvasW - 4, canvasH - 4);
+    ctx.rotate(-Math.PI / 2);
+    ctx.fillText('Time ↑', 0, 0);
+    ctx.restore();
+  }
+
+  /**
+   * Draw embedding space 2D projection
+   * @param {CanvasRenderingContext2D} ctx
+   * @param {{ video: Array, csi: Array, fused: Array }} points
+   * @param {number} w
+   * @param {number} h
+   */
+  drawEmbeddingSpace(ctx, points, w, h) {
+    ctx.fillStyle = '#050810';
+    ctx.fillRect(0, 0, w, h);
+
+    // Grid
+    ctx.strokeStyle = 'rgba(255,255,255,0.05)';
+    ctx.lineWidth = 0.5;
+    for (let i = 0; i <= 4; i++) {
+      const x = (i / 4) * w;
+      ctx.beginPath(); ctx.moveTo(x, 0); ctx.lineTo(x, h); ctx.stroke();
+      const y = (i / 4) * h;
+      ctx.beginPath(); ctx.moveTo(0, y); ctx.lineTo(w, y); ctx.stroke();
+    }
+
+    // Axes
+    ctx.strokeStyle = 'rgba(255,255,255,0.1)';
+    ctx.lineWidth = 1;
+    ctx.beginPath(); ctx.moveTo(w / 2, 0); ctx.lineTo(w / 2, h); ctx.stroke();
+    ctx.beginPath(); ctx.moveTo(0, h / 2); ctx.lineTo(w, h / 2); ctx.stroke();
+
+    // Auto-scale: find max extent across all point sets
+    let maxExtent = 0.01;
+    for (const pts of [points.video, points.csi, points.fused]) {
+      if (!pts) continue;
+      for (const p of pts) {
+        if (!p) continue;
+        maxExtent = Math.max(maxExtent, Math.abs(p[0]), Math.abs(p[1]));
+      }
+    }
+    const scale = 0.42 / maxExtent; // Fill ~84% of half-width
+
+    const drawPoints = (pts, color, size) => {
+      if (!pts || pts.length === 0) return;
+      const len = pts.length;
+
+      // Draw trail line connecting recent points
+      if (len >= 2) {
+        ctx.beginPath();
+        let started = false;
+        for (let i = 0; i < len; i++) {
+          const p = pts[i];
+          if (!p) continue;
+          const px = w / 2 + p[0] * scale * w;
+          const py = h / 2 + p[1] * scale * h;
+          if (px < -10 || px > w + 10 || py < -10 || py > h + 10) continue;
+          if (!started) { ctx.moveTo(px, py); started = true; }
+          else ctx.lineTo(px, py);
+        }
+        ctx.strokeStyle = color;
+        ctx.globalAlpha = 0.2;
+        ctx.lineWidth = 1;
+        ctx.stroke();
+      }
+
+      // Draw dots with glow on newest
+      for (let i = 0; i < len; i++) {
+        const p = pts[i];
+        if (!p) continue;
+        const age = 1 - (i / len) * 0.7;
+        const px = w / 2 + p[0] * scale * w;
+        const py = h / 2 + p[1] * scale * h;
+
+        if (px < -10 || px > w + 10 || py < -10 || py > h + 10) continue;
+
+        // Glow on newest point
+        if (i === len - 1) {
+          ctx.beginPath();
+          ctx.arc(px, py, size + 4, 0, Math.PI * 2);
+          ctx.fillStyle = color;
+          ctx.globalAlpha = 0.3;
+          ctx.fill();
+        }
+
+        ctx.beginPath();
+        ctx.arc(px, py, i === len - 1 ? size + 1 : size, 0, Math.PI * 2);
+        ctx.fillStyle = color;
+        ctx.globalAlpha = age * 0.8;
+        ctx.fill();
+      }
+    };
+
+    drawPoints(points.video, this.colors.videoEmb, 3);
+    drawPoints(points.csi, this.colors.csiEmb, 3);
+    drawPoints(points.fused, this.colors.fusedEmb, 4);
+    ctx.globalAlpha = 1;
+
+    // Legend
+    ctx.font = '9px "JetBrains Mono", monospace';
+    const legends = [
+      { color: this.colors.videoEmb, label: 'Video' },
+      { color: this.colors.csiEmb, label: 'CSI' },
+      { color: this.colors.fusedEmb, label: 'Fused' },
+    ];
+    legends.forEach((l, i) => {
+      const ly = 12 + i * 14;
+      ctx.fillStyle = l.color;
+      ctx.beginPath();
+      ctx.arc(10, ly - 3, 3, 0, Math.PI * 2);
+      ctx.fill();
+      ctx.fillStyle = 'rgba(255,255,255,0.5)';
+      ctx.fillText(l.label, 18, ly);
+    });
+  }
+
+  _heatmapColor(val) {
+    // Dark blue → cyan → green → yellow → red
+    if (val < 0.25) {
+      const t = val / 0.25;
+      return `rgb(${Math.floor(t * 20)}, ${Math.floor(20 + t * 60)}, ${Math.floor(60 + t * 100)})`;
+    } else if (val < 0.5) {
+      const t = (val - 0.25) / 0.25;
+      return `rgb(${Math.floor(20 + t * 20)}, ${Math.floor(80 + t * 100)}, ${Math.floor(160 - t * 60)})`;
+    } else if (val < 0.75) {
+      const t = (val - 0.5) / 0.25;
+      return `rgb(${Math.floor(40 + t * 180)}, ${Math.floor(180 + t * 75)}, ${Math.floor(100 - t * 80)})`;
+    } else {
+      const t = (val - 0.75) / 0.25;
+      return `rgb(${Math.floor(220 + t * 35)}, ${Math.floor(255 - t * 120)}, ${Math.floor(20 - t * 20)})`;
+    }
+  }
+}

ui/pose-fusion/js/cnn-embedder.js

@@ -0,0 +1,443 @@
+/**
+ * CNN Embedder — RuVector Attention-powered feature extractor.
+ *
+ * Uses the real ruvector-attention-wasm WASM module for Multi-Head Attention
+ * and Flash Attention on CSI/video data. Falls back to a JS Conv2D pipeline
+ * when WASM is not available.
+ *
+ * Pipeline: Conv2D → BatchNorm → ReLU → Pool → RuVector Attention → Project → L2 Normalize
+ * Two instances are created: one for video frames, one for CSI pseudo-images.
+ */
+
+// Seeded PRNG for deterministic weight initialization
+function mulberry32(seed) {
+  return function() {
+    let t = (seed += 0x6D2B79F5);
+    t = Math.imul(t ^ (t >>> 15), t | 1);
+    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
+  };
+}
+
+export class CnnEmbedder {
+  /**
+   * @param {object} opts
+   * @param {number} opts.inputSize   - Square input dimension (default 56 for speed)
+   * @param {number} opts.embeddingDim - Output embedding dimension (default 128)
+   * @param {boolean} opts.normalize  - L2 normalize output
+   * @param {number} opts.seed        - PRNG seed for weight init
+   */
+  constructor(opts = {}) {
+    this.inputSize = opts.inputSize || 56;
+    this.embeddingDim = opts.embeddingDim || 128;
+    this.normalize = opts.normalize !== false;
+    this.wasmEmbedder = null;
+    this.rvAttention = null;      // RuVector Multi-Head Attention (WASM)
+    this.rvFlash = null;          // RuVector Flash Attention (WASM)
+    this.rvHyperbolic = null;     // RuVector Hyperbolic Attention (hierarchical body)
+    this.rvMoE = null;            // RuVector Mixture-of-Experts (body-region routing)
+    this.rvLinear = null;         // RuVector Linear Attention (O(n) fast hand refinement)
+    this.rvLocalGlobal = null;    // RuVector Local-Global Attention (detail + context)
+    this.rvModule = null;         // RuVector WASM module reference
+    this.useRuVector = false;
+
+    // Initialize weights with deterministic PRNG
+    const rng = mulberry32(opts.seed || 42);
+    const randRange = (lo, hi) => lo + rng() * (hi - lo);
+
+    // Conv 3x3: 3 input channels → 16 output channels
+    this.convWeights = new Float32Array(3 * 3 * 3 * 16);
+    for (let i = 0; i < this.convWeights.length; i++) {
+      this.convWeights[i] = randRange(-0.15, 0.15);
+    }
+
+    // BatchNorm params (16 channels)
+    this.bnGamma = new Float32Array(16).fill(1.0);
+    this.bnBeta = new Float32Array(16).fill(0.0);
+    this.bnMean = new Float32Array(16).fill(0.0);
+    this.bnVar = new Float32Array(16).fill(1.0);
+
+    // Projection: 16 → embeddingDim (used when RuVector not available)
+    this.projWeights = new Float32Array(16 * this.embeddingDim);
+    for (let i = 0; i < this.projWeights.length; i++) {
+      this.projWeights[i] = randRange(-0.1, 0.1);
+    }
+
+    // Attention projection: attention_dim → embeddingDim
+    this.attnProjWeights = new Float32Array(16 * this.embeddingDim);
+    for (let i = 0; i < this.attnProjWeights.length; i++) {
+      this.attnProjWeights[i] = randRange(-0.08, 0.08);
+    }
+  }
+
+  /**
+   * Try to load RuVector attention WASM, then fall back to ruvector-cnn-wasm
+   * @param {string} wasmPath - Path to the WASM package directory
+   */
+  async tryLoadWasm(wasmPath) {
+    // First try: RuVector Attention WASM (the real thing — browser ESM build)
+    try {
+      const attnBase = new URL('../pkg/ruvector-attention/ruvector_attention_browser.js', import.meta.url).href;
+      const mod = await import(attnBase);
+      await mod.default();  // async WASM init via fetch
+      mod.init();
+
+      // Create all 6 attention mechanisms
+      this.rvAttention = new mod.WasmMultiHeadAttention(16, 4);
+      this.rvFlash = new mod.WasmFlashAttention(16, 8);
+      this.rvHyperbolic = new mod.WasmHyperbolicAttention(16, -1.0);
+      this.rvMoE = new mod.WasmMoEAttention(16, 3, 2);
+      this.rvLinear = new mod.WasmLinearAttention(16, 16);
+      this.rvLocalGlobal = new mod.WasmLocalGlobalAttention(16, 4, 2);
+      this.rvModule = mod;
+      this.useRuVector = true;
+
+      // Log available mechanisms
+      const mechs = mod.available_mechanisms();
+      console.log(`[CNN] RuVector WASM v${mod.version()} — all 6 attention mechanisms active`, mechs);
+      return true;
+    } catch (e) {
+      console.log('[CNN] RuVector Attention WASM not available:', e.message);
+    }
+
+    // Second try: ruvector-cnn-wasm (legacy path)
+    try {
+      const mod = await import(`${wasmPath}/ruvector_cnn_wasm.js`);
+      await mod.default();
+      const config = new mod.EmbedderConfig();
+      config.input_size = this.inputSize;
+      config.embedding_dim = this.embeddingDim;
+      config.normalize = this.normalize;
+      this.wasmEmbedder = new mod.WasmCnnEmbedder(config);
+      console.log('[CNN] WASM CNN embedder loaded successfully');
+      return true;
+    } catch (e) {
+      console.log('[CNN] WASM CNN not available, using JS fallback:', e.message);
+      return false;
+    }
+  }
+
+  /**
+   * Extract embedding from RGB image data
+   * @param {Uint8Array} rgbData - RGB pixel data (H*W*3)
+   * @param {number} width
+   * @param {number} height
+   * @returns {Float32Array} embedding vector
+   */
+  extract(rgbData, width, height) {
+    if (this.wasmEmbedder) {
+      try {
+        const result = this.wasmEmbedder.extract(rgbData, width, height);
+        return new Float32Array(result);
+      } catch (_) { /* fallback to JS */ }
+    }
+    return this._extractJS(rgbData, width, height);
+  }
+
+  _extractJS(rgbData, width, height) {
+    // 1. Resize to inputSize × inputSize if needed
+    const sz = this.inputSize;
+    let input;
+    if (width === sz && height === sz) {
+      input = new Float32Array(rgbData.length);
+      for (let i = 0; i < rgbData.length; i++) input[i] = rgbData[i] / 255.0;
+    } else {
+      input = this._resize(rgbData, width, height, sz, sz);
+    }
+
+    // 2. ImageNet normalization
+    const mean = [0.485, 0.456, 0.406];
+    const std = [0.229, 0.224, 0.225];
+    const pixels = sz * sz;
+    for (let i = 0; i < pixels; i++) {
+      input[i * 3]     = (input[i * 3]     - mean[0]) / std[0];
+      input[i * 3 + 1] = (input[i * 3 + 1] - mean[1]) / std[1];
+      input[i * 3 + 2] = (input[i * 3 + 2] - mean[2]) / std[2];
+    }
+
+    // 3. Conv2D 3x3 (3 → 16 channels)
+    const convOut = this._conv2d3x3(input, sz, sz, 3, 16);
+
+    // 4. BatchNorm
+    this._batchNorm(convOut, 16);
+
+    // 5. ReLU
+    for (let i = 0; i < convOut.length; i++) {
+      if (convOut[i] < 0) convOut[i] = 0;
+    }
+
+    // 6. Global average pooling → spatial tokens (each 16-dim)
+    const outH = sz - 2, outW = sz - 2;
+    const spatial = outH * outW;
+
+    // 7. RuVector Attention (if loaded) — apply attention over spatial tokens
+    if (this.useRuVector && this.rvAttention) {
+      return this._extractWithAttention(convOut, spatial, 16);
+    }
+
+    // Fallback: simple global average pool + linear projection
+    const pooled = new Float32Array(16);
+    for (let i = 0; i < spatial; i++) {
+      for (let c = 0; c < 16; c++) {
+        pooled[c] += convOut[i * 16 + c];
+      }
+    }
+    for (let c = 0; c < 16; c++) pooled[c] /= spatial;
+
+    // Linear projection → embeddingDim
+    const emb = new Float32Array(this.embeddingDim);
+    for (let o = 0; o < this.embeddingDim; o++) {
+      let sum = 0;
+      for (let i = 0; i < 16; i++) {
+        sum += pooled[i] * this.projWeights[i * this.embeddingDim + o];
+      }
+      emb[o] = sum;
+    }
+
+    // L2 normalize
+    if (this.normalize) {
+      let norm = 0;
+      for (let i = 0; i < emb.length; i++) norm += emb[i] * emb[i];
+      norm = Math.sqrt(norm);
+      if (norm > 1e-8) {
+        for (let i = 0; i < emb.length; i++) emb[i] /= norm;
+      }
+    }
+
+    return emb;
+  }
+
+  /**
+   * Full 6-stage RuVector WASM attention pipeline:
+   * 1. Flash Attention (efficient O(n) pre-screening of spatial tokens)
+   * 2. Multi-Head Attention (global spatial reasoning)
+   * 3. Hyperbolic Attention (hierarchical body-part structure, Poincaré ball)
+   * 4. Linear Attention (O(n) refinement for fine detail — hands/extremities)
+   * 5. MoE Attention (body-region specialized expert routing)
+   * 6. Local-Global Attention (local detail + global context fusion)
+   * → Weighted blend + batch_normalize + project + L2 normalize
+   */
+  _extractWithAttention(convOut, numTokens, channels) {
+    const mod = this.rvModule;
+
+    // Subsample spatial tokens for attention (max 64 for speed)
+    const maxTokens = 64;
+    const step = numTokens > maxTokens ? Math.floor(numTokens / maxTokens) : 1;
+    const tokens = [];
+    for (let i = 0; i < numTokens && tokens.length < maxTokens; i += step) {
+      const token = new Float32Array(channels);
+      for (let c = 0; c < channels; c++) {
+        token[c] = convOut[i * channels + c];
+      }
+      tokens.push(token);
+    }
+
+    const numQueries = Math.min(4, tokens.length);
+    const queryStride = Math.floor(tokens.length / numQueries);
+
+    // === Stage 1: Flash Attention (efficient pre-screening) ===
+    const flashOut = new Float32Array(channels);
+    try {
+      // Flash attention with block size 8 for efficient O(n) screening
+      const result = this.rvFlash.compute(tokens[0], tokens, tokens);
+      for (let c = 0; c < channels; c++) flashOut[c] = result[c];
+    } catch (_) {
+      flashOut.set(tokens[0]);
+    }
+
+    // === Stage 2: Multi-Head Attention (global spatial reasoning) ===
+    const mhaOut = new Float32Array(channels);
+    for (let q = 0; q < numQueries; q++) {
+      const queryToken = tokens[q * queryStride];
+      try {
+        const result = this.rvAttention.compute(queryToken, tokens, tokens);
+        for (let c = 0; c < channels; c++) mhaOut[c] += result[c] / numQueries;
+      } catch (_) {
+        for (let c = 0; c < channels; c++) mhaOut[c] += queryToken[c] / numQueries;
+      }
+    }
+
+    // === Stage 3: Hyperbolic Attention (hierarchical body structure) ===
+    const hyOut = new Float32Array(channels);
+    try {
+      const result = this.rvHyperbolic.compute(mhaOut, tokens, tokens);
+      for (let c = 0; c < channels; c++) hyOut[c] = result[c];
+    } catch (_) {
+      hyOut.set(mhaOut);
+    }
+
+    // === Stage 4: Linear Attention (O(n) fast refinement for extremities) ===
+    const linOut = new Float32Array(channels);
+    try {
+      const result = this.rvLinear.compute(hyOut, tokens, tokens);
+      for (let c = 0; c < channels; c++) linOut[c] = result[c];
+    } catch (_) {
+      linOut.set(hyOut);
+    }
+
+    // === Stage 5: MoE Attention (body-region expert routing) ===
+    const moeOut = new Float32Array(channels);
+    try {
+      const result = this.rvMoE.compute(linOut, tokens, tokens);
+      for (let c = 0; c < channels; c++) moeOut[c] = result[c];
+    } catch (_) {
+      moeOut.set(linOut);
+    }
+
+    // === Stage 6: Local-Global Attention (detail + context) ===
+    const lgOut = new Float32Array(channels);
+    try {
+      const result = this.rvLocalGlobal.compute(moeOut, tokens, tokens);
+      for (let c = 0; c < channels; c++) lgOut[c] = result[c];
+    } catch (_) {
+      lgOut.set(moeOut);
+    }
+
+    // === Blend all 6 outputs ===
+    // Use WASM softmax on log-energy scores for dynamic stage weighting
+    const blended = new Float32Array(channels);
+    const stages = [flashOut, mhaOut, hyOut, linOut, moeOut, lgOut];
+    // Use log-energy to prevent exp() overflow in softmax
+    const logEnergies = new Float32Array(6);
+    for (let s = 0; s < 6; s++) {
+      const e = this._energy(stages[s]);
+      logEnergies[s] = e > 1e-10 ? Math.log(e) : -20;
+    }
+    try { mod.softmax(logEnergies); } catch (_) {
+      let max = -Infinity;
+      for (let i = 0; i < 6; i++) max = Math.max(max, logEnergies[i]);
+      let sum = 0;
+      for (let i = 0; i < 6; i++) { logEnergies[i] = Math.exp(logEnergies[i] - max); sum += logEnergies[i]; }
+      for (let i = 0; i < 6; i++) logEnergies[i] /= sum;
+    }
+    for (let c = 0; c < channels; c++) {
+      for (let s = 0; s < 6; s++) {
+        blended[c] += logEnergies[s] * stages[s][c];
+      }
+    }
+
+    // Batch normalize only when we have enough diversity (skip for single vectors)
+    // Single-vector batch norm collapses to zeros, killing embedding space
+    let normed = blended;
+
+    // Project to embeddingDim
+    const emb = new Float32Array(this.embeddingDim);
+    for (let o = 0; o < this.embeddingDim; o++) {
+      let sum = 0;
+      for (let i = 0; i < channels; i++) {
+        sum += normed[i] * this.attnProjWeights[i * this.embeddingDim + o];
+      }
+      emb[o] = sum;
+    }
+
+    // L2 normalize using RuVector WASM
+    if (this.normalize) {
+      try { mod.normalize(emb); } catch (_) {
+        let norm = 0;
+        for (let i = 0; i < emb.length; i++) norm += emb[i] * emb[i];
+        norm = Math.sqrt(norm);
+        if (norm > 1e-8) for (let i = 0; i < emb.length; i++) emb[i] /= norm;
+      }
+    }
+
+    return emb;
+  }
+
+  /** Compute vector energy (L2 norm squared) for attention weighting */
+  _energy(vec) {
+    let e = 0;
+    for (let i = 0; i < vec.length; i++) e += vec[i] * vec[i];
+    return e;
+  }
+
+  _conv2d3x3(input, H, W, Cin, Cout) {
+    const outH = H - 2, outW = W - 2;
+    const output = new Float32Array(outH * outW * Cout);
+    for (let y = 0; y < outH; y++) {
+      for (let x = 0; x < outW; x++) {
+        for (let co = 0; co < Cout; co++) {
+          let sum = 0;
+          for (let ky = 0; ky < 3; ky++) {
+            for (let kx = 0; kx < 3; kx++) {
+              for (let ci = 0; ci < Cin; ci++) {
+                const px = ((y + ky) * W + (x + kx)) * Cin + ci;
+                const wt = (((ky * 3 + kx) * Cin) + ci) * Cout + co;
+                sum += input[px] * this.convWeights[wt];
+              }
+            }
+          }
+          output[(y * outW + x) * Cout + co] = sum;
+        }
+      }
+    }
+    return output;
+  }
+
+  _batchNorm(data, channels) {
+    const spatial = data.length / channels;
+    for (let i = 0; i < spatial; i++) {
+      for (let c = 0; c < channels; c++) {
+        const idx = i * channels + c;
+        data[idx] = this.bnGamma[c] * (data[idx] - this.bnMean[c]) / Math.sqrt(this.bnVar[c] + 1e-5) + this.bnBeta[c];
+      }
+    }
+  }
+
+  _resize(rgbData, srcW, srcH, dstW, dstH) {
+    const output = new Float32Array(dstW * dstH * 3);
+    const xRatio = srcW / dstW;
+    const yRatio = srcH / dstH;
+    for (let y = 0; y < dstH; y++) {
+      for (let x = 0; x < dstW; x++) {
+        const sx = Math.min(Math.floor(x * xRatio), srcW - 1);
+        const sy = Math.min(Math.floor(y * yRatio), srcH - 1);
+        const srcIdx = (sy * srcW + sx) * 3;
+        const dstIdx = (y * dstW + x) * 3;
+        output[dstIdx]     = rgbData[srcIdx]     / 255.0;
+        output[dstIdx + 1] = rgbData[srcIdx + 1] / 255.0;
+        output[dstIdx + 2] = rgbData[srcIdx + 2] / 255.0;
+      }
+    }
+    return output;
+  }
+
+  /** Cosine similarity using WASM when available, JS fallback */
+  cosineSim(a, b) {
+    if (this.rvModule) {
+      try { return this.rvModule.cosine_similarity(a, b); } catch (_) { /* fallback */ }
+    }
+    return CnnEmbedder.cosineSimilarity(a, b);
+  }
+
+  /** L2 norm using WASM when available */
+  l2Norm(vec) {
+    if (this.rvModule) {
+      try { return this.rvModule.l2_norm(vec); } catch (_) { /* fallback */ }
+    }
+    let norm = 0;
+    for (let i = 0; i < vec.length; i++) norm += vec[i] * vec[i];
+    return Math.sqrt(norm);
+  }
+
+  /** Pairwise distance matrix using WASM (for skeleton validation) */
+  pairwiseDistances(vectors) {
+    if (this.rvModule) {
+      try { return this.rvModule.pairwise_distances(vectors); } catch (_) { /* fallback */ }
+    }
+    return null;
+  }
+
+  /** Static JS fallback for cosine similarity */
+  static cosineSimilarity(a, b) {
+    let dot = 0, normA = 0, normB = 0;
+    for (let i = 0; i < a.length; i++) {
+      dot += a[i] * b[i];
+      normA += a[i] * a[i];
+      normB += b[i] * b[i];
+    }
+    normA = Math.sqrt(normA);
+    normB = Math.sqrt(normB);
+    if (normA < 1e-8 || normB < 1e-8) return 0;
+    return dot / (normA * normB);
+  }
+}

ui/pose-fusion/js/csi-simulator.js

@@ -0,0 +1,363 @@
+/**
+ * CSI Simulator — Generates realistic WiFi Channel State Information data.
+ *
+ * In live mode, connects to the sensing server via WebSocket.
+ * In demo mode, generates synthetic CSI that correlates with detected motion.
+ *
+ * Outputs: 3-channel pseudo-image (amplitude, phase, temporal diff)
+ * matching the ADR-018 frame format expectations.
+ */
+
+export class CsiSimulator {
+  static VERSION = 'v4-drift';  // Cache-bust verification
+
+  constructor(opts = {}) {
+    this.subcarriers = opts.subcarriers || 52; // 802.11n HT20
+    this.timeWindow = opts.timeWindow || 56;   // frames in sliding window
+    this.mode = 'demo'; // 'demo' | 'live'
+    this.ws = null;
+
+    // Circular buffer for CSI frames
+    this.amplitudeBuffer = [];
+    this.phaseBuffer = [];
+    this.frameCount = 0;
+
+    // Noise parameters
+    this._rng = this._mulberry32(opts.seed || 7);
+    this._noiseState = new Float32Array(this.subcarriers);
+    this._baseAmplitude = new Float32Array(this.subcarriers);
+    this._basePhase = new Float32Array(this.subcarriers);
+
+    // Initialize base CSI profile (empty room)
+    for (let i = 0; i < this.subcarriers; i++) {
+      this._baseAmplitude[i] = 0.5 + 0.3 * Math.sin(i * 0.12);
+      this._basePhase[i] = (i / this.subcarriers) * Math.PI * 2;
+    }
+
+    // RSSI tracking
+    this.rssiDbm = -70; // default mid-range
+    this._rssiTarget = -70;
+
+    // Person influence (updated from video motion)
+    this.personPresence = 0;
+    this.personX = 0.5;
+    this.personY = 0.5;
+    this.personMotion = 0;
+  }
+
+  /**
+   * Connect to live sensing server WebSocket
+   * @param {string} url - WebSocket URL (e.g. ws://localhost:3030/ws/csi)
+   */
+  async connectLive(url) {
+    return new Promise((resolve) => {
+      try {
+        this.ws = new WebSocket(url);
+        this.ws.binaryType = 'arraybuffer';
+        this.ws.onmessage = (evt) => this._handleLiveFrame(evt.data);
+        this.ws.onopen = () => { this.mode = 'live'; resolve(true); };
+        this.ws.onerror = () => resolve(false);
+        this.ws.onclose = () => { this.mode = 'demo'; };
+        // Timeout after 3s
+        setTimeout(() => { if (this.mode !== 'live') resolve(false); }, 3000);
+      } catch {
+        resolve(false);
+      }
+    });
+  }
+
+  disconnect() {
+    if (this.ws) { this.ws.close(); this.ws = null; }
+    this.mode = 'demo';
+  }
+
+  get isLive() { return this.mode === 'live'; }
+
+  /**
+   * Update person state from video detection (for correlated demo data).
+   * When person exits frame, CSI maintains presence with slow decay
+   * (simulating through-wall sensing capability).
+   */
+  updatePersonState(presence, x, y, motion) {
+    // Don't override real CSI sensing with synthetic video-derived state
+    if (this.mode === 'live') return;
+
+    if (presence > 0.1) {
+      // Person detected in video — update CSI state directly
+      this.personPresence = presence;
+      this.personX = x;
+      this.personY = y;
+      this.personMotion = motion;
+      this._lastSeenTime = performance.now();
+      this._lastSeenX = x;
+      this._lastSeenY = y;
+    } else if (this._lastSeenTime) {
+      // Person NOT in video — CSI "through-wall" persistence
+      const elapsed = (performance.now() - this._lastSeenTime) / 1000;
+      // CSI can sense through walls for ~10 seconds with decaying confidence
+      const decayRate = 0.15; // Lose ~15% per second
+      this.personPresence = Math.max(0, 1.0 - elapsed * decayRate);
+      // Position slowly drifts (person walking behind wall)
+      this.personX = this._lastSeenX;
+      this.personY = this._lastSeenY;
+      this.personMotion = Math.max(0, motion * 0.5 + this.personPresence * 0.2);
+
+      if (this.personPresence < 0.05) {
+        this._lastSeenTime = null;
+      }
+    } else {
+      this.personPresence = 0;
+      this.personMotion = 0;
+    }
+  }
+
+  /**
+   * Generate next CSI frame (demo mode) or return latest live frame
+   * @param {number} elapsed - Time in seconds
+   * @returns {{ amplitude: Float32Array, phase: Float32Array, snr: number }}
+   */
+  nextFrame(elapsed) {
+    const amp = new Float32Array(this.subcarriers);
+    const phase = new Float32Array(this.subcarriers);
+
+    if (this.mode === 'live' && this._liveAmplitude) {
+      amp.set(this._liveAmplitude);
+      phase.set(this._livePhase);
+    } else {
+      this._generateDemoFrame(amp, phase, elapsed);
+    }
+
+    // Push to circular buffer
+    this.amplitudeBuffer.push(new Float32Array(amp));
+    this.phaseBuffer.push(new Float32Array(phase));
+    if (this.amplitudeBuffer.length > this.timeWindow) {
+      this.amplitudeBuffer.shift();
+      this.phaseBuffer.shift();
+    }
+
+    // RSSI: smooth toward target (demo mode generates synthetic RSSI)
+    if (this.mode === 'demo') {
+      // Simulate RSSI based on person presence and slow drift
+      this._rssiTarget = -55 - 25 * (1 - this.personPresence) + Math.sin(elapsed * 0.3) * 3;
+    }
+    this.rssiDbm += (this._rssiTarget - this.rssiDbm) * 0.1;
+
+    // SNR estimate
+    let signalPower = 0, noisePower = 0;
+    for (let i = 0; i < this.subcarriers; i++) {
+      signalPower += amp[i] * amp[i];
+      noisePower += this._noiseState[i] * this._noiseState[i];
+    }
+    const snr = noisePower > 0 ? 10 * Math.log10(signalPower / noisePower) : 30;
+
+    this.frameCount++;
+    return { amplitude: amp, phase, snr: Math.max(0, Math.min(40, snr)) };
+  }
+
+  /**
+   * Build 3-channel pseudo-image for CNN input
+   * @param {number} targetSize - Output image dimension (square)
+   * @returns {Uint8Array} RGB data (targetSize * targetSize * 3)
+   */
+  buildPseudoImage(targetSize = 56) {
+    const buf = this.amplitudeBuffer;
+    const pBuf = this.phaseBuffer;
+    const frames = buf.length;
+    if (frames < 2) {
+      return new Uint8Array(targetSize * targetSize * 3);
+    }
+
+    const rgb = new Uint8Array(targetSize * targetSize * 3);
+
+    for (let y = 0; y < targetSize; y++) {
+      const fi = Math.min(Math.floor(y / targetSize * frames), frames - 1);
+      for (let x = 0; x < targetSize; x++) {
+        const si = Math.min(Math.floor(x / targetSize * this.subcarriers), this.subcarriers - 1);
+        const idx = (y * targetSize + x) * 3;
+
+        // R: Amplitude (normalized to 0-255)
+        const ampVal = buf[fi][si];
+        rgb[idx] = Math.min(255, Math.max(0, Math.floor(ampVal * 255)));
+
+        // G: Phase (wrapped to 0-255)
+        const phaseVal = (pBuf[fi][si] % (2 * Math.PI) + 2 * Math.PI) % (2 * Math.PI);
+        rgb[idx + 1] = Math.floor(phaseVal / (2 * Math.PI) * 255);
+
+        // B: Temporal difference
+        if (fi > 0) {
+          const diff = Math.abs(buf[fi][si] - buf[fi - 1][si]);
+          rgb[idx + 2] = Math.min(255, Math.floor(diff * 500));
+        }
+      }
+    }
+
+    return rgb;
+  }
+
+  /**
+   * Get heatmap data for visualization
+   * @returns {{ data: Float32Array, width: number, height: number }}
+   */
+  getHeatmapData() {
+    const frames = this.amplitudeBuffer.length;
+    const w = this.subcarriers;
+    const h = Math.min(frames, this.timeWindow);
+    const data = new Float32Array(w * h);
+    for (let y = 0; y < h; y++) {
+      const fi = frames - h + y;
+      if (fi >= 0 && fi < frames) {
+        for (let x = 0; x < w; x++) {
+          data[y * w + x] = this.amplitudeBuffer[fi][x];
+        }
+      }
+    }
+    return { data, width: w, height: h };
+  }
+
+  // === Private ===
+
+  _generateDemoFrame(amp, phase, elapsed) {
+    const rng = this._rng;
+    const presence = this.personPresence;
+    const motion = this.personMotion;
+    const px = this.personX;
+
+    for (let i = 0; i < this.subcarriers; i++) {
+      // Base CSI profile (frequency-selective channel)
+      let a = this._baseAmplitude[i];
+      let p = this._basePhase[i] + elapsed * 0.05;
+
+      // Environmental noise (correlated across subcarriers)
+      this._noiseState[i] = 0.95 * this._noiseState[i] + 0.05 * (rng() * 2 - 1) * 0.03;
+      a += this._noiseState[i];
+
+      // Ambient temporal drift (multipath fading even in empty room)
+      a += 0.06 * Math.sin(elapsed * 0.7 + i * 0.25)
+         + 0.04 * Math.sin(elapsed * 1.3 - i * 0.18)
+         + 0.03 * Math.cos(elapsed * 2.1 + i * 0.4);
+
+      // Person-induced CSI perturbation
+      if (presence > 0.1) {
+        // Subcarrier-dependent body reflection (Fresnel zone model)
+        const freqOffset = (i - this.subcarriers * px) / (this.subcarriers * 0.3);
+        const bodyReflection = presence * 0.25 * Math.exp(-freqOffset * freqOffset);
+
+        // Motion causes amplitude fluctuation
+        const motionEffect = motion * 0.15 * Math.sin(elapsed * 3.5 + i * 0.3);
+
+        // Breathing modulation (0.2-0.3 Hz)
+        const breathing = presence * 0.02 * Math.sin(elapsed * 1.5 + i * 0.05);
+
+        a += bodyReflection + motionEffect + breathing;
+        p += presence * 0.4 * Math.sin(elapsed * 2.1 + i * 0.15);
+      }
+
+      amp[i] = Math.max(0, Math.min(1, a));
+      phase[i] = p;
+    }
+  }
+
+  _handleLiveFrame(data) {
+    // Handle JSON text frames from the sensing server
+    if (typeof data === 'string') {
+      try {
+        const msg = JSON.parse(data);
+        this._handleJsonFrame(msg);
+      } catch (_) { /* ignore malformed JSON */ }
+      return;
+    }
+
+    // Handle Blob data (convert to ArrayBuffer and re-process)
+    if (data instanceof Blob) {
+      data.arrayBuffer().then(ab => this._handleLiveFrame(ab)).catch(() => {});
+      return;
+    }
+
+    // Handle binary ArrayBuffer frames (ADR-018 format)
+    if (!(data instanceof ArrayBuffer)) return;
+    const view = new DataView(data);
+    // Check ADR-018 magic: 0xC5110001
+    if (data.byteLength < 20) return;
+    const magic = view.getUint32(0, true);
+    if (magic !== 0xC5110001) return;
+
+    const numSub = Math.min(view.getUint16(8, true), this.subcarriers);
+    this._liveAmplitude = new Float32Array(this.subcarriers);
+    this._livePhase = new Float32Array(this.subcarriers);
+
+    const headerSize = 20;
+    for (let i = 0; i < numSub && (headerSize + i * 4 + 3) < data.byteLength; i++) {
+      const real = view.getInt16(headerSize + i * 4, true);
+      const imag = view.getInt16(headerSize + i * 4 + 2, true);
+      this._liveAmplitude[i] = Math.sqrt(real * real + imag * imag) / 2048;
+      this._livePhase[i] = Math.atan2(imag, real);
+    }
+  }
+
+  _handleJsonFrame(msg) {
+    // Sensing server sends: { type: "sensing_update", nodes: [{ amplitude: [...], subcarrier_count }], classification, features }
+    this._liveAmplitude = new Float32Array(this.subcarriers);
+    this._livePhase = new Float32Array(this.subcarriers);
+
+    // Extract amplitude from sensing_update node data
+    const node = (msg.nodes && msg.nodes[0]) || msg;
+    const ampArr = node.amplitude || msg.amplitude;
+    if (ampArr && Array.isArray(ampArr)) {
+      const n = Math.min(ampArr.length, this.subcarriers);
+      // Server sends raw amplitude (already magnitude), normalize to 0-1
+      let maxAmp = 0;
+      for (let i = 0; i < n; i++) maxAmp = Math.max(maxAmp, Math.abs(ampArr[i]));
+      const scale = maxAmp > 0 ? 1.0 / maxAmp : 1.0;
+      for (let i = 0; i < n; i++) {
+        this._liveAmplitude[i] = Math.abs(ampArr[i]) * scale;
+      }
+    }
+
+    // Phase from node (if available)
+    const phaseArr = node.phase || msg.phase;
+    if (phaseArr && Array.isArray(phaseArr)) {
+      const n = Math.min(phaseArr.length, this.subcarriers);
+      for (let i = 0; i < n; i++) this._livePhase[i] = phaseArr[i];
+    } else if (ampArr) {
+      // Synthesize phase from amplitude variation (Hilbert-like estimate)
+      for (let i = 1; i < this.subcarriers; i++) {
+        this._livePhase[i] = this._livePhase[i - 1] + (this._liveAmplitude[i] - this._liveAmplitude[i - 1]) * Math.PI;
+      }
+    }
+
+    // Handle raw I/Q pairs
+    const iq = node.iq || msg.iq;
+    if (iq && Array.isArray(iq)) {
+      const n = Math.min(iq.length / 2, this.subcarriers);
+      for (let i = 0; i < n; i++) {
+        const real = iq[i * 2], imag = iq[i * 2 + 1];
+        this._liveAmplitude[i] = Math.sqrt(real * real + imag * imag) / 2048;
+        this._livePhase[i] = Math.atan2(imag, real);
+      }
+    }
+
+    // Extract RSSI from node data
+    if (typeof node.rssi_dbm === 'number') {
+      this._rssiTarget = node.rssi_dbm;
+    } else if (msg.features && typeof msg.features.mean_rssi === 'number') {
+      this._rssiTarget = msg.features.mean_rssi;
+    }
+
+    // Update presence from server classification
+    const cls = msg.classification;
+    if (cls) {
+      if (typeof cls.confidence === 'number') {
+        this.personPresence = cls.presence ? cls.confidence : 0;
+      }
+    }
+  }
+
+  _mulberry32(seed) {
+    return function() {
+      let t = (seed += 0x6D2B79F5);
+      t = Math.imul(t ^ (t >>> 15), t | 1);
+      t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
+      return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
+    };
+  }
+}

ui/pose-fusion/js/fusion-engine.js

@@ -0,0 +1,183 @@
+/**
+ * FusionEngine — Attention-weighted dual-modal embedding fusion.
+ *
+ * Combines visual (camera) and CSI (WiFi) embeddings with dynamic
+ * confidence gating based on signal quality.
+ */
+
+export class FusionEngine {
+  /**
+   * @param {number} embeddingDim
+   * @param {object} opts
+   * @param {object} opts.wasmModule - RuVector WASM module for cosine_similarity etc.
+   */
+  constructor(embeddingDim = 128, opts = {}) {
+    this.embeddingDim = embeddingDim;
+    this.wasmModule = opts.wasmModule || null;
+
+    // Learnable attention weights (initialized to balanced 0.5)
+    this.attentionWeights = new Float32Array(embeddingDim).fill(0.5);
+
+    // Dynamic modality confidence [0, 1]
+    this.videoConfidence = 1.0;
+    this.csiConfidence = 0.0;
+    this.fusedConfidence = 0.5;
+
+    // Smoothing for confidence transitions
+    this._smoothAlpha = 0.85;
+
+    // Embedding history for visualization
+    this.recentVideoEmbeddings = [];
+    this.recentCsiEmbeddings = [];
+    this.recentFusedEmbeddings = [];
+    this.maxHistory = 50;
+  }
+
+  /** Set the WASM module reference (called after WASM loads) */
+  setWasmModule(mod) { this.wasmModule = mod; }
+
+  /**
+   * Update quality-based confidence scores
+   * @param {number} videoBrightness - [0,1] video brightness quality
+   * @param {number} videoMotion     - [0,1] motion detected
+   * @param {number} csiSnr          - CSI signal-to-noise ratio in dB
+   * @param {boolean} csiActive      - Whether CSI source is connected
+   */
+  updateConfidence(videoBrightness, videoMotion, csiSnr, csiActive) {
+    // Video confidence: drops with low brightness, boosted by motion
+    let vc = 0;
+    if (videoBrightness > 0.05) {
+      vc = Math.min(1, videoBrightness * 1.5) * 0.7 + Math.min(1, videoMotion * 3) * 0.3;
+    }
+
+    // CSI confidence: based on SNR and connection status
+    let cc = 0;
+    if (csiActive) {
+      cc = Math.min(1, csiSnr / 25); // 25dB = full confidence
+    }
+
+    // Smooth transitions
+    this.videoConfidence = this._smoothAlpha * this.videoConfidence + (1 - this._smoothAlpha) * vc;
+    this.csiConfidence = this._smoothAlpha * this.csiConfidence + (1 - this._smoothAlpha) * cc;
+
+    // Fused confidence is the max of either (fusion can only help)
+    this.fusedConfidence = Math.min(1, Math.sqrt(
+      this.videoConfidence * this.videoConfidence + this.csiConfidence * this.csiConfidence
+    ));
+  }
+
+  /**
+   * Fuse video and CSI embeddings
+   * @param {Float32Array|null} videoEmb - Visual embedding (or null if video-off)
+   * @param {Float32Array|null} csiEmb   - CSI embedding (or null if CSI-off)
+   * @param {string} mode                - 'dual' | 'video' | 'csi'
+   * @returns {Float32Array} Fused embedding
+   */
+  fuse(videoEmb, csiEmb, mode = 'dual') {
+    const dim = this.embeddingDim;
+    const fused = new Float32Array(dim);
+
+    if (mode === 'video' || !csiEmb) {
+      if (videoEmb) fused.set(videoEmb);
+      this._recordEmbedding(videoEmb, null, fused);
+      return fused;
+    }
+
+    if (mode === 'csi' || !videoEmb) {
+      if (csiEmb) fused.set(csiEmb);
+      this._recordEmbedding(null, csiEmb, fused);
+      return fused;
+    }
+
+    // Dual mode: attention-weighted fusion with confidence gating
+    const totalConf = this.videoConfidence + this.csiConfidence;
+    const videoWeight = totalConf > 0 ? this.videoConfidence / totalConf : 0.5;
+
+    for (let i = 0; i < dim; i++) {
+      const alpha = this.attentionWeights[i] * videoWeight +
+                    (1 - this.attentionWeights[i]) * (1 - videoWeight);
+      fused[i] = alpha * videoEmb[i] + (1 - alpha) * csiEmb[i];
+    }
+
+    // Re-normalize using WASM when available
+    if (this.wasmModule) {
+      try { this.wasmModule.normalize(fused); } catch (_) { this._jsNormalize(fused); }
+    } else {
+      this._jsNormalize(fused);
+    }
+
+    this._recordEmbedding(videoEmb, csiEmb, fused);
+    return fused;
+  }
+
+  /**
+   * Get embedding pairs for 2D visualization (PCA projection)
+   * @returns {{ video: Array, csi: Array, fused: Array }}
+   */
+  getEmbeddingPoints() {
+    // Sparse random projection: pick a few dimensions with fixed coefficients
+    // to get visible 2D spread (avoids cancellation from summing all 128 dims)
+    const project = (emb) => {
+      if (!emb || emb.length < 4) return null;
+      // Use 8 sparse dimensions with predetermined signs (seeded, not random)
+      const dim = emb.length;
+      const x = emb[0] * 3.2 - emb[3] * 2.8 + emb[7] * 2.1 - emb[12] * 1.9
+              + (dim > 30 ? emb[29] * 1.5 - emb[31] * 1.3 : 0)
+              + (dim > 60 ? emb[55] * 1.1 - emb[60] * 0.9 : 0);
+      const y = emb[1] * 3.0 - emb[5] * 2.5 + emb[9] * 2.3 - emb[15] * 1.7
+              + (dim > 40 ? emb[37] * 1.4 - emb[42] * 1.2 : 0)
+              + (dim > 80 ? emb[73] * 1.0 - emb[80] * 0.8 : 0);
+      return [x, y];
+    };
+
+    return {
+      video: this.recentVideoEmbeddings.map(project).filter(Boolean),
+      csi: this.recentCsiEmbeddings.map(project).filter(Boolean),
+      fused: this.recentFusedEmbeddings.map(project).filter(Boolean)
+    };
+  }
+
+  /**
+   * Cross-modal similarity score
+   * @returns {number} Cosine similarity between latest video and CSI embeddings
+   */
+  getCrossModalSimilarity() {
+    const v = this.recentVideoEmbeddings[this.recentVideoEmbeddings.length - 1];
+    const c = this.recentCsiEmbeddings[this.recentCsiEmbeddings.length - 1];
+    if (!v || !c) return 0;
+
+    // Use WASM cosine_similarity when available
+    if (this.wasmModule) {
+      try { return this.wasmModule.cosine_similarity(v, c); } catch (_) { /* fallback */ }
+    }
+
+    let dot = 0, na = 0, nb = 0;
+    for (let i = 0; i < v.length; i++) {
+      dot += v[i] * c[i];
+      na += v[i] * v[i];
+      nb += c[i] * c[i];
+    }
+    na = Math.sqrt(na); nb = Math.sqrt(nb);
+    return (na > 1e-8 && nb > 1e-8) ? dot / (na * nb) : 0;
+  }
+
+  _jsNormalize(vec) {
+    let norm = 0;
+    for (let i = 0; i < vec.length; i++) norm += vec[i] * vec[i];
+    norm = Math.sqrt(norm);
+    if (norm > 1e-8) for (let i = 0; i < vec.length; i++) vec[i] /= norm;
+  }
+
+  _recordEmbedding(video, csi, fused) {
+    if (video) {
+      this.recentVideoEmbeddings.push(new Float32Array(video));
+      if (this.recentVideoEmbeddings.length > this.maxHistory) this.recentVideoEmbeddings.shift();
+    }
+    if (csi) {
+      this.recentCsiEmbeddings.push(new Float32Array(csi));
+      if (this.recentCsiEmbeddings.length > this.maxHistory) this.recentCsiEmbeddings.shift();
+    }
+    this.recentFusedEmbeddings.push(new Float32Array(fused));
+    if (this.recentFusedEmbeddings.length > this.maxHistory) this.recentFusedEmbeddings.shift();
+  }
+}

ui/pose-fusion/js/main.js

@@ -0,0 +1,472 @@
+/**
+ * RuView — Dual-Modal Pose Estimation Demo
+ *
+ * Main orchestration: video capture → CNN embedding → CSI processing → fusion → rendering
+ */
+
+import { VideoCapture } from './video-capture.js?v=13';
+import { CsiSimulator } from './csi-simulator.js?v=13';
+import { CnnEmbedder } from './cnn-embedder.js?v=13';
+import { FusionEngine } from './fusion-engine.js?v=13';
+import { PoseDecoder } from './pose-decoder.js?v=13';
+import { CanvasRenderer } from './canvas-renderer.js?v=13';
+
+// === State ===
+let mode = 'dual';  // 'dual' | 'video' | 'csi'
+let isRunning = false;
+let isPaused = false;
+let startTime = 0;
+let frameCount = 0;
+let fps = 0;
+let lastFpsTime = 0;
+let confidenceThreshold = 0.3;
+
+// Latency tracking
+const latency = { video: 0, csi: 0, fusion: 0, total: 0 };
+
+// === Components ===
+const videoCapture = new VideoCapture(document.getElementById('webcam'));
+const csiSimulator = new CsiSimulator({ subcarriers: 52, timeWindow: 56 });
+const visualCnn = new CnnEmbedder({ inputSize: 56, embeddingDim: 128, seed: 42 });
+const csiCnn = new CnnEmbedder({ inputSize: 56, embeddingDim: 128, seed: 137 });
+const fusionEngine = new FusionEngine(128);
+const poseDecoder = new PoseDecoder(128);
+const renderer = new CanvasRenderer();
+
+// === Canvas Elements ===
+const skeletonCanvas = document.getElementById('skeleton-canvas');
+const skeletonCtx = skeletonCanvas.getContext('2d');
+const csiCanvas = document.getElementById('csi-canvas');
+const csiCtx = csiCanvas.getContext('2d');
+const embeddingCanvas = document.getElementById('embedding-canvas');
+const embeddingCtx = embeddingCanvas.getContext('2d');
+
+// === UI Elements ===
+const modeSelect = document.getElementById('mode-select');
+const statusDot = document.getElementById('status-dot');
+const statusLabel = document.getElementById('status-label');
+const fpsDisplay = document.getElementById('fps-display');
+const cameraPrompt = document.getElementById('camera-prompt');
+const startCameraBtn = document.getElementById('start-camera-btn');
+const pauseBtn = document.getElementById('pause-btn');
+const confSlider = document.getElementById('confidence-slider');
+const confValue = document.getElementById('confidence-value');
+const wsUrlInput = document.getElementById('ws-url');
+const connectWsBtn = document.getElementById('connect-ws-btn');
+
+// Fusion bar elements
+const videoBar = document.getElementById('video-bar');
+const csiBar = document.getElementById('csi-bar');
+const fusedBar = document.getElementById('fused-bar');
+const videoBarVal = document.getElementById('video-bar-val');
+const csiBarVal = document.getElementById('csi-bar-val');
+const fusedBarVal = document.getElementById('fused-bar-val');
+
+// Latency elements
+const latVideoEl = document.getElementById('lat-video');
+const latCsiEl = document.getElementById('lat-csi');
+const latFusionEl = document.getElementById('lat-fusion');
+const latTotalEl = document.getElementById('lat-total');
+
+// Cross-modal similarity
+const crossModalEl = document.getElementById('cross-modal-sim');
+
+// RSSI elements
+const rssiBarEl = document.getElementById('rssi-bar');
+const rssiValueEl = document.getElementById('rssi-value');
+const rssiQualityEl = document.getElementById('rssi-quality');
+const rssiSparkCanvas = document.getElementById('rssi-sparkline');
+const rssiSparkCtx = rssiSparkCanvas ? rssiSparkCanvas.getContext('2d') : null;
+const rssiHistory = [];
+const RSSI_HISTORY_MAX = 80;
+
+// === Initialize ===
+function init() {
+  console.log(`[PoseFusion] init() v4 — CsiSimulator=${CsiSimulator.VERSION || 'OLD'}, starting...`);
+  resizeCanvases();
+  console.log(`[PoseFusion] canvases: skeleton=${skeletonCanvas.width}x${skeletonCanvas.height}, csi=${csiCanvas.width}x${csiCanvas.height}, emb=${embeddingCanvas.width}x${embeddingCanvas.height}`);
+  window.addEventListener('resize', resizeCanvases);
+
+  // Mode change
+  modeSelect.addEventListener('change', (e) => {
+    mode = e.target.value;
+    updateModeUI();
+  });
+
+  // Camera start
+  startCameraBtn.addEventListener('click', startCamera);
+
+  // Pause
+  pauseBtn.addEventListener('click', () => {
+    isPaused = !isPaused;
+    pauseBtn.textContent = isPaused ? '▶ Resume' : '⏸ Pause';
+    pauseBtn.classList.toggle('active', isPaused);
+  });
+
+  // Confidence slider
+  confSlider.addEventListener('input', (e) => {
+    confidenceThreshold = parseFloat(e.target.value);
+    confValue.textContent = confidenceThreshold.toFixed(2);
+  });
+
+  // WebSocket connect
+  connectWsBtn.addEventListener('click', async () => {
+    const url = wsUrlInput.value.trim();
+    if (!url) return;
+    connectWsBtn.textContent = 'Connecting...';
+    const ok = await csiSimulator.connectLive(url);
+    connectWsBtn.textContent = ok ? '✓ Connected' : 'Connect';
+    if (ok) {
+      connectWsBtn.classList.add('active');
+    }
+  });
+
+  // Try to load RuVector Attention WASM embedders (non-blocking)
+  const wasmBase = new URL('../pkg/ruvector-attention', import.meta.url).href;
+  visualCnn.tryLoadWasm(wasmBase).then((ok) => {
+    // Share the WASM module with FusionEngine for cosine_similarity, normalize, etc.
+    if (visualCnn.rvModule) fusionEngine.setWasmModule(visualCnn.rvModule);
+    // Update footer backend label
+    const backendEl = document.getElementById('cnn-backend');
+    if (backendEl) {
+      backendEl.textContent = ok && visualCnn.useRuVector
+        ? `RuVector WASM v${visualCnn.rvModule.version()} — 6 attention mechanisms`
+        : 'ruvector-cnn (JS fallback)';
+    }
+  });
+  csiCnn.tryLoadWasm(wasmBase);
+
+  // Auto-connect to local sensing server WebSocket if available
+  const defaultWsUrl = 'ws://localhost:8765/ws/sensing';
+  if (wsUrlInput) wsUrlInput.value = defaultWsUrl;
+  csiSimulator.connectLive(defaultWsUrl).then(ok => {
+    if (ok && connectWsBtn) {
+      connectWsBtn.textContent = '✓ Live ESP32';
+      connectWsBtn.classList.add('active');
+      statusLabel.textContent = 'LIVE CSI';
+      statusDot.classList.remove('offline');
+    }
+  });
+
+  // Auto-start camera for video/dual modes
+  updateModeUI();
+  startTime = performance.now() / 1000;
+  isRunning = true;
+  requestAnimationFrame(mainLoop);
+}
+
+async function startCamera() {
+  cameraPrompt.style.display = 'none';
+  const ok = await videoCapture.start();
+  if (ok) {
+    statusDot.classList.remove('offline');
+    statusLabel.textContent = 'LIVE';
+    resizeCanvases();
+  } else {
+    cameraPrompt.style.display = 'flex';
+    cameraPrompt.querySelector('p').textContent = 'Camera access denied. Try CSI-only mode.';
+  }
+}
+
+function updateModeUI() {
+  const needsVideo = mode !== 'csi';
+
+  // Show/hide camera prompt
+  if (needsVideo && !videoCapture.isActive) {
+    cameraPrompt.style.display = 'flex';
+  } else {
+    cameraPrompt.style.display = 'none';
+  }
+
+  // Update mode label in both the overlay and the camera prompt
+  const labelMap = { dual: 'DUAL FUSION', video: 'VIDEO ONLY', csi: 'CSI ONLY' };
+  const modeLabel = document.getElementById('mode-label');
+  const promptLabel = document.getElementById('prompt-mode-label');
+  if (modeLabel) modeLabel.textContent = labelMap[mode] || mode;
+  if (promptLabel) promptLabel.textContent = labelMap[mode] || mode;
+}
+
+function resizeCanvases() {
+  const videoPanel = document.querySelector('.video-panel');
+  if (videoPanel) {
+    const rect = videoPanel.getBoundingClientRect();
+    skeletonCanvas.width = rect.width;
+    skeletonCanvas.height = rect.height;
+  }
+
+  // CSI canvas (min 200px width)
+  csiCanvas.width = Math.max(200, csiCanvas.parentElement.clientWidth);
+  csiCanvas.height = 120;
+
+  // Embedding canvas (min 200px width)
+  embeddingCanvas.width = Math.max(200, embeddingCanvas.parentElement.clientWidth);
+  embeddingCanvas.height = 140;
+}
+
+// === Main Loop ===
+let _loopErrorShown = false;
+let _diagDone = false;
+function mainLoop(timestamp) {
+  if (!isRunning) return;
+  requestAnimationFrame(mainLoop);
+
+  if (isPaused) return;
+
+  try {
+  const elapsed = performance.now() / 1000 - startTime;
+  const totalStart = performance.now();
+
+  // --- Video Pipeline ---
+  let videoEmb = null;
+  let motionRegion = null;
+  if (mode !== 'csi' && videoCapture.isActive) {
+    const t0 = performance.now();
+    const frame = videoCapture.captureFrame(56, 56);
+    if (frame) {
+      videoEmb = visualCnn.extract(frame.rgb, frame.width, frame.height);
+      motionRegion = videoCapture.detectMotionRegion(56, 56);
+
+      // Feed motion to CSI simulator for correlated demo data
+      // When detected=false, CSI simulator handles through-wall persistence
+      csiSimulator.updatePersonState(
+        motionRegion.detected ? 1.0 : 0,
+        motionRegion.detected ? motionRegion.x + motionRegion.w / 2 : 0.5,
+        motionRegion.detected ? motionRegion.y + motionRegion.h / 2 : 0.5,
+        frame.motion
+      );
+
+      fusionEngine.updateConfidence(
+        frame.brightness, frame.motion,
+        0, csiSimulator.isLive || mode === 'dual'
+      );
+    }
+    latency.video = performance.now() - t0;
+  }
+
+  // --- CSI Pipeline ---
+  let csiEmb = null;
+  if (mode !== 'video') {
+    const t0 = performance.now();
+    const csiFrame = csiSimulator.nextFrame(elapsed);
+    const pseudoImage = csiSimulator.buildPseudoImage(56);
+    csiEmb = csiCnn.extract(pseudoImage, 56, 56);
+
+    fusionEngine.updateConfidence(
+      videoCapture.brightnessScore,
+      videoCapture.motionScore,
+      csiFrame.snr,
+      true
+    );
+
+    // Draw CSI heatmap
+    const heatmap = csiSimulator.getHeatmapData();
+    renderer.drawCsiHeatmap(csiCtx, heatmap, csiCanvas.width, csiCanvas.height);
+
+    latency.csi = performance.now() - t0;
+  }
+
+  // --- Fusion ---
+  const t0f = performance.now();
+  const fusedEmb = fusionEngine.fuse(videoEmb, csiEmb, mode);
+  latency.fusion = performance.now() - t0f;
+
+  // --- Pose Decode ---
+  // For CSI-only mode, generate a synthetic motion region from CSI energy
+  if (mode === 'csi' && (!motionRegion || !motionRegion.detected)) {
+    const csiPresence = csiSimulator.personPresence;
+    if (csiPresence > 0.1) {
+      motionRegion = {
+        detected: true,
+        x: 0.25, y: 0.15, w: 0.5, h: 0.7,
+        coverage: csiPresence,
+        motionGrid: null,
+        gridCols: 10,
+        gridRows: 8
+      };
+    }
+  }
+
+  // CSI state for through-wall tracking
+  const csiState = {
+    csiPresence: csiSimulator.personPresence,
+    isLive: csiSimulator.isLive
+  };
+
+  const keypoints = poseDecoder.decode(fusedEmb, motionRegion, elapsed, csiState);
+
+  // --- Render Skeleton ---
+  const labelMap = { dual: 'DUAL FUSION', video: 'VIDEO ONLY', csi: 'CSI ONLY' };
+  renderer.drawSkeleton(skeletonCtx, keypoints, skeletonCanvas.width, skeletonCanvas.height, {
+    minConfidence: confidenceThreshold,
+    color: mode === 'csi' ? 'amber' : 'green',
+    label: labelMap[mode]
+  });
+
+  // --- Render Embedding Space ---
+  const embPoints = fusionEngine.getEmbeddingPoints();
+  renderer.drawEmbeddingSpace(embeddingCtx, embPoints, embeddingCanvas.width, embeddingCanvas.height);
+
+  // --- Update UI ---
+  latency.total = performance.now() - totalStart;
+
+  // FPS
+  frameCount++;
+  if (timestamp - lastFpsTime > 500) {
+    fps = Math.round(frameCount * 1000 / (timestamp - lastFpsTime));
+    lastFpsTime = timestamp;
+    frameCount = 0;
+    fpsDisplay.textContent = `${fps} FPS`;
+  }
+
+  // Fusion bars
+  const vc = fusionEngine.videoConfidence;
+  const cc = fusionEngine.csiConfidence;
+  const fc = fusionEngine.fusedConfidence;
+  videoBar.style.width = `${vc * 100}%`;
+  csiBar.style.width = `${cc * 100}%`;
+  fusedBar.style.width = `${fc * 100}%`;
+  videoBarVal.textContent = `${Math.round(vc * 100)}%`;
+  csiBarVal.textContent = `${Math.round(cc * 100)}%`;
+  fusedBarVal.textContent = `${Math.round(fc * 100)}%`;
+
+  // Latency
+  latVideoEl.textContent = `${latency.video.toFixed(1)}ms`;
+  latCsiEl.textContent = `${latency.csi.toFixed(1)}ms`;
+  latFusionEl.textContent = `${latency.fusion.toFixed(1)}ms`;
+  latTotalEl.textContent = `${latency.total.toFixed(1)}ms`;
+
+  // Cross-modal similarity
+  const sim = fusionEngine.getCrossModalSimilarity();
+  crossModalEl.textContent = sim.toFixed(3);
+
+  // RuVector attention pipeline stats
+  const rvStats = poseDecoder.attentionStats;
+  const rvEnergyEl = document.getElementById('rv-energy');
+  const rvRefineEl = document.getElementById('rv-refine');
+  const rvImpactEl = document.getElementById('rv-impact');
+  if (rvEnergyEl) rvEnergyEl.textContent = (rvStats.energy || 0).toFixed(2);
+  if (rvRefineEl) rvRefineEl.textContent = ((rvStats.refinementMag || 0) * 1000).toFixed(1) + 'px';
+  if (rvImpactEl) {
+    const impact = Math.min(100, (rvStats.refinementMag || 0) * 5000);
+    rvImpactEl.textContent = impact.toFixed(0) + '%';
+  }
+  // Pulse the pipeline stages when active
+  if (visualCnn.useRuVector && rvStats.energy > 0.1) {
+    document.querySelectorAll('.rv-stage').forEach(el => el.classList.add('active'));
+  }
+
+  // RSSI update
+  updateRssi(csiSimulator.rssiDbm);
+
+  // One-time diagnostic
+  if (!_diagDone) {
+    _diagDone = true;
+    console.log(`[PoseFusion] frame 1 OK — mode=${mode}, csi.bufLen=${csiSimulator.amplitudeBuffer.length}, embPts=${embPoints?.fused?.length ?? 0}, rssi=${(csiSimulator.rssiDbm ?? -99).toFixed(1)}`);
+  }
+
+  } catch (err) {
+    if (!_loopErrorShown) {
+      _loopErrorShown = true;
+      console.error('[MainLoop]', err);
+      // Show error visually on page
+      const errDiv = document.createElement('div');
+      errDiv.style.cssText = 'position:fixed;bottom:60px;left:24px;right:24px;background:rgba(255,48,64,0.95);color:#fff;padding:12px 16px;border-radius:8px;font:12px/1.4 "JetBrains Mono",monospace;z-index:9999;max-height:120px;overflow:auto';
+      errDiv.textContent = `[MainLoop Error] ${err.message}\n${err.stack?.split('\n').slice(0,3).join('\n')}`;
+      document.body.appendChild(errDiv);
+    }
+  }
+}
+
+// === RSSI Visualization ===
+function updateRssi(dbm) {
+  if (!rssiBarEl) return;
+
+  // Clamp to typical WiFi range: -100 (worst) to -30 (best)
+  const clamped = Math.max(-100, Math.min(-30, dbm));
+  const pct = ((clamped + 100) / 70) * 100; // 0-100%
+
+  rssiBarEl.style.width = `${pct}%`;
+  rssiValueEl.textContent = `${Math.round(clamped)} dBm`;
+
+  // Quality label
+  let quality;
+  if (clamped > -50) quality = 'Excellent';
+  else if (clamped > -60) quality = 'Good';
+  else if (clamped > -70) quality = 'Fair';
+  else if (clamped > -80) quality = 'Weak';
+  else quality = 'Poor';
+  rssiQualityEl.textContent = quality;
+
+  // Color the dBm value based on quality
+  if (clamped > -60) rssiValueEl.style.color = 'var(--green-glow)';
+  else if (clamped > -75) rssiValueEl.style.color = 'var(--amber)';
+  else rssiValueEl.style.color = 'var(--red-alert)';
+
+  // Sparkline history
+  rssiHistory.push(clamped);
+  if (rssiHistory.length > RSSI_HISTORY_MAX) rssiHistory.shift();
+  drawRssiSparkline();
+}
+
+function drawRssiSparkline() {
+  if (!rssiSparkCtx || rssiHistory.length < 2) return;
+  const w = rssiSparkCanvas.width;
+  const h = rssiSparkCanvas.height;
+  const ctx = rssiSparkCtx;
+
+  ctx.clearRect(0, 0, w, h);
+
+  // Draw signal strength line
+  const len = rssiHistory.length;
+  const step = w / (RSSI_HISTORY_MAX - 1);
+
+  // Gradient fill under line
+  const grad = ctx.createLinearGradient(0, 0, 0, h);
+  grad.addColorStop(0, 'rgba(0,210,120,0.3)');
+  grad.addColorStop(1, 'rgba(0,210,120,0)');
+
+  ctx.beginPath();
+  for (let i = 0; i < len; i++) {
+    const x = (RSSI_HISTORY_MAX - len + i) * step;
+    const y = h - ((rssiHistory[i] + 100) / 70) * h;
+    if (i === 0) ctx.moveTo(x, y);
+    else ctx.lineTo(x, y);
+  }
+  // Fill area
+  const lastX = (RSSI_HISTORY_MAX - 1) * step;
+  const firstX = (RSSI_HISTORY_MAX - len) * step;
+  ctx.lineTo(lastX, h);
+  ctx.lineTo(firstX, h);
+  ctx.closePath();
+  ctx.fillStyle = grad;
+  ctx.fill();
+
+  // Draw line on top
+  ctx.beginPath();
+  for (let i = 0; i < len; i++) {
+    const x = (RSSI_HISTORY_MAX - len + i) * step;
+    const y = h - ((rssiHistory[i] + 100) / 70) * h;
+    if (i === 0) ctx.moveTo(x, y);
+    else ctx.lineTo(x, y);
+  }
+  ctx.strokeStyle = '#00d878';
+  ctx.lineWidth = 1.5;
+  ctx.stroke();
+
+  // Pulsing dot at latest value
+  const latestX = lastX;
+  const latestY = h - ((rssiHistory[len - 1] + 100) / 70) * h;
+  const pulse = 0.5 + 0.5 * Math.sin(performance.now() / 300);
+  ctx.beginPath();
+  ctx.arc(latestX, latestY, 2 + pulse, 0, Math.PI * 2);
+  ctx.fillStyle = '#00d878';
+  ctx.fill();
+  ctx.beginPath();
+  ctx.arc(latestX, latestY, 4 + pulse * 2, 0, Math.PI * 2);
+  ctx.strokeStyle = `rgba(0,216,120,${0.3 + pulse * 0.3})`;
+  ctx.lineWidth = 1;
+  ctx.stroke();
+}
+
+// Boot
+document.addEventListener('DOMContentLoaded', init);

ui/pose-fusion/js/pose-decoder.js

@@ -0,0 +1,553 @@
+/**
+ * PoseDecoder — Maps motion detection grid → 17 COCO keypoints.
+ *
+ * Uses per-cell motion intensity to track actual body part positions:
+ * - Head: top-center motion cluster
+ * - Shoulders/Elbows/Wrists: lateral motion in upper body zone
+ * - Hips/Knees/Ankles: lower body motion distribution
+ *
+ * When person exits frame, CSI data continues tracking (through-wall mode).
+ */
+
+// Extended keypoint definitions: 17 COCO + 9 hand/fingertip approximations = 26 total
+export const KEYPOINT_NAMES = [
+  'nose', 'left_eye', 'right_eye', 'left_ear', 'right_ear',
+  'left_shoulder', 'right_shoulder', 'left_elbow', 'right_elbow',
+  'left_wrist', 'right_wrist', 'left_hip', 'right_hip',
+  'left_knee', 'right_knee', 'left_ankle', 'right_ankle',
+  // Extended: hand keypoints (17-25)
+  'left_thumb', 'left_index', 'left_pinky',       // 17, 18, 19
+  'right_thumb', 'right_index', 'right_pinky',    // 20, 21, 22
+  'left_foot_index', 'right_foot_index',           // 23, 24 (toe tips)
+  'neck',                                          // 25 (mid-shoulder)
+];
+
+// Skeleton connections (pairs of keypoint indices)
+export const SKELETON_CONNECTIONS = [
+  [0, 1], [0, 2], [1, 3], [2, 4],           // Head
+  [0, 25],                                    // Nose → neck
+  [25, 5], [25, 6],                           // Neck → shoulders
+  [5, 7], [7, 9],                             // Left arm
+  [6, 8], [8, 10],                            // Right arm
+  [5, 11], [6, 12],                           // Torso
+  [11, 12],                                   // Hips
+  [11, 13], [13, 15],                         // Left leg
+  [12, 14], [14, 16],                         // Right leg
+  // Hand connections
+  [9, 17], [9, 18], [9, 19],                 // Left wrist → fingers
+  [10, 20], [10, 21], [10, 22],              // Right wrist → fingers
+  // Foot connections
+  [15, 23], [16, 24],                         // Ankles → toes
+];
+
+// Standard body proportions (relative to body height)
+const PROPORTIONS = {
+  headToShoulder: 0.15,
+  shoulderWidth: 0.25,
+  shoulderToElbow: 0.18,
+  elbowToWrist: 0.16,
+  shoulderToHip: 0.30,
+  hipWidth: 0.18,
+  hipToKnee: 0.24,
+  kneeToAnkle: 0.24,
+  eyeSpacing: 0.04,
+  earSpacing: 0.07,
+  // Hand proportions
+  wristToFinger: 0.09,
+  fingerSpread: 0.04,
+  thumbAngle: 0.6,    // radians from wrist-elbow axis
+  // Foot proportions
+  ankleToToe: 0.06,
+};
+
+export class PoseDecoder {
+  constructor(embeddingDim = 128) {
+    this.embeddingDim = embeddingDim;
+    this.smoothedKeypoints = null;
+    this.smoothingFactor = 0.25; // Low = responsive to real movement
+    this._time = 0;
+
+    // Through-wall tracking state
+    this._lastBodyState = null;
+    this._ghostState = null;
+    this._ghostConfidence = 0;
+    this._ghostVelocity = { x: 0, y: 0 };
+
+    // Zone centroid tracking (normalized 0-1 positions)
+    this._headCx = 0.5;
+    this._headCy = 0.15;
+    this._leftArmCx = 0.3;
+    this._leftArmCy = 0.35;
+    this._rightArmCx = 0.7;
+    this._rightArmCy = 0.35;
+    this._leftLegCx = 0.4;
+    this._leftLegCy = 0.8;
+    this._rightLegCx = 0.6;
+    this._rightLegCy = 0.8;
+    this._torsoCx = 0.5;
+    this._torsoCy = 0.45;
+
+    // RuVector embedding → joint mapping
+    // Each joint gets 2 consecutive embedding dimensions (dx, dy offset)
+    // and 1 dimension for confidence modulation. 26 joints × 3 = 78 dims used from 128.
+    // Remaining 50 dims encode global pose features (body scale, rotation, lean).
+    this._jointEmbMap = this._buildJointEmbeddingMap(embeddingDim);
+
+    // Attention contribution tracking (for UI overlay)
+    this.attentionStats = { energy: 0, maxDim: 0, refinementMag: 0 };
+  }
+
+  /**
+   * Build the mapping from embedding dimensions to joint refinement signals.
+   * This maps the RuVector attention output to anatomically meaningful joint offsets.
+   */
+  _buildJointEmbeddingMap(dim) {
+    const map = [];
+    // 26 joints × 3 dims each (dx, dy, confidence_mod) = 78 dims
+    for (let j = 0; j < 26; j++) {
+      const base = j * 3;
+      if (base + 2 < dim) {
+        map.push({ dxDim: base, dyDim: base + 1, confDim: base + 2 });
+      } else {
+        map.push({ dxDim: j % dim, dyDim: (j + 1) % dim, confDim: (j + 2) % dim });
+      }
+    }
+    // Global pose features from dims 78-127
+    return {
+      joints: map,
+      scaleDim: Math.min(78, dim - 1),       // body scale factor
+      rotDim: Math.min(79, dim - 1),          // body rotation
+      leanXDim: Math.min(80, dim - 1),        // lateral lean
+      leanYDim: Math.min(81, dim - 1),        // forward/back lean
+    };
+  }
+
+  /**
+   * Decode motion data into 17 keypoints
+   * @param {Float32Array} embedding - Fused embedding vector
+   * @param {{ detected, x, y, w, h, motionGrid, gridCols, gridRows, motionCx, motionCy, exitDirection }} motionRegion
+   * @param {number} elapsed - Time in seconds
+   * @param {{ csiPresence: number }} csiState - CSI sensing state for through-wall
+   * @returns {Array<{x: number, y: number, confidence: number, name: string}>}
+   */
+  decode(embedding, motionRegion, elapsed, csiState = {}) {
+    this._time = elapsed;
+
+    const hasMotion = motionRegion && motionRegion.detected;
+    const hasCsi = csiState && csiState.csiPresence > 0.1;
+
+    if (hasMotion) {
+      // Active tracking from video motion grid
+      this._ghostConfidence = 0;
+      const rawKeypoints = this._trackFromMotionGrid(motionRegion, embedding, elapsed);
+      this._lastBodyState = { keypoints: rawKeypoints.map(kp => ({...kp})), time: elapsed };
+
+      // Track exit velocity
+      if (motionRegion.exitDirection) {
+        const speed = 0.008;
+        this._ghostVelocity = {
+          x: motionRegion.exitDirection === 'left' ? -speed : motionRegion.exitDirection === 'right' ? speed : 0,
+          y: motionRegion.exitDirection === 'up' ? -speed : motionRegion.exitDirection === 'down' ? speed : 0
+        };
+      }
+
+      // Apply temporal smoothing
+      if (this.smoothedKeypoints && this.smoothedKeypoints.length === rawKeypoints.length) {
+        const alpha = this.smoothingFactor;
+        for (let i = 0; i < rawKeypoints.length; i++) {
+          rawKeypoints[i].x = alpha * this.smoothedKeypoints[i].x + (1 - alpha) * rawKeypoints[i].x;
+          rawKeypoints[i].y = alpha * this.smoothedKeypoints[i].y + (1 - alpha) * rawKeypoints[i].y;
+        }
+      }
+
+      this.smoothedKeypoints = rawKeypoints;
+      return rawKeypoints;
+
+    } else if (this._lastBodyState && (hasCsi || this._ghostConfidence > 0.05)) {
+      // Through-wall mode: person left frame but CSI still senses them
+      return this._trackThroughWall(elapsed, csiState);
+
+    } else if (this.smoothedKeypoints) {
+      // Fade out
+      const faded = this.smoothedKeypoints.map(kp => ({
+        ...kp,
+        confidence: kp.confidence * 0.88
+      })).filter(kp => kp.confidence > 0.05);
+      if (faded.length === 0) this.smoothedKeypoints = null;
+      else this.smoothedKeypoints = faded;
+      return faded;
+    }
+
+    return [];
+  }
+
+  /**
+   * Track body parts from the motion grid.
+   * Finds the centroid of motion in each body zone and positions joints there.
+   */
+  _trackFromMotionGrid(region, embedding, elapsed) {
+    const grid = region.motionGrid;
+    const cols = region.gridCols || 10;
+    const rows = region.gridRows || 8;
+
+    // Body bounding box (in normalized 0-1 coords)
+    const bx = region.x, by = region.y, bw = region.w, bh = region.h;
+    const cx = bx + bw / 2;
+    const cy = by + bh / 2;
+    const bodyH = Math.max(bh, 0.3);
+    const bodyW = Math.max(bw, 0.15);
+
+    // Find motion centroids per body zone from the grid
+    if (grid) {
+      const zones = this._findZoneCentroids(grid, cols, rows, bx, by, bw, bh);
+      // Smooth with low alpha for responsiveness
+      const a = 0.3; // 30% old, 70% new → responsive
+      this._headCx    = a * this._headCx    + (1 - a) * zones.head.x;
+      this._headCy    = a * this._headCy    + (1 - a) * zones.head.y;
+      this._leftArmCx = a * this._leftArmCx + (1 - a) * zones.leftArm.x;
+      this._leftArmCy = a * this._leftArmCy + (1 - a) * zones.leftArm.y;
+      this._rightArmCx= a * this._rightArmCx+ (1 - a) * zones.rightArm.x;
+      this._rightArmCy= a * this._rightArmCy+ (1 - a) * zones.rightArm.y;
+      this._leftLegCx = a * this._leftLegCx + (1 - a) * zones.leftLeg.x;
+      this._leftLegCy = a * this._leftLegCy + (1 - a) * zones.leftLeg.y;
+      this._rightLegCx= a * this._rightLegCx+ (1 - a) * zones.rightLeg.x;
+      this._rightLegCy= a * this._rightLegCy+ (1 - a) * zones.rightLeg.y;
+      this._torsoCx   = a * this._torsoCx   + (1 - a) * zones.torso.x;
+      this._torsoCy   = a * this._torsoCy   + (1 - a) * zones.torso.y;
+    }
+
+    const P = PROPORTIONS;
+
+    // Breathing (subtle)
+    const breathe = Math.sin(elapsed * 1.5) * 0.002;
+
+    // === Position joints using tracked centroids ===
+
+    // HEAD: tracked centroid (top zone)
+    const headX = this._headCx;
+    const headY = this._headCy;
+
+    // TORSO center drives shoulder/hip
+    const torsoX = this._torsoCx;
+    const shoulderY = this._torsoCy - bodyH * 0.08 + breathe;
+    const halfW = P.shoulderWidth * bodyH / 2;
+    const hipHalfW = P.hipWidth * bodyH / 2;
+    const hipY = shoulderY + P.shoulderToHip * bodyH;
+
+    // ARMS: elbow + wrist driven toward arm zone centroids
+    // Left arm: shoulder is fixed, elbow/wrist pulled toward left arm centroid
+    const lShX = torsoX - halfW;
+    const lShY = shoulderY;
+    // Vector from shoulder toward arm centroid
+    const lArmDx = this._leftArmCx - lShX;
+    const lArmDy = this._leftArmCy - lShY;
+    const lArmDist = Math.sqrt(lArmDx * lArmDx + lArmDy * lArmDy) || 0.01;
+    const lArmNx = lArmDx / lArmDist;
+    const lArmNy = lArmDy / lArmDist;
+    // Elbow at shoulderToElbow distance along that direction
+    const elbowLen = P.shoulderToElbow * bodyH;
+    const lElbowX = lShX + lArmNx * elbowLen;
+    const lElbowY = lShY + lArmNy * elbowLen;
+    // Wrist continues further
+    const wristLen = P.elbowToWrist * bodyH;
+    const lWristX = lElbowX + lArmNx * wristLen;
+    const lWristY = lElbowY + lArmNy * wristLen;
+
+    // Right arm: same approach
+    const rShX = torsoX + halfW;
+    const rShY = shoulderY;
+    const rArmDx = this._rightArmCx - rShX;
+    const rArmDy = this._rightArmCy - rShY;
+    const rArmDist = Math.sqrt(rArmDx * rArmDx + rArmDy * rArmDy) || 0.01;
+    const rArmNx = rArmDx / rArmDist;
+    const rArmNy = rArmDy / rArmDist;
+    const rElbowX = rShX + rArmNx * elbowLen;
+    const rElbowY = rShY + rArmNy * elbowLen;
+    const rWristX = rElbowX + rArmNx * wristLen;
+    const rWristY = rElbowY + rArmNy * wristLen;
+
+    // LEGS: knees/ankles pulled toward leg zone centroids
+    const lHipX = torsoX - hipHalfW;
+    const rHipX = torsoX + hipHalfW;
+    const lLegDx = this._leftLegCx - lHipX;
+    const lLegDy = Math.max(0.05, this._leftLegCy - hipY); // always downward
+    const lLegDist = Math.sqrt(lLegDx * lLegDx + lLegDy * lLegDy) || 0.01;
+    const lLegNx = lLegDx / lLegDist;
+    const lLegNy = lLegDy / lLegDist;
+    const kneeLen = P.hipToKnee * bodyH;
+    const ankleLen = P.kneeToAnkle * bodyH;
+    const lKneeX = lHipX + lLegNx * kneeLen;
+    const lKneeY = hipY + lLegNy * kneeLen;
+    const lAnkleX = lKneeX + lLegNx * ankleLen;
+    const lAnkleY = lKneeY + lLegNy * ankleLen;
+
+    const rLegDx = this._rightLegCx - rHipX;
+    const rLegDy = Math.max(0.05, this._rightLegCy - hipY);
+    const rLegDist = Math.sqrt(rLegDx * rLegDx + rLegDy * rLegDy) || 0.01;
+    const rLegNx = rLegDx / rLegDist;
+    const rLegNy = rLegDy / rLegDist;
+    const rKneeX = rHipX + rLegNx * kneeLen;
+    const rKneeY = hipY + rLegNy * kneeLen;
+    const rAnkleX = rKneeX + rLegNx * ankleLen;
+    const rAnkleY = rKneeY + rLegNy * ankleLen;
+
+    // Arm raise amount (for hand openness)
+    const leftArmRaise = Math.max(0, Math.min(1, (shoulderY - this._leftArmCy) / (bodyH * 0.3)));
+    const rightArmRaise = Math.max(0, Math.min(1, (shoulderY - this._rightArmCy) / (bodyH * 0.3)));
+
+    // Compute hand finger positions from wrist-elbow axis
+    const lHandAngle = Math.atan2(lWristY - lElbowY, lWristX - lElbowX);
+    const rHandAngle = Math.atan2(rWristY - rElbowY, rWristX - rElbowX);
+    const fingerLen = P.wristToFinger * bodyH;
+    const fingerSpr = P.fingerSpread * bodyH;
+
+    // Hand openness driven by arm raise + arm lateral spread
+    const lArmSpread = Math.abs(this._leftArmCx - (bx + bw * 0.3)) / (bw * 0.3);
+    const rArmSpread = Math.abs(this._rightArmCx - (bx + bw * 0.7)) / (bw * 0.3);
+    const lHandOpen = Math.min(1, leftArmRaise * 0.5 + lArmSpread * 0.5);
+    const rHandOpen = Math.min(1, rightArmRaise * 0.5 + rArmSpread * 0.5);
+
+    const keypoints = [
+      // 0: nose
+      { x: headX, y: headY + 0.01, confidence: 0.92 },
+      // 1: left_eye
+      { x: headX - P.eyeSpacing * bodyH, y: headY - 0.005, confidence: 0.88 },
+      // 2: right_eye
+      { x: headX + P.eyeSpacing * bodyH, y: headY - 0.005, confidence: 0.88 },
+      // 3: left_ear
+      { x: headX - P.earSpacing * bodyH, y: headY + 0.005, confidence: 0.72 },
+      // 4: right_ear
+      { x: headX + P.earSpacing * bodyH, y: headY + 0.005, confidence: 0.72 },
+      // 5: left_shoulder
+      { x: lShX, y: lShY, confidence: 0.94 },
+      // 6: right_shoulder
+      { x: rShX, y: rShY, confidence: 0.94 },
+      // 7: left_elbow
+      { x: lElbowX, y: lElbowY, confidence: 0.87 },
+      // 8: right_elbow
+      { x: rElbowX, y: rElbowY, confidence: 0.87 },
+      // 9: left_wrist
+      { x: lWristX, y: lWristY, confidence: 0.82 },
+      // 10: right_wrist
+      { x: rWristX, y: rWristY, confidence: 0.82 },
+      // 11: left_hip
+      { x: lHipX, y: hipY, confidence: 0.91 },
+      // 12: right_hip
+      { x: rHipX, y: hipY, confidence: 0.91 },
+      // 13: left_knee
+      { x: lKneeX, y: lKneeY, confidence: 0.88 },
+      // 14: right_knee
+      { x: rKneeX, y: rKneeY, confidence: 0.88 },
+      // 15: left_ankle
+      { x: lAnkleX, y: lAnkleY, confidence: 0.83 },
+      // 16: right_ankle
+      { x: rAnkleX, y: rAnkleY, confidence: 0.83 },
+
+      // === Extended keypoints (17-25) ===
+
+      // 17: left_thumb — offset at thumb angle from wrist-elbow axis
+      { x: lWristX + fingerLen * Math.cos(lHandAngle + P.thumbAngle) * (0.6 + lHandOpen * 0.4),
+        y: lWristY + fingerLen * Math.sin(lHandAngle + P.thumbAngle) * (0.6 + lHandOpen * 0.4),
+        confidence: 0.68 * (0.5 + lHandOpen * 0.5) },
+      // 18: left_index — extends along wrist-elbow axis
+      { x: lWristX + fingerLen * Math.cos(lHandAngle) + fingerSpr * lHandOpen * Math.cos(lHandAngle + 0.3),
+        y: lWristY + fingerLen * Math.sin(lHandAngle) + fingerSpr * lHandOpen * Math.sin(lHandAngle + 0.3),
+        confidence: 0.72 * (0.5 + lHandOpen * 0.5) },
+      // 19: left_pinky — offset opposite thumb
+      { x: lWristX + fingerLen * 0.85 * Math.cos(lHandAngle - P.thumbAngle * 0.7),
+        y: lWristY + fingerLen * 0.85 * Math.sin(lHandAngle - P.thumbAngle * 0.7),
+        confidence: 0.60 * (0.5 + lHandOpen * 0.5) },
+
+      // 20: right_thumb
+      { x: rWristX + fingerLen * Math.cos(rHandAngle - P.thumbAngle) * (0.6 + rHandOpen * 0.4),
+        y: rWristY + fingerLen * Math.sin(rHandAngle - P.thumbAngle) * (0.6 + rHandOpen * 0.4),
+        confidence: 0.68 * (0.5 + rHandOpen * 0.5) },
+      // 21: right_index
+      { x: rWristX + fingerLen * Math.cos(rHandAngle) + fingerSpr * rHandOpen * Math.cos(rHandAngle - 0.3),
+        y: rWristY + fingerLen * Math.sin(rHandAngle) + fingerSpr * rHandOpen * Math.sin(rHandAngle - 0.3),
+        confidence: 0.72 * (0.5 + rHandOpen * 0.5) },
+      // 22: right_pinky
+      { x: rWristX + fingerLen * 0.85 * Math.cos(rHandAngle + P.thumbAngle * 0.7),
+        y: rWristY + fingerLen * 0.85 * Math.sin(rHandAngle + P.thumbAngle * 0.7),
+        confidence: 0.60 * (0.5 + rHandOpen * 0.5) },
+
+      // 23: left_foot_index (toe tip) — extends forward from ankle
+      { x: lAnkleX + P.ankleToToe * bodyH * 0.5,
+        y: lAnkleY + P.ankleToToe * bodyH * 0.3,
+        confidence: 0.65 },
+      // 24: right_foot_index
+      { x: rAnkleX + P.ankleToToe * bodyH * 0.5,
+        y: rAnkleY + P.ankleToToe * bodyH * 0.3,
+        confidence: 0.65 },
+
+      // 25: neck (midpoint between shoulders, slightly above)
+      { x: (lShX + rShX) / 2, y: shoulderY - P.headToShoulder * bodyH * 0.35, confidence: 0.93 },
+    ];
+
+    for (let i = 0; i < keypoints.length; i++) {
+      keypoints[i].name = KEYPOINT_NAMES[i];
+    }
+
+    // === RuVector Attention Embedding Refinement ===
+    // Compute attention stats for the UI pipeline display, but only apply
+    // positional refinement when a trained model is loaded (random-weight
+    // embeddings carry no meaningful spatial signal and distort the skeleton).
+    if (embedding && embedding.length >= 26 * 3) {
+      this._computeEmbeddingStats(keypoints, embedding, bodyH);
+    }
+
+    return keypoints;
+  }
+
+  /**
+   * Apply RuVector attention embedding to refine joint positions and confidence.
+   *
+   * The 128-dim fused embedding is decoded as:
+   * - Dims 0-77:  Per-joint (dx, dy, confidence_mod) × 26 joints
+   * - Dims 78-81: Global pose parameters (scale, rotation, lean)
+   * - Dims 82-127: Reserved for cross-modal fusion features
+   *
+   * The attention mechanism determines HOW MUCH each spatial region contributes
+   * to each joint's refinement. Multi-Head captures global relationships,
+   * Hyperbolic captures hierarchical (torso→limb→hand) dependencies,
+   * MoE routes different body regions to specialized experts,
+   * Linear provides fast extremity refinement, Local-Global balances detail/context.
+   */
+  /**
+   * Compute embedding statistics for UI display without modifying joint positions.
+   * The 6-stage attention pipeline stats are shown in the RuVector panel.
+   * Position refinement is disabled until a trained model replaces random weights.
+   */
+  _computeEmbeddingStats(keypoints, emb, bodyH) {
+    const map = this._jointEmbMap;
+    const tc = (v) => Math.tanh(Number(v) || 0);
+
+    // Embedding energy (L2 norm of the used dims)
+    let energy = 0;
+    for (let i = 0; i < Math.min(emb.length, 82); i++) {
+      energy += emb[i] * emb[i];
+    }
+    energy = Math.sqrt(energy);
+
+    // Simulated per-joint refinement magnitude (what WOULD be applied)
+    const scale = bodyH * 0.015;
+    let totalRefinement = 0;
+    let maxDimVal = 0;
+
+    for (let j = 0; j < Math.min(keypoints.length, 26); j++) {
+      const jmap = map.joints[j];
+      if (!jmap) continue;
+      const dx = tc(emb[jmap.dxDim]) * scale;
+      const dy = tc(emb[jmap.dyDim]) * scale;
+      totalRefinement += Math.sqrt(dx * dx + dy * dy);
+      maxDimVal = Math.max(maxDimVal, Math.abs(tc(emb[jmap.dxDim])), Math.abs(tc(emb[jmap.dyDim])));
+    }
+
+    this.attentionStats.energy = energy;
+    this.attentionStats.maxDim = maxDimVal;
+    this.attentionStats.refinementMag = totalRefinement / 26;
+  }
+
+  /**
+   * Find weighted motion centroids for each body zone.
+   * Divides the bounding box into 6 zones: head, left arm, right arm, torso, left leg, right leg.
+   * Returns the (x,y) centroid of motion intensity for each zone.
+   */
+  _findZoneCentroids(grid, cols, rows, bx, by, bw, bh) {
+    // Zone definitions (in grid-relative fractions)
+    const zones = {
+      head:     { rMin: 0,    rMax: 0.2,  cMin: 0.25, cMax: 0.75, wx: 0, wy: 0, wt: 0 },
+      leftArm:  { rMin: 0.1,  rMax: 0.6,  cMin: 0,    cMax: 0.35, wx: 0, wy: 0, wt: 0 },
+      rightArm: { rMin: 0.1,  rMax: 0.6,  cMin: 0.65, cMax: 1.0,  wx: 0, wy: 0, wt: 0 },
+      torso:    { rMin: 0.15, rMax: 0.55, cMin: 0.3,  cMax: 0.7,  wx: 0, wy: 0, wt: 0 },
+      leftLeg:  { rMin: 0.5,  rMax: 1.0,  cMin: 0.1,  cMax: 0.5,  wx: 0, wy: 0, wt: 0 },
+      rightLeg: { rMin: 0.5,  rMax: 1.0,  cMin: 0.5,  cMax: 0.9,  wx: 0, wy: 0, wt: 0 },
+    };
+
+    // Accumulate weighted centroids per zone
+    for (let r = 0; r < rows; r++) {
+      const ry = r / rows; // 0-1 within grid
+      for (let c = 0; c < cols; c++) {
+        const cx_g = c / cols; // 0-1 within grid
+        const val = grid[r][c];
+        if (val < 0.005) continue; // skip near-zero motion
+
+        // Map grid position to body-space coordinates (0-1)
+        const worldX = bx + cx_g * bw;
+        const worldY = by + ry * bh;
+
+        // Assign to matching zones (a cell can contribute to multiple overlapping zones)
+        for (const z of Object.values(zones)) {
+          if (ry >= z.rMin && ry < z.rMax && cx_g >= z.cMin && cx_g < z.cMax) {
+            z.wx += worldX * val;
+            z.wy += worldY * val;
+            z.wt += val;
+          }
+        }
+      }
+    }
+
+    // Compute centroids with fallback defaults
+    const centroid = (z, defX, defY) => ({
+      x: z.wt > 0.01 ? z.wx / z.wt : defX,
+      y: z.wt > 0.01 ? z.wy / z.wt : defY,
+      weight: z.wt
+    });
+
+    const midX = bx + bw / 2;
+    const midY = by + bh / 2;
+
+    return {
+      head:     centroid(zones.head,     midX,           by + bh * 0.1),
+      leftArm:  centroid(zones.leftArm,  bx + bw * 0.2, midY - bh * 0.05),
+      rightArm: centroid(zones.rightArm, bx + bw * 0.8, midY - bh * 0.05),
+      torso:    centroid(zones.torso,    midX,           midY),
+      leftLeg:  centroid(zones.leftLeg,  bx + bw * 0.35,by + bh * 0.75),
+      rightLeg: centroid(zones.rightLeg, bx + bw * 0.65,by + bh * 0.75),
+    };
+  }
+
+  /**
+   * Through-wall tracking: continue showing pose via CSI when person left video frame.
+   * The skeleton drifts in the exit direction with decreasing confidence.
+   */
+  _trackThroughWall(elapsed, csiState) {
+    if (!this._lastBodyState) return [];
+
+    const dt = elapsed - this._lastBodyState.time;
+    const csiPresence = csiState.csiPresence || 0;
+
+    // Initialize ghost on first call
+    if (this._ghostConfidence <= 0.05) {
+      this._ghostConfidence = 0.8;
+      this._ghostState = this._lastBodyState.keypoints.map(kp => ({...kp}));
+    }
+
+    // Ghost confidence decays, but CSI presence sustains it
+    const csiBoost = Math.min(0.7, csiPresence * 0.8);
+    this._ghostConfidence = Math.max(0.05, this._ghostConfidence * 0.995 - 0.001 + csiBoost * 0.002);
+
+    // Drift the ghost in exit direction
+    const vx = this._ghostVelocity.x;
+    const vy = this._ghostVelocity.y;
+
+    // Breathing continues via CSI
+    const breathe = Math.sin(elapsed * 1.5) * 0.003 * csiPresence;
+
+    const keypoints = this._ghostState.map((kp, i) => {
+      return {
+        x: kp.x + vx * dt * 0.3,
+        y: kp.y + vy * dt * 0.3 + (i >= 5 && i <= 6 ? breathe : 0),
+        confidence: kp.confidence * this._ghostConfidence * (0.5 + csiPresence * 0.5),
+        name: kp.name
+      };
+    });
+
+    // Slow down drift over time
+    this._ghostVelocity.x *= 0.998;
+    this._ghostVelocity.y *= 0.998;
+
+    this.smoothedKeypoints = keypoints;
+    return keypoints;
+  }
+}

ui/pose-fusion/js/video-capture.js

@@ -0,0 +1,235 @@
+/**
+ * VideoCapture — getUserMedia webcam capture with frame extraction.
+ * Provides quality metrics (brightness, motion) for fusion confidence gating.
+ */
+
+export class VideoCapture {
+  constructor(videoElement) {
+    this.video = videoElement;
+    this.stream = null;
+    this.offscreen = document.createElement('canvas');
+    this.offCtx = this.offscreen.getContext('2d', { willReadFrequently: true });
+    this.prevFrame = null;
+    this.motionScore = 0;
+    this.brightnessScore = 0;
+  }
+
+  async start(constraints = {}) {
+    const defaultConstraints = {
+      video: {
+        width: { ideal: 640 },
+        height: { ideal: 480 },
+        facingMode: 'user',
+        frameRate: { ideal: 30 }
+      },
+      audio: false
+    };
+
+    try {
+      this.stream = await navigator.mediaDevices.getUserMedia(
+        Object.keys(constraints).length ? constraints : defaultConstraints
+      );
+      this.video.srcObject = this.stream;
+      await this.video.play();
+
+      this.offscreen.width = this.video.videoWidth;
+      this.offscreen.height = this.video.videoHeight;
+
+      return true;
+    } catch (err) {
+      console.error('[Video] Camera access failed:', err.message);
+      return false;
+    }
+  }
+
+  stop() {
+    if (this.stream) {
+      this.stream.getTracks().forEach(t => t.stop());
+      this.stream = null;
+    }
+    this.video.srcObject = null;
+  }
+
+  get isActive() {
+    return this.stream !== null && this.video.readyState >= 2;
+  }
+
+  get width() { return this.video.videoWidth || 640; }
+  get height() { return this.video.videoHeight || 480; }
+
+  /**
+   * Capture current frame as RGB Uint8Array + compute quality metrics.
+   * @param {number} targetW - Target width for CNN input
+   * @param {number} targetH - Target height for CNN input
+   * @returns {{ rgb: Uint8Array, width: number, height: number, motion: number, brightness: number }}
+   */
+  captureFrame(targetW = 56, targetH = 56) {
+    if (!this.isActive) return null;
+
+    // Draw to offscreen at target resolution
+    this.offscreen.width = targetW;
+    this.offscreen.height = targetH;
+    this.offCtx.drawImage(this.video, 0, 0, targetW, targetH);
+    const imageData = this.offCtx.getImageData(0, 0, targetW, targetH);
+    const rgba = imageData.data;
+
+    // Convert RGBA → RGB
+    const pixels = targetW * targetH;
+    const rgb = new Uint8Array(pixels * 3);
+    let brightnessSum = 0;
+    let motionSum = 0;
+
+    for (let i = 0; i < pixels; i++) {
+      const r = rgba[i * 4];
+      const g = rgba[i * 4 + 1];
+      const b = rgba[i * 4 + 2];
+      rgb[i * 3] = r;
+      rgb[i * 3 + 1] = g;
+      rgb[i * 3 + 2] = b;
+
+      // Luminance for brightness
+      const lum = 0.299 * r + 0.587 * g + 0.114 * b;
+      brightnessSum += lum;
+
+      // Motion: diff from previous frame
+      if (this.prevFrame) {
+        const pr = this.prevFrame[i * 3];
+        const pg = this.prevFrame[i * 3 + 1];
+        const pb = this.prevFrame[i * 3 + 2];
+        motionSum += Math.abs(r - pr) + Math.abs(g - pg) + Math.abs(b - pb);
+      }
+    }
+
+    this.brightnessScore = brightnessSum / (pixels * 255);
+    this.motionScore = this.prevFrame ? Math.min(1, motionSum / (pixels * 100)) : 0;
+    this.prevFrame = new Uint8Array(rgb);
+
+    return {
+      rgb,
+      width: targetW,
+      height: targetH,
+      motion: this.motionScore,
+      brightness: this.brightnessScore
+    };
+  }
+
+  /**
+   * Capture full-resolution RGBA for overlay rendering
+   * @returns {ImageData|null}
+   */
+  captureFullFrame() {
+    if (!this.isActive) return null;
+    this.offscreen.width = this.width;
+    this.offscreen.height = this.height;
+    this.offCtx.drawImage(this.video, 0, 0);
+    return this.offCtx.getImageData(0, 0, this.width, this.height);
+  }
+
+  /**
+   * Detect motion region + detailed motion grid for body-part tracking.
+   * Returns bounding box + a grid showing WHERE motion is concentrated.
+   * @returns {{ x, y, w, h, detected: boolean, motionGrid: number[][], gridCols: number, gridRows: number, exitDirection: string|null }}
+   */
+  detectMotionRegion(targetW = 56, targetH = 56) {
+    if (!this.isActive || !this.prevFrame) return { detected: false, motionGrid: null };
+
+    this.offscreen.width = targetW;
+    this.offscreen.height = targetH;
+    this.offCtx.drawImage(this.video, 0, 0, targetW, targetH);
+    const rgba = this.offCtx.getImageData(0, 0, targetW, targetH).data;
+
+    let minX = targetW, minY = targetH, maxX = 0, maxY = 0;
+    let motionPixels = 0;
+    const threshold = 25;
+
+    // Motion grid: divide frame into cells and track motion intensity per cell
+    const gridCols = 10;
+    const gridRows = 8;
+    const cellW = targetW / gridCols;
+    const cellH = targetH / gridRows;
+    const motionGrid = Array.from({ length: gridRows }, () => new Float32Array(gridCols));
+    const cellPixels = cellW * cellH;
+
+    // Also track motion centroid weighted by intensity
+    let motionCxSum = 0, motionCySum = 0, motionWeightSum = 0;
+
+    for (let y = 0; y < targetH; y++) {
+      for (let x = 0; x < targetW; x++) {
+        const i = y * targetW + x;
+        const r = rgba[i * 4], g = rgba[i * 4 + 1], b = rgba[i * 4 + 2];
+        const pr = this.prevFrame[i * 3], pg = this.prevFrame[i * 3 + 1], pb = this.prevFrame[i * 3 + 2];
+        const diff = Math.abs(r - pr) + Math.abs(g - pg) + Math.abs(b - pb);
+
+        if (diff > threshold * 3) {
+          motionPixels++;
+          if (x < minX) minX = x;
+          if (y < minY) minY = y;
+          if (x > maxX) maxX = x;
+          if (y > maxY) maxY = y;
+        }
+
+        // Accumulate per-cell motion intensity
+        const gc = Math.min(Math.floor(x / cellW), gridCols - 1);
+        const gr = Math.min(Math.floor(y / cellH), gridRows - 1);
+        const intensity = diff / (3 * 255); // Normalize 0-1
+        motionGrid[gr][gc] += intensity / cellPixels;
+
+        // Weighted centroid
+        if (diff > threshold) {
+          motionCxSum += x * diff;
+          motionCySum += y * diff;
+          motionWeightSum += diff;
+        }
+      }
+    }
+
+    const detected = motionPixels > (targetW * targetH * 0.02);
+
+    // Motion centroid (normalized 0-1)
+    const motionCx = motionWeightSum > 0 ? motionCxSum / (motionWeightSum * targetW) : 0.5;
+    const motionCy = motionWeightSum > 0 ? motionCySum / (motionWeightSum * targetH) : 0.5;
+
+    // Detect exit direction: if centroid is near edges
+    let exitDirection = null;
+    if (detected && motionCx < 0.1) exitDirection = 'left';
+    else if (detected && motionCx > 0.9) exitDirection = 'right';
+    else if (detected && motionCy < 0.1) exitDirection = 'up';
+    else if (detected && motionCy > 0.9) exitDirection = 'down';
+
+    // Track last known position for through-wall persistence
+    if (detected) {
+      this._lastDetected = {
+        x: minX / targetW,
+        y: minY / targetH,
+        w: (maxX - minX) / targetW,
+        h: (maxY - minY) / targetH,
+        cx: motionCx,
+        cy: motionCy,
+        exitDirection,
+        time: performance.now()
+      };
+    }
+
+    return {
+      detected,
+      x: minX / targetW,
+      y: minY / targetH,
+      w: (maxX - minX) / targetW,
+      h: (maxY - minY) / targetH,
+      coverage: motionPixels / (targetW * targetH),
+      motionGrid,
+      gridCols,
+      gridRows,
+      motionCx,
+      motionCy,
+      exitDirection
+    };
+  }
+
+  /**
+   * Get the last known detection info (for through-wall persistence)
+   */
+  get lastDetection() {
+    return this._lastDetected || null;
+  }
+}

ui/pose-fusion/pkg/ruvector-attention/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 rUv
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

ui/pose-fusion/pkg/ruvector-attention/README.md

@@ -0,0 +1,220 @@
+# ruvector-attention-wasm
+
+WebAssembly bindings for the ruvector-attention package, providing high-performance attention mechanisms for browser and Node.js environments.
+
+## Features
+
+- **Multiple Attention Mechanisms**:
+  - Scaled Dot-Product Attention
+  - Multi-Head Attention
+  - Hyperbolic Attention (for hierarchical data)
+  - Linear Attention (Performer-style)
+  - Flash Attention (memory-efficient)
+  - Local-Global Attention
+  - Mixture of Experts (MoE) Attention
+  - **CGT Sheaf Attention** (coherence-gated via Prime-Radiant)
+
+- **Training Utilities**:
+  - InfoNCE contrastive loss
+  - Adam optimizer
+  - AdamW optimizer (with decoupled weight decay)
+  - Learning rate scheduler (warmup + cosine decay)
+
+- **TypeScript Support**: Full type definitions and modern API
+
+## Installation
+
+```bash
+npm install ruvector-attention-wasm
+```
+
+## Usage
+
+### TypeScript/JavaScript
+
+```typescript
+import { initialize, MultiHeadAttention, utils } from 'ruvector-attention-wasm';
+
+// Initialize WASM module
+await initialize();
+
+// Create multi-head attention
+const attention = new MultiHeadAttention({ dim: 64, numHeads: 8 });
+
+// Prepare inputs
+const query = new Float32Array(64);
+const keys = [new Float32Array(64), new Float32Array(64)];
+const values = [new Float32Array(64), new Float32Array(64)];
+
+// Compute attention
+const output = attention.compute(query, keys, values);
+
+// Use utilities
+const similarity = utils.cosineSimilarity(query, keys[0]);
+```
+
+### Advanced Examples
+
+#### Hyperbolic Attention
+
+```typescript
+import { HyperbolicAttention } from 'ruvector-attention-wasm';
+
+const hyperbolic = new HyperbolicAttention({
+  dim: 128,
+  curvature: 1.0
+});
+
+const output = hyperbolic.compute(query, keys, values);
+```
+
+#### MoE Attention with Expert Stats
+
+```typescript
+import { MoEAttention } from 'ruvector-attention-wasm';
+
+const moe = new MoEAttention({
+  dim: 64,
+  numExperts: 4,
+  topK: 2
+});
+
+const output = moe.compute(query, keys, values);
+
+// Get expert utilization
+const stats = moe.getExpertStats();
+console.log('Load balance:', stats.loadBalance);
+```
+
+#### Training with InfoNCE Loss
+
+```typescript
+import { InfoNCELoss, Adam } from 'ruvector-attention-wasm';
+
+const loss = new InfoNCELoss(0.07);
+const optimizer = new Adam(paramCount, {
+  learningRate: 0.001,
+  beta1: 0.9,
+  beta2: 0.999,
+});
+
+// Training loop
+const lossValue = loss.compute(anchor, positive, negatives);
+optimizer.step(params, gradients);
+```
+
+#### Learning Rate Scheduling
+
+```typescript
+import { LRScheduler, AdamW } from 'ruvector-attention-wasm';
+
+const scheduler = new LRScheduler({
+  initialLR: 0.001,
+  warmupSteps: 1000,
+  totalSteps: 10000,
+});
+
+const optimizer = new AdamW(paramCount, {
+  learningRate: scheduler.getLR(),
+  weightDecay: 0.01,
+});
+
+// Training loop
+for (let step = 0; step < 10000; step++) {
+  optimizer.learningRate = scheduler.getLR();
+  optimizer.step(params, gradients);
+  scheduler.step();
+}
+```
+
+## Building from Source
+
+### Prerequisites
+
+- Rust 1.70+
+- wasm-pack
+
+### Build Commands
+
+```bash
+# Build for web (ES modules)
+wasm-pack build --target web --out-dir pkg
+
+# Build for Node.js
+wasm-pack build --target nodejs --out-dir pkg-node
+
+# Build for bundlers (webpack, vite, etc.)
+wasm-pack build --target bundler --out-dir pkg-bundler
+
+# Run tests
+wasm-pack test --headless --firefox
+```
+
+## API Reference
+
+### Attention Mechanisms
+
+- `MultiHeadAttention` - Standard multi-head attention
+- `HyperbolicAttention` - Attention in hyperbolic space
+- `LinearAttention` - Linear complexity attention (Performer)
+- `FlashAttention` - Memory-efficient attention
+- `LocalGlobalAttention` - Combined local and global attention
+- `MoEAttention` - Mixture of Experts attention
+- `CGTSheafAttention` - Coherence-gated via Prime-Radiant energy
+- `scaledDotAttention()` - Functional API for basic attention
+
+### CGT Sheaf Attention (Prime-Radiant Integration)
+
+The CGT (Coherence-Gated Transformer) Sheaf Attention mechanism uses Prime-Radiant's sheaf Laplacian energy to gate attention based on mathematical consistency:
+
+```typescript
+import { CGTSheafAttention } from 'ruvector-attention-wasm';
+
+const cgtAttention = new CGTSheafAttention({
+  dim: 128,
+  numHeads: 8,
+  coherenceThreshold: 0.3,  // Block if energy > threshold
+});
+
+// Attention is gated by coherence energy
+const result = cgtAttention.compute(query, keys, values);
+console.log('Coherence energy:', result.energy);
+console.log('Is coherent:', result.isCoherent);
+```
+
+**Key features:**
+- Energy-weighted attention: Lower coherence energy → higher attention
+- Automatic hallucination detection via residual analysis
+- GPU-accelerated with wgpu WGSL shaders (vec4 optimized)
+- SIMD fallback (AVX-512/AVX2/NEON)
+
+### Training
+
+- `InfoNCELoss` - Contrastive loss function
+- `Adam` - Adam optimizer
+- `AdamW` - AdamW optimizer with weight decay
+- `LRScheduler` - Learning rate scheduler
+
+### Utilities
+
+- `utils.cosineSimilarity()` - Cosine similarity between vectors
+- `utils.l2Norm()` - L2 norm of a vector
+- `utils.normalize()` - Normalize vector to unit length
+- `utils.softmax()` - Apply softmax transformation
+- `utils.attentionWeights()` - Compute attention weights from scores
+- `utils.batchNormalize()` - Batch normalization
+- `utils.randomOrthogonalMatrix()` - Generate random orthogonal matrix
+- `utils.pairwiseDistances()` - Compute pairwise distances
+
+## Performance
+
+The WASM bindings provide near-native performance for attention computations:
+
+- Optimized with `opt-level = "s"` and LTO
+- SIMD acceleration where available
+- Efficient memory management
+- Zero-copy data transfer where possible
+
+## License
+
+MIT OR Apache-2.0

ui/pose-fusion/pkg/ruvector-attention/package.json

@@ -0,0 +1,28 @@
+{
+  "name": "ruvector-attention-wasm",
+  "collaborators": [
+    "Ruvector Team"
+  ],
+  "description": "High-performance WebAssembly attention mechanisms: Multi-Head, Flash, Hyperbolic, MoE, CGT Sheaf Attention with GPU acceleration for transformers and LLMs",
+  "version": "2.0.5",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/ruvnet/ruvector"
+  },
+  "files": [
+    "ruvector_attention_wasm_bg.wasm",
+    "ruvector_attention_wasm.js",
+    "ruvector_attention_wasm.d.ts"
+  ],
+  "main": "ruvector_attention_wasm.js",
+  "homepage": "https://ruv.io/ruvector",
+  "types": "ruvector_attention_wasm.d.ts",
+  "keywords": [
+    "wasm",
+    "attention",
+    "transformer",
+    "flash-attention",
+    "llm"
+  ]
+}

ui/pose-fusion/pkg/ruvector-attention/ruvector_attention_browser.js

@@ -0,0 +1,642 @@
+/**
+ * Browser ESM wrapper for ruvector-attention-wasm v2.0.5
+ *
+ * The upstream pkg/ was built with wasm-pack --target nodejs (CJS + fs.readFileSync).
+ * This wrapper loads the same WASM binary via fetch() for browser use.
+ *
+ * Usage:
+ *   import initWasm, { WasmMultiHeadAttention, ... } from './ruvector_attention_browser.js';
+ *   await initWasm();
+ *   const attn = new WasmMultiHeadAttention(dim, heads);
+ */
+
+let _wasm;
+let _initialized = false;
+
+// The entire CJS module runs inside this IIFE to avoid polluting global scope.
+// We capture all exports in _mod.
+const _mod = {};
+
+(function(exports, wasm_getter) {
+
+// ── wasm-bindgen heap management ──────────────────────────────────
+const heap = new Array(128).fill(undefined);
+heap.push(undefined, null, true, false);
+let heap_next = heap.length;
+
+function addHeapObject(obj) {
+  if (heap_next === heap.length) heap.push(heap.length + 1);
+  const idx = heap_next;
+  heap_next = heap[idx];
+  heap[idx] = obj;
+  return idx;
+}
+function getObject(idx) { return heap[idx]; }
+function dropObject(idx) {
+  if (idx < 132) return;
+  heap[idx] = heap_next;
+  heap_next = idx;
+}
+function takeObject(idx) {
+  const ret = getObject(idx);
+  dropObject(idx);
+  return ret;
+}
+function isLikeNone(x) { return x === undefined || x === null; }
+
+// ── Memory views ──────────────────────────────────────────────────
+let cachedDataViewMemory0 = null;
+let cachedUint8ArrayMemory0 = null;
+let cachedFloat32ArrayMemory0 = null;
+
+function wasm() { return wasm_getter(); }
+
+function getDataViewMemory0() {
+  if (cachedDataViewMemory0 === null || cachedDataViewMemory0.buffer !== wasm().memory.buffer)
+    cachedDataViewMemory0 = new DataView(wasm().memory.buffer);
+  return cachedDataViewMemory0;
+}
+function getUint8ArrayMemory0() {
+  if (cachedUint8ArrayMemory0 === null || cachedUint8ArrayMemory0.buffer !== wasm().memory.buffer)
+    cachedUint8ArrayMemory0 = new Uint8Array(wasm().memory.buffer);
+  return cachedUint8ArrayMemory0;
+}
+function getFloat32ArrayMemory0() {
+  if (cachedFloat32ArrayMemory0 === null || cachedFloat32ArrayMemory0.buffer !== wasm().memory.buffer)
+    cachedFloat32ArrayMemory0 = new Float32Array(wasm().memory.buffer);
+  return cachedFloat32ArrayMemory0;
+}
+function getArrayF32FromWasm0(ptr, len) {
+  ptr = ptr >>> 0;
+  return getFloat32ArrayMemory0().subarray(ptr / 4, ptr / 4 + len);
+}
+function getArrayU8FromWasm0(ptr, len) {
+  ptr = ptr >>> 0;
+  return getUint8ArrayMemory0().subarray(ptr, ptr + len);
+}
+
+let WASM_VECTOR_LEN = 0;
+
+function passArrayF32ToWasm0(arg, malloc) {
+  const ptr = malloc(arg.length * 4, 4) >>> 0;
+  getFloat32ArrayMemory0().set(arg, ptr / 4);
+  WASM_VECTOR_LEN = arg.length;
+  return ptr;
+}
+
+const cachedTextEncoder = new TextEncoder();
+const cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+cachedTextDecoder.decode();
+
+function getStringFromWasm0(ptr, len) {
+  ptr = ptr >>> 0;
+  return cachedTextDecoder.decode(getUint8ArrayMemory0().subarray(ptr, ptr + len));
+}
+
+function passStringToWasm0(arg, malloc, realloc) {
+  const buf = cachedTextEncoder.encode(arg);
+  const ptr = malloc(buf.length, 1) >>> 0;
+  getUint8ArrayMemory0().subarray(ptr, ptr + buf.length).set(buf);
+  WASM_VECTOR_LEN = buf.length;
+  return ptr;
+}
+
+function debugString(val) {
+  const type = typeof val;
+  if (type == 'number' || type == 'boolean' || val == null) return `${val}`;
+  if (type == 'string') return `"${val}"`;
+  if (type == 'symbol') return val.description ? `Symbol(${val.description})` : 'Symbol';
+  if (type == 'function') return 'Function';
+  if (Array.isArray(val)) return `[${val.map(debugString).join(', ')}]`;
+  try {
+    const keys = Object.keys(val);
+    return `{${keys.map(k => `${k}: ${debugString(val[k])}`).join(', ')}}`;
+  } catch (_) { return Object.prototype.toString.call(val); }
+}
+
+function handleError(f, args) {
+  try { return f.apply(this, args); }
+  catch (e) { wasm().__wbindgen_export3(addHeapObject(e)); }
+}
+
+// ── FinalizationRegistry ──────────────────────────────────────────
+const FR = typeof FinalizationRegistry !== 'undefined'
+  ? FinalizationRegistry
+  : class { register() {} unregister() {} };
+
+const WasmMultiHeadAttentionFinalization = new FR(ptr => wasm().__wbg_wasmmultiheadattention_free(ptr >>> 0, 1));
+const WasmFlashAttentionFinalization = new FR(ptr => wasm().__wbg_wasmflashattention_free(ptr >>> 0, 1));
+const WasmHyperbolicAttentionFinalization = new FR(ptr => wasm().__wbg_wasmhyperbolicattention_free(ptr >>> 0, 1));
+const WasmMoEAttentionFinalization = new FR(ptr => wasm().__wbg_wasmmoeattention_free(ptr >>> 0, 1));
+const WasmLinearAttentionFinalization = new FR(ptr => wasm().__wbg_wasmlinearattention_free(ptr >>> 0, 1));
+const WasmLocalGlobalAttentionFinalization = new FR(ptr => wasm().__wbg_wasmlocalglobalattention_free(ptr >>> 0, 1));
+
+// ── Classes ───────────────────────────────────────────────────────
+
+class WasmMultiHeadAttention {
+  constructor(dim, num_heads) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      wasm().wasmmultiheadattention_new(retptr, dim, num_heads);
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      if (r2) throw takeObject(r1);
+      this.__wbg_ptr = r0 >>> 0;
+      WasmMultiHeadAttentionFinalization.register(this, this.__wbg_ptr, this);
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmMultiHeadAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmmultiheadattention_free(ptr, 0);
+  }
+  get dim() { return wasm().wasmmultiheadattention_dim(this.__wbg_ptr); }
+  get num_heads() { return wasm().wasmmultiheadattention_num_heads(this.__wbg_ptr); }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmmultiheadattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+class WasmFlashAttention {
+  constructor(dim, block_size) {
+    const ret = wasm().wasmflashattention_new(dim, block_size);
+    this.__wbg_ptr = ret >>> 0;
+    WasmFlashAttentionFinalization.register(this, this.__wbg_ptr, this);
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmFlashAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmflashattention_free(ptr, 0);
+  }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmflashattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+class WasmHyperbolicAttention {
+  constructor(dim, curvature) {
+    const ret = wasm().wasmhyperbolicattention_new(dim, curvature);
+    this.__wbg_ptr = ret >>> 0;
+    WasmHyperbolicAttentionFinalization.register(this, this.__wbg_ptr, this);
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmHyperbolicAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmhyperbolicattention_free(ptr, 0);
+  }
+  get curvature() { return wasm().wasmhyperbolicattention_curvature(this.__wbg_ptr); }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmhyperbolicattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+class WasmMoEAttention {
+  constructor(dim, num_experts, top_k) {
+    const ret = wasm().wasmmoeattention_new(dim, num_experts, top_k);
+    this.__wbg_ptr = ret >>> 0;
+    WasmMoEAttentionFinalization.register(this, this.__wbg_ptr, this);
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmMoEAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmmoeattention_free(ptr, 0);
+  }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmmoeattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+class WasmLinearAttention {
+  constructor(dim, num_features) {
+    const ret = wasm().wasmlinearattention_new(dim, num_features || dim);
+    this.__wbg_ptr = ret >>> 0;
+    WasmLinearAttentionFinalization.register(this, this.__wbg_ptr, this);
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmLinearAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmlinearattention_free(ptr, 0);
+  }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmlinearattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+class WasmLocalGlobalAttention {
+  constructor(dim, local_window, global_tokens) {
+    const ret = wasm().wasmlocalglobalattention_new(dim, local_window || 4, global_tokens || 2);
+    this.__wbg_ptr = ret >>> 0;
+    WasmLocalGlobalAttentionFinalization.register(this, this.__wbg_ptr, this);
+  }
+  free() {
+    const ptr = this.__wbg_ptr; this.__wbg_ptr = 0;
+    WasmLocalGlobalAttentionFinalization.unregister(this);
+    wasm().__wbg_wasmlocalglobalattention_free(ptr, 0);
+  }
+  compute(query, keys, values) {
+    const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+    try {
+      const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+      const len0 = WASM_VECTOR_LEN;
+      wasm().wasmlocalglobalattention_compute(retptr, this.__wbg_ptr, ptr0, len0, addHeapObject(keys), addHeapObject(values));
+      var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+      var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+      var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+      var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+      if (r3) throw takeObject(r2);
+      var v1 = getArrayF32FromWasm0(r0, r1).slice();
+      wasm().__wbindgen_export4(r0, r1 * 4, 4);
+      return v1;
+    } finally {
+      wasm().__wbindgen_add_to_stack_pointer(16);
+    }
+  }
+}
+
+// ── Standalone functions ──────────────────────────────────────────
+
+function cosine_similarity(a, b) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    const ptr0 = passArrayF32ToWasm0(a, wasm().__wbindgen_export);
+    const len0 = WASM_VECTOR_LEN;
+    const ptr1 = passArrayF32ToWasm0(b, wasm().__wbindgen_export);
+    const len1 = WASM_VECTOR_LEN;
+    wasm().cosine_similarity(retptr, ptr0, len0, ptr1, len1);
+    var r0 = getDataViewMemory0().getFloat64(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 8, true);
+    var r2 = getDataViewMemory0().getInt32(retptr + 12, true);
+    if (r2) throw takeObject(r1);
+    return r0;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function normalize(vec) {
+  const ptr0 = passArrayF32ToWasm0(vec, wasm().__wbindgen_export);
+  const len0 = WASM_VECTOR_LEN;
+  wasm().normalize(ptr0, len0, addHeapObject(vec));
+}
+
+function l2_norm(vec) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    const ptr0 = passArrayF32ToWasm0(vec, wasm().__wbindgen_export);
+    const len0 = WASM_VECTOR_LEN;
+    wasm().l2_norm(retptr, ptr0, len0);
+    var r0 = getDataViewMemory0().getFloat64(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 8, true);
+    var r2 = getDataViewMemory0().getInt32(retptr + 12, true);
+    if (r2) throw takeObject(r1);
+    return r0;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function softmax(vec) {
+  const ptr0 = passArrayF32ToWasm0(vec, wasm().__wbindgen_export);
+  const len0 = WASM_VECTOR_LEN;
+  wasm().softmax(ptr0, len0, addHeapObject(vec));
+}
+
+function batch_normalize(vectors, epsilon) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    wasm().batch_normalize(retptr, addHeapObject(vectors), isLikeNone(epsilon) ? 0x100000001 : Math.fround(epsilon));
+    var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+    var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+    var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+    if (r3) throw takeObject(r2);
+    var v1 = getArrayF32FromWasm0(r0, r1).slice();
+    wasm().__wbindgen_export4(r0, r1 * 4, 4);
+    return v1;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function pairwise_distances(vectors) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    wasm().pairwise_distances(retptr, addHeapObject(vectors));
+    var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+    var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+    var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+    if (r3) throw takeObject(r2);
+    var v1 = getArrayF32FromWasm0(r0, r1).slice();
+    wasm().__wbindgen_export4(r0, r1 * 4, 4);
+    return v1;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function scaled_dot_attention(query, keys, values, scale) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    const ptr0 = passArrayF32ToWasm0(query, wasm().__wbindgen_export);
+    const len0 = WASM_VECTOR_LEN;
+    wasm().scaled_dot_attention(retptr, ptr0, len0, addHeapObject(keys), addHeapObject(values), isLikeNone(scale) ? 0x100000001 : Math.fround(scale));
+    var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+    var r2 = getDataViewMemory0().getInt32(retptr + 8, true);
+    var r3 = getDataViewMemory0().getInt32(retptr + 12, true);
+    if (r3) throw takeObject(r2);
+    var v1 = getArrayF32FromWasm0(r0, r1).slice();
+    wasm().__wbindgen_export4(r0, r1 * 4, 4);
+    return v1;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function attention_weights(scores, temperature) {
+  const ptr0 = passArrayF32ToWasm0(scores, wasm().__wbindgen_export);
+  const len0 = WASM_VECTOR_LEN;
+  wasm().attention_weights(ptr0, len0, addHeapObject(scores), isLikeNone(temperature) ? 0x100000001 : Math.fround(temperature));
+}
+
+function available_mechanisms() {
+  const ret = wasm().available_mechanisms();
+  return takeObject(ret);
+}
+
+function random_orthogonal_matrix(dim) {
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    wasm().random_orthogonal_matrix(retptr, dim);
+    var r0 = getDataViewMemory0().getInt32(retptr + 0, true);
+    var r1 = getDataViewMemory0().getInt32(retptr + 4, true);
+    var v1 = getArrayF32FromWasm0(r0, r1).slice();
+    wasm().__wbindgen_export4(r0, r1 * 4, 4);
+    return v1;
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+  }
+}
+
+function rv_init() { wasm().init(); }
+
+function rv_version() {
+  let d0, d1;
+  const retptr = wasm().__wbindgen_add_to_stack_pointer(-16);
+  try {
+    wasm().version(retptr);
+    d0 = getDataViewMemory0().getInt32(retptr + 0, true);
+    d1 = getDataViewMemory0().getInt32(retptr + 4, true);
+    return getStringFromWasm0(d0, d1);
+  } finally {
+    wasm().__wbindgen_add_to_stack_pointer(16);
+    if (d0 !== undefined) wasm().__wbindgen_export4(d0, d1, 1);
+  }
+}
+
+// ── Collect exports ───────────────────────────────────────────────
+exports.WasmMultiHeadAttention = WasmMultiHeadAttention;
+exports.WasmFlashAttention = WasmFlashAttention;
+exports.WasmHyperbolicAttention = WasmHyperbolicAttention;
+exports.WasmMoEAttention = WasmMoEAttention;
+exports.WasmLinearAttention = WasmLinearAttention;
+exports.WasmLocalGlobalAttention = WasmLocalGlobalAttention;
+exports.cosine_similarity = cosine_similarity;
+exports.normalize = normalize;
+exports.l2_norm = l2_norm;
+exports.softmax = softmax;
+exports.batch_normalize = batch_normalize;
+exports.pairwise_distances = pairwise_distances;
+exports.scaled_dot_attention = scaled_dot_attention;
+exports.attention_weights = attention_weights;
+exports.available_mechanisms = available_mechanisms;
+exports.random_orthogonal_matrix = random_orthogonal_matrix;
+exports.init = rv_init;
+exports.version = rv_version;
+
+// ── Build WASM import object ──────────────────────────────────────
+exports.__wbg_get_imports = function() {
+  const import0 = {
+    __proto__: null,
+    __wbg_Error_4577686b3a6d9b3a: (arg0, arg1) => addHeapObject(Error(getStringFromWasm0(arg0, arg1))),
+    __wbg_String_8564e559799eccda: (arg0, arg1) => {
+      const ret = String(getObject(arg1));
+      const ptr1 = passStringToWasm0(ret, wasm().__wbindgen_export, wasm().__wbindgen_export2);
+      const len1 = WASM_VECTOR_LEN;
+      getDataViewMemory0().setInt32(arg0 + 4, len1, true);
+      getDataViewMemory0().setInt32(arg0, ptr1, true);
+    },
+    __wbg___wbindgen_boolean_get_18c4ed9422296fff: (arg0) => {
+      const v = getObject(arg0);
+      const ret = typeof v === 'boolean' ? v : undefined;
+      return isLikeNone(ret) ? 0xFFFFFF : ret ? 1 : 0;
+    },
+    __wbg___wbindgen_copy_to_typed_array_5294f8e46aecc086: (arg0, arg1, arg2) => {
+      new Uint8Array(getObject(arg2).buffer, getObject(arg2).byteOffset, getObject(arg2).byteLength).set(getArrayU8FromWasm0(arg0, arg1));
+    },
+    __wbg___wbindgen_debug_string_ddde1867f49c2442: (arg0, arg1) => {
+      const ret = debugString(getObject(arg1));
+      const ptr1 = passStringToWasm0(ret, wasm().__wbindgen_export, wasm().__wbindgen_export2);
+      const len1 = WASM_VECTOR_LEN;
+      getDataViewMemory0().setInt32(arg0 + 4, len1, true);
+      getDataViewMemory0().setInt32(arg0, ptr1, true);
+    },
+    __wbg___wbindgen_is_function_d633e708baf0d146: (arg0) => typeof getObject(arg0) === 'function',
+    __wbg___wbindgen_is_object_4b3de556756ee8a8: (arg0) => {
+      const val = getObject(arg0);
+      return typeof val === 'object' && val !== null;
+    },
+    __wbg___wbindgen_jsval_loose_eq_1562ceb9af84e990: (arg0, arg1) => getObject(arg0) == getObject(arg1),
+    __wbg___wbindgen_number_get_5854912275df1894: (arg0, arg1) => {
+      const obj = getObject(arg1);
+      const ret = typeof obj === 'number' ? obj : undefined;
+      getDataViewMemory0().setFloat64(arg0 + 8, isLikeNone(ret) ? 0 : ret, true);
+      getDataViewMemory0().setInt32(arg0, !isLikeNone(ret), true);
+    },
+    __wbg___wbindgen_string_get_3e5751597f39a112: (arg0, arg1) => {
+      const obj = getObject(arg1);
+      const ret = typeof obj === 'string' ? obj : undefined;
+      var ptr1 = isLikeNone(ret) ? 0 : passStringToWasm0(ret, wasm().__wbindgen_export, wasm().__wbindgen_export2);
+      var len1 = WASM_VECTOR_LEN;
+      getDataViewMemory0().setInt32(arg0 + 4, len1, true);
+      getDataViewMemory0().setInt32(arg0, ptr1, true);
+    },
+    __wbg___wbindgen_throw_39bc967c0e5a9b58: (arg0, arg1) => { throw new Error(getStringFromWasm0(arg0, arg1)); },
+    __wbg_call_73af281463ec8b58: function() { return handleError(function(arg0, arg1) {
+      return addHeapObject(getObject(arg0).call(getObject(arg1)));
+    }, arguments); },
+    __wbg_done_5aad55ec6b1954b1: (arg0) => getObject(arg0).done,
+    __wbg_error_a6fa202b58aa1cd3: (arg0, arg1) => {
+      try { console.error(getStringFromWasm0(arg0, arg1)); }
+      finally { wasm().__wbindgen_export4(arg0, arg1, 1); }
+    },
+    __wbg_error_ad28debb48b5c6bb: (arg0) => console.error(getObject(arg0)),
+    __wbg_get_4920fefd3451364b: function() { return handleError(function(arg0, arg1) {
+      return addHeapObject(Reflect.get(getObject(arg0), getObject(arg1)));
+    }, arguments); },
+    __wbg_get_unchecked_3d0f4b91c8eca4f0: (arg0, arg1) => addHeapObject(getObject(arg0)[arg1 >>> 0]),
+    __wbg_instanceof_ArrayBuffer_15859862b80b732d: (arg0) => {
+      try { return getObject(arg0) instanceof ArrayBuffer; } catch (_) { return false; }
+    },
+    __wbg_instanceof_Uint8Array_2240b7046ac16f05: (arg0) => {
+      try { return getObject(arg0) instanceof Uint8Array; } catch (_) { return false; }
+    },
+    __wbg_isArray_fad08a0d12828686: (arg0) => Array.isArray(getObject(arg0)),
+    __wbg_iterator_fc7ad8d33bab9e26: () => addHeapObject(Symbol.iterator),
+    __wbg_length_5855c1f289dfffc1: (arg0) => getObject(arg0).length,
+    __wbg_length_a31e05262e09b7f8: (arg0) => getObject(arg0).length,
+    __wbg_log_3c5e4b64af29e724: (arg0) => console.log(getObject(arg0)),
+    __wbg_new_09959f7b4c92c246: (arg0) => addHeapObject(new Uint8Array(getObject(arg0))),
+    __wbg_new_227d7c05414eb861: () => addHeapObject(new Error()),
+    __wbg_new_cbee8c0d5c479eac: () => addHeapObject(new Array()),
+    __wbg_next_a5fe6f328f7affc2: (arg0) => addHeapObject(getObject(arg0).next),
+    __wbg_next_e592122bb4ed4c67: function() { return handleError(function(arg0) {
+      return addHeapObject(getObject(arg0).next());
+    }, arguments); },
+    __wbg_prototypesetcall_f034d444741426c3: (arg0, arg1, arg2) => {
+      Uint8Array.prototype.set.call(getArrayU8FromWasm0(arg0, arg1), getObject(arg2));
+    },
+    __wbg_random_2b7bed8995d680fb: () => Math.random(),
+    __wbg_set_4c81cfb5dc3a333c: (arg0, arg1, arg2) => { getObject(arg0)[arg1 >>> 0] = takeObject(arg2); },
+    __wbg_stack_3b0d974bbf31e44f: (arg0, arg1) => {
+      const ret = getObject(arg1).stack;
+      const ptr1 = passStringToWasm0(ret, wasm().__wbindgen_export, wasm().__wbindgen_export2);
+      const len1 = WASM_VECTOR_LEN;
+      getDataViewMemory0().setInt32(arg0 + 4, len1, true);
+      getDataViewMemory0().setInt32(arg0, ptr1, true);
+    },
+    __wbg_value_667dcb90597486a6: (arg0) => addHeapObject(getObject(arg0).value),
+    __wbindgen_cast_0000000000000001: (arg0, arg1) => addHeapObject(getStringFromWasm0(arg0, arg1)),
+    __wbindgen_object_drop_ref: (arg0) => takeObject(arg0),
+  };
+  return { __proto__: null, "./ruvector_attention_wasm_bg.js": import0 };
+};
+
+})(_mod, () => _wasm);
+
+
+// ── Async WASM init (fetch-based for browsers) ───────────────────
+
+export default async function initWasm() {
+  if (_initialized) return;
+  const wasmUrl = new URL('ruvector_attention_wasm_bg.wasm', import.meta.url);
+  const imports = _mod.__wbg_get_imports();
+  let result;
+  if (typeof WebAssembly.instantiateStreaming === 'function') {
+    try {
+      result = await WebAssembly.instantiateStreaming(fetch(wasmUrl), imports);
+    } catch (e) {
+      // Fallback if streaming fails (e.g. wrong MIME type)
+      const bytes = await (await fetch(wasmUrl)).arrayBuffer();
+      result = await WebAssembly.instantiate(bytes, imports);
+    }
+  } else {
+    const bytes = await (await fetch(wasmUrl)).arrayBuffer();
+    result = await WebAssembly.instantiate(bytes, imports);
+  }
+  _wasm = result.instance.exports;
+  _wasm.__wbindgen_start();
+  _initialized = true;
+}
+
+// ── ESM re-exports ────────────────────────────────────────────────
+// Attention mechanism classes
+export const WasmMultiHeadAttention = _mod.WasmMultiHeadAttention;
+export const WasmFlashAttention = _mod.WasmFlashAttention;
+export const WasmHyperbolicAttention = _mod.WasmHyperbolicAttention;
+export const WasmMoEAttention = _mod.WasmMoEAttention;
+export const WasmLinearAttention = _mod.WasmLinearAttention;
+export const WasmLocalGlobalAttention = _mod.WasmLocalGlobalAttention;
+// Utility functions
+export const cosine_similarity = _mod.cosine_similarity;
+export const normalize = _mod.normalize;
+export const l2_norm = _mod.l2_norm;
+export const softmax = _mod.softmax;
+export const batch_normalize = _mod.batch_normalize;
+export const pairwise_distances = _mod.pairwise_distances;
+export const scaled_dot_attention = _mod.scaled_dot_attention;
+export const attention_weights = _mod.attention_weights;
+export const random_orthogonal_matrix = _mod.random_orthogonal_matrix;
+export const available_mechanisms = _mod.available_mechanisms;
+// Lifecycle
+export const init = _mod.init;
+export const version = _mod.version;

ui/pose-fusion/pkg/ruvector-attention/ruvector_attention_wasm.d.ts

@@ -0,0 +1,359 @@
+/* tslint:disable */
+/* eslint-disable */
+
+/**
+ * Adam optimizer
+ */
+export class WasmAdam {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Create a new Adam optimizer
+     *
+     * # Arguments
+     * * `param_count` - Number of parameters
+     * * `learning_rate` - Learning rate
+     */
+    constructor(param_count: number, learning_rate: number);
+    /**
+     * Reset optimizer state
+     */
+    reset(): void;
+    /**
+     * Perform optimization step
+     *
+     * # Arguments
+     * * `params` - Current parameter values (will be updated in-place)
+     * * `gradients` - Gradient values
+     */
+    step(params: Float32Array, gradients: Float32Array): void;
+    /**
+     * Get current learning rate
+     */
+    learning_rate: number;
+}
+
+/**
+ * AdamW optimizer (Adam with decoupled weight decay)
+ */
+export class WasmAdamW {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Create a new AdamW optimizer
+     *
+     * # Arguments
+     * * `param_count` - Number of parameters
+     * * `learning_rate` - Learning rate
+     * * `weight_decay` - Weight decay coefficient
+     */
+    constructor(param_count: number, learning_rate: number, weight_decay: number);
+    /**
+     * Reset optimizer state
+     */
+    reset(): void;
+    /**
+     * Perform optimization step with weight decay
+     */
+    step(params: Float32Array, gradients: Float32Array): void;
+    /**
+     * Get current learning rate
+     */
+    learning_rate: number;
+    /**
+     * Get weight decay
+     */
+    readonly weight_decay: number;
+}
+
+/**
+ * Flash attention mechanism
+ */
+export class WasmFlashAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute flash attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new flash attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `block_size` - Block size for tiling
+     */
+    constructor(dim: number, block_size: number);
+}
+
+/**
+ * Hyperbolic attention mechanism
+ */
+export class WasmHyperbolicAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute hyperbolic attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new hyperbolic attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `curvature` - Hyperbolic curvature parameter
+     */
+    constructor(dim: number, curvature: number);
+    /**
+     * Get the curvature
+     */
+    readonly curvature: number;
+}
+
+/**
+ * InfoNCE contrastive loss for training
+ */
+export class WasmInfoNCELoss {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute InfoNCE loss
+     *
+     * # Arguments
+     * * `anchor` - Anchor embedding
+     * * `positive` - Positive example embedding
+     * * `negatives` - Array of negative example embeddings
+     */
+    compute(anchor: Float32Array, positive: Float32Array, negatives: any): number;
+    /**
+     * Create a new InfoNCE loss instance
+     *
+     * # Arguments
+     * * `temperature` - Temperature parameter for softmax
+     */
+    constructor(temperature: number);
+}
+
+/**
+ * Learning rate scheduler
+ */
+export class WasmLRScheduler {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Get learning rate for current step
+     */
+    get_lr(): number;
+    /**
+     * Create a new learning rate scheduler with warmup and cosine decay
+     *
+     * # Arguments
+     * * `initial_lr` - Initial learning rate
+     * * `warmup_steps` - Number of warmup steps
+     * * `total_steps` - Total training steps
+     */
+    constructor(initial_lr: number, warmup_steps: number, total_steps: number);
+    /**
+     * Reset scheduler
+     */
+    reset(): void;
+    /**
+     * Advance to next step
+     */
+    step(): void;
+}
+
+/**
+ * Linear attention (Performer-style)
+ */
+export class WasmLinearAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute linear attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new linear attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `num_features` - Number of random features
+     */
+    constructor(dim: number, num_features: number);
+}
+
+/**
+ * Local-global attention mechanism
+ */
+export class WasmLocalGlobalAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute local-global attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new local-global attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `local_window` - Size of local attention window
+     * * `global_tokens` - Number of global attention tokens
+     */
+    constructor(dim: number, local_window: number, global_tokens: number);
+}
+
+/**
+ * Mixture of Experts (MoE) attention
+ */
+export class WasmMoEAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute MoE attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new MoE attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `num_experts` - Number of expert attention mechanisms
+     * * `top_k` - Number of experts to use per query
+     */
+    constructor(dim: number, num_experts: number, top_k: number);
+}
+
+/**
+ * Multi-head attention mechanism
+ */
+export class WasmMultiHeadAttention {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Compute multi-head attention
+     */
+    compute(query: Float32Array, keys: any, values: any): Float32Array;
+    /**
+     * Create a new multi-head attention instance
+     *
+     * # Arguments
+     * * `dim` - Embedding dimension
+     * * `num_heads` - Number of attention heads
+     */
+    constructor(dim: number, num_heads: number);
+    /**
+     * Get the dimension
+     */
+    readonly dim: number;
+    /**
+     * Get the number of heads
+     */
+    readonly num_heads: number;
+}
+
+/**
+ * SGD optimizer with momentum
+ */
+export class WasmSGD {
+    free(): void;
+    [Symbol.dispose](): void;
+    /**
+     * Create a new SGD optimizer
+     *
+     * # Arguments
+     * * `param_count` - Number of parameters
+     * * `learning_rate` - Learning rate
+     * * `momentum` - Momentum coefficient (default: 0)
+     */
+    constructor(param_count: number, learning_rate: number, momentum?: number | null);
+    /**
+     * Reset optimizer state
+     */
+    reset(): void;
+    /**
+     * Perform optimization step
+     */
+    step(params: Float32Array, gradients: Float32Array): void;
+    /**
+     * Get current learning rate
+     */
+    learning_rate: number;
+}
+
+/**
+ * Compute attention weights from scores
+ */
+export function attention_weights(scores: Float32Array, temperature?: number | null): void;
+
+/**
+ * Get information about available attention mechanisms
+ */
+export function available_mechanisms(): any;
+
+/**
+ * Batch normalize vectors
+ */
+export function batch_normalize(vectors: any, epsilon?: number | null): Float32Array;
+
+/**
+ * Compute cosine similarity between two vectors
+ */
+export function cosine_similarity(a: Float32Array, b: Float32Array): number;
+
+/**
+ * Initialize the WASM module with panic hook
+ */
+export function init(): void;
+
+/**
+ * Compute L2 norm of a vector
+ */
+export function l2_norm(vec: Float32Array): number;
+
+/**
+ * Log a message to the browser console
+ */
+export function log(message: string): void;
+
+/**
+ * Log an error to the browser console
+ */
+export function log_error(message: string): void;
+
+/**
+ * Normalize a vector to unit length
+ */
+export function normalize(vec: Float32Array): void;
+
+/**
+ * Compute pairwise distances between vectors
+ */
+export function pairwise_distances(vectors: any): Float32Array;
+
+/**
+ * Generate random orthogonal matrix (for initialization)
+ */
+export function random_orthogonal_matrix(dim: number): Float32Array;
+
+/**
+ * Compute scaled dot-product attention
+ *
+ * # Arguments
+ * * `query` - Query vector as Float32Array
+ * * `keys` - Array of key vectors
+ * * `values` - Array of value vectors
+ * * `scale` - Optional scaling factor (defaults to 1/sqrt(dim))
+ */
+export function scaled_dot_attention(query: Float32Array, keys: any, values: any, scale?: number | null): Float32Array;
+
+/**
+ * Compute softmax of a vector
+ */
+export function softmax(vec: Float32Array): void;
+
+/**
+ * Get the version of the ruvector-attention-wasm crate
+ */
+export function version(): string;

ui/pose-fusion/pkg/ruvector-attention/ruvector_attention_wasm_bg.wasm.d.ts

@@ -0,0 +1,71 @@
+/* tslint:disable */
+/* eslint-disable */
+export const memory: WebAssembly.Memory;
+export const __wbg_wasmadam_free: (a: number, b: number) => void;
+export const __wbg_wasmadamw_free: (a: number, b: number) => void;
+export const __wbg_wasmflashattention_free: (a: number, b: number) => void;
+export const __wbg_wasmhyperbolicattention_free: (a: number, b: number) => void;
+export const __wbg_wasminfonceloss_free: (a: number, b: number) => void;
+export const __wbg_wasmlinearattention_free: (a: number, b: number) => void;
+export const __wbg_wasmmoeattention_free: (a: number, b: number) => void;
+export const __wbg_wasmmultiheadattention_free: (a: number, b: number) => void;
+export const __wbg_wasmsgd_free: (a: number, b: number) => void;
+export const attention_weights: (a: number, b: number, c: number, d: number) => void;
+export const available_mechanisms: () => number;
+export const batch_normalize: (a: number, b: number, c: number) => void;
+export const cosine_similarity: (a: number, b: number, c: number, d: number, e: number) => void;
+export const l2_norm: (a: number, b: number) => number;
+export const log: (a: number, b: number) => void;
+export const log_error: (a: number, b: number) => void;
+export const normalize: (a: number, b: number, c: number, d: number) => void;
+export const pairwise_distances: (a: number, b: number) => void;
+export const random_orthogonal_matrix: (a: number, b: number) => void;
+export const scaled_dot_attention: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const softmax: (a: number, b: number, c: number) => void;
+export const version: (a: number) => void;
+export const wasmadam_learning_rate: (a: number) => number;
+export const wasmadam_new: (a: number, b: number) => number;
+export const wasmadam_reset: (a: number) => void;
+export const wasmadam_set_learning_rate: (a: number, b: number) => void;
+export const wasmadam_step: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmadamw_new: (a: number, b: number, c: number) => number;
+export const wasmadamw_reset: (a: number) => void;
+export const wasmadamw_step: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmadamw_weight_decay: (a: number) => number;
+export const wasmflashattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmflashattention_new: (a: number, b: number) => number;
+export const wasmhyperbolicattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmhyperbolicattention_curvature: (a: number) => number;
+export const wasmhyperbolicattention_new: (a: number, b: number) => number;
+export const wasminfonceloss_compute: (a: number, b: number, c: number, d: number, e: number, f: number, g: number) => void;
+export const wasminfonceloss_new: (a: number) => number;
+export const wasmlinearattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmlinearattention_new: (a: number, b: number) => number;
+export const wasmlocalglobalattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmlocalglobalattention_new: (a: number, b: number, c: number) => number;
+export const wasmlrscheduler_get_lr: (a: number) => number;
+export const wasmlrscheduler_new: (a: number, b: number, c: number) => number;
+export const wasmlrscheduler_reset: (a: number) => void;
+export const wasmlrscheduler_step: (a: number) => void;
+export const wasmmoeattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmmoeattention_new: (a: number, b: number, c: number) => number;
+export const wasmmultiheadattention_compute: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const wasmmultiheadattention_dim: (a: number) => number;
+export const wasmmultiheadattention_new: (a: number, b: number, c: number) => void;
+export const wasmmultiheadattention_num_heads: (a: number) => number;
+export const wasmsgd_learning_rate: (a: number) => number;
+export const wasmsgd_new: (a: number, b: number, c: number) => number;
+export const wasmsgd_reset: (a: number) => void;
+export const wasmsgd_set_learning_rate: (a: number, b: number) => void;
+export const wasmsgd_step: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const init: () => void;
+export const wasmadamw_set_learning_rate: (a: number, b: number) => void;
+export const wasmadamw_learning_rate: (a: number) => number;
+export const __wbg_wasmlocalglobalattention_free: (a: number, b: number) => void;
+export const __wbg_wasmlrscheduler_free: (a: number, b: number) => void;
+export const __wbindgen_export: (a: number, b: number) => number;
+export const __wbindgen_export2: (a: number, b: number, c: number, d: number) => number;
+export const __wbindgen_export3: (a: number) => void;
+export const __wbindgen_export4: (a: number, b: number, c: number) => void;
+export const __wbindgen_add_to_stack_pointer: (a: number) => number;
+export const __wbindgen_start: () => void;

ui/pose-fusion/pkg/ruvector_cnn_wasm/package.json

@@ -0,0 +1,26 @@
+{
+  "name": "ruvector-cnn-wasm",
+  "type": "module",
+  "description": "WASM bindings for ruvector-cnn - CNN feature extraction for image embeddings",
+  "version": "0.1.0",
+  "license": "MIT OR Apache-2.0",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/ruvnet/ruvector"
+  },
+  "files": [
+    "ruvector_cnn_wasm_bg.wasm",
+    "ruvector_cnn_wasm.js"
+  ],
+  "main": "ruvector_cnn_wasm.js",
+  "sideEffects": [
+    "./snippets/*"
+  ],
+  "keywords": [
+    "cnn",
+    "embeddings",
+    "wasm",
+    "simd",
+    "machine-learning"
+  ]
+}

ui/pose-fusion/pkg/ruvector_cnn_wasm/ruvector_cnn_wasm.js

@@ -0,0 +1,802 @@
+/**
+ * Configuration for CNN embedder
+ */
+export class EmbedderConfig {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        EmbedderConfigFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_embedderconfig_free(ptr, 0);
+    }
+    constructor() {
+        const ret = wasm.embedderconfig_new();
+        this.__wbg_ptr = ret >>> 0;
+        EmbedderConfigFinalization.register(this, this.__wbg_ptr, this);
+        return this;
+    }
+    /**
+     * Output embedding dimension
+     * @returns {number}
+     */
+    get embedding_dim() {
+        const ret = wasm.__wbg_get_embedderconfig_embedding_dim(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Input image size (square)
+     * @returns {number}
+     */
+    get input_size() {
+        const ret = wasm.__wbg_get_embedderconfig_input_size(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Whether to L2 normalize embeddings
+     * @returns {boolean}
+     */
+    get normalize() {
+        const ret = wasm.__wbg_get_embedderconfig_normalize(this.__wbg_ptr);
+        return ret !== 0;
+    }
+    /**
+     * Output embedding dimension
+     * @param {number} arg0
+     */
+    set embedding_dim(arg0) {
+        wasm.__wbg_set_embedderconfig_embedding_dim(this.__wbg_ptr, arg0);
+    }
+    /**
+     * Input image size (square)
+     * @param {number} arg0
+     */
+    set input_size(arg0) {
+        wasm.__wbg_set_embedderconfig_input_size(this.__wbg_ptr, arg0);
+    }
+    /**
+     * Whether to L2 normalize embeddings
+     * @param {boolean} arg0
+     */
+    set normalize(arg0) {
+        wasm.__wbg_set_embedderconfig_normalize(this.__wbg_ptr, arg0);
+    }
+}
+if (Symbol.dispose) EmbedderConfig.prototype[Symbol.dispose] = EmbedderConfig.prototype.free;
+
+/**
+ * Layer operations for building custom networks
+ */
+export class LayerOps {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        LayerOpsFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_layerops_free(ptr, 0);
+    }
+    /**
+     * Apply batch normalization (returns new array)
+     * @param {Float32Array} input
+     * @param {Float32Array} gamma
+     * @param {Float32Array} beta
+     * @param {Float32Array} mean
+     * @param {Float32Array} _var
+     * @param {number} epsilon
+     * @returns {Float32Array}
+     */
+    static batch_norm(input, gamma, beta, mean, _var, epsilon) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(input, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            const ptr1 = passArrayF32ToWasm0(gamma, wasm.__wbindgen_export2);
+            const len1 = WASM_VECTOR_LEN;
+            const ptr2 = passArrayF32ToWasm0(beta, wasm.__wbindgen_export2);
+            const len2 = WASM_VECTOR_LEN;
+            const ptr3 = passArrayF32ToWasm0(mean, wasm.__wbindgen_export2);
+            const len3 = WASM_VECTOR_LEN;
+            const ptr4 = passArrayF32ToWasm0(_var, wasm.__wbindgen_export2);
+            const len4 = WASM_VECTOR_LEN;
+            wasm.layerops_batch_norm(retptr, ptr0, len0, ptr1, len1, ptr2, len2, ptr3, len3, ptr4, len4, epsilon);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var v6 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v6;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Apply global average pooling
+     * Returns one value per channel
+     * @param {Float32Array} input
+     * @param {number} height
+     * @param {number} width
+     * @param {number} channels
+     * @returns {Float32Array}
+     */
+    static global_avg_pool(input, height, width, channels) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(input, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.layerops_global_avg_pool(retptr, ptr0, len0, height, width, channels);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var v2 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v2;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+}
+if (Symbol.dispose) LayerOps.prototype[Symbol.dispose] = LayerOps.prototype.free;
+
+/**
+ * SIMD-optimized operations
+ */
+export class SimdOps {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        SimdOpsFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_simdops_free(ptr, 0);
+    }
+    /**
+     * Dot product of two vectors
+     * @param {Float32Array} a
+     * @param {Float32Array} b
+     * @returns {number}
+     */
+    static dot_product(a, b) {
+        const ptr0 = passArrayF32ToWasm0(a, wasm.__wbindgen_export2);
+        const len0 = WASM_VECTOR_LEN;
+        const ptr1 = passArrayF32ToWasm0(b, wasm.__wbindgen_export2);
+        const len1 = WASM_VECTOR_LEN;
+        const ret = wasm.simdops_dot_product(ptr0, len0, ptr1, len1);
+        return ret;
+    }
+    /**
+     * L2 normalize a vector (returns new array)
+     * @param {Float32Array} data
+     * @returns {Float32Array}
+     */
+    static l2_normalize(data) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(data, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.simdops_l2_normalize(retptr, ptr0, len0);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var v2 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v2;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * ReLU activation (returns new array)
+     * @param {Float32Array} data
+     * @returns {Float32Array}
+     */
+    static relu(data) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(data, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.simdops_relu(retptr, ptr0, len0);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var v2 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v2;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * ReLU6 activation (returns new array)
+     * @param {Float32Array} data
+     * @returns {Float32Array}
+     */
+    static relu6(data) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(data, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.simdops_relu6(retptr, ptr0, len0);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var v2 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v2;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+}
+if (Symbol.dispose) SimdOps.prototype[Symbol.dispose] = SimdOps.prototype.free;
+
+/**
+ * WASM CNN Embedder for image feature extraction
+ */
+export class WasmCnnEmbedder {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        WasmCnnEmbedderFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_wasmcnnembedder_free(ptr, 0);
+    }
+    /**
+     * Compute cosine similarity between two embeddings
+     * @param {Float32Array} a
+     * @param {Float32Array} b
+     * @returns {number}
+     */
+    cosine_similarity(a, b) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(a, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            const ptr1 = passArrayF32ToWasm0(b, wasm.__wbindgen_export2);
+            const len1 = WASM_VECTOR_LEN;
+            wasm.wasmcnnembedder_cosine_similarity(retptr, this.__wbg_ptr, ptr0, len0, ptr1, len1);
+            var r0 = getDataViewMemory0().getFloat32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            if (r2) {
+                throw takeObject(r1);
+            }
+            return r0;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Get the embedding dimension
+     * @returns {number}
+     */
+    get embedding_dim() {
+        const ret = wasm.wasmcnnembedder_embedding_dim(this.__wbg_ptr);
+        return ret >>> 0;
+    }
+    /**
+     * Extract embedding from image data (RGB format, row-major)
+     * @param {Uint8Array} image_data
+     * @param {number} width
+     * @param {number} height
+     * @returns {Float32Array}
+     */
+    extract(image_data, width, height) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArray8ToWasm0(image_data, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.wasmcnnembedder_extract(retptr, this.__wbg_ptr, ptr0, len0, width, height);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            var r3 = getDataViewMemory0().getInt32(retptr + 4 * 3, true);
+            if (r3) {
+                throw takeObject(r2);
+            }
+            var v2 = getArrayF32FromWasm0(r0, r1).slice();
+            wasm.__wbindgen_export(r0, r1 * 4, 4);
+            return v2;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Create a new CNN embedder
+     * @param {EmbedderConfig | null} [config]
+     */
+    constructor(config) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            let ptr0 = 0;
+            if (!isLikeNone(config)) {
+                _assertClass(config, EmbedderConfig);
+                ptr0 = config.__destroy_into_raw();
+            }
+            wasm.wasmcnnembedder_new(retptr, ptr0);
+            var r0 = getDataViewMemory0().getInt32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            if (r2) {
+                throw takeObject(r1);
+            }
+            this.__wbg_ptr = r0 >>> 0;
+            WasmCnnEmbedderFinalization.register(this, this.__wbg_ptr, this);
+            return this;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+}
+if (Symbol.dispose) WasmCnnEmbedder.prototype[Symbol.dispose] = WasmCnnEmbedder.prototype.free;
+
+/**
+ * InfoNCE loss for contrastive learning (SimCLR style)
+ */
+export class WasmInfoNCELoss {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        WasmInfoNCELossFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_wasminfonceloss_free(ptr, 0);
+    }
+    /**
+     * Compute loss for a batch of embedding pairs
+     * embeddings: [2N, D] flattened where (i, i+N) are positive pairs
+     * @param {Float32Array} embeddings
+     * @param {number} batch_size
+     * @param {number} dim
+     * @returns {number}
+     */
+    forward(embeddings, batch_size, dim) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(embeddings, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            wasm.wasminfonceloss_forward(retptr, this.__wbg_ptr, ptr0, len0, batch_size, dim);
+            var r0 = getDataViewMemory0().getFloat32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            if (r2) {
+                throw takeObject(r1);
+            }
+            return r0;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Create new InfoNCE loss with temperature parameter
+     * @param {number} temperature
+     */
+    constructor(temperature) {
+        const ret = wasm.wasminfonceloss_new(temperature);
+        this.__wbg_ptr = ret >>> 0;
+        WasmInfoNCELossFinalization.register(this, this.__wbg_ptr, this);
+        return this;
+    }
+    /**
+     * Get the temperature parameter
+     * @returns {number}
+     */
+    get temperature() {
+        const ret = wasm.wasminfonceloss_temperature(this.__wbg_ptr);
+        return ret;
+    }
+}
+if (Symbol.dispose) WasmInfoNCELoss.prototype[Symbol.dispose] = WasmInfoNCELoss.prototype.free;
+
+/**
+ * Triplet loss for metric learning
+ */
+export class WasmTripletLoss {
+    __destroy_into_raw() {
+        const ptr = this.__wbg_ptr;
+        this.__wbg_ptr = 0;
+        WasmTripletLossFinalization.unregister(this);
+        return ptr;
+    }
+    free() {
+        const ptr = this.__destroy_into_raw();
+        wasm.__wbg_wasmtripletloss_free(ptr, 0);
+    }
+    /**
+     * Compute loss for a batch of triplets
+     * @param {Float32Array} anchors
+     * @param {Float32Array} positives
+     * @param {Float32Array} negatives
+     * @param {number} dim
+     * @returns {number}
+     */
+    forward(anchors, positives, negatives, dim) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(anchors, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            const ptr1 = passArrayF32ToWasm0(positives, wasm.__wbindgen_export2);
+            const len1 = WASM_VECTOR_LEN;
+            const ptr2 = passArrayF32ToWasm0(negatives, wasm.__wbindgen_export2);
+            const len2 = WASM_VECTOR_LEN;
+            wasm.wasmtripletloss_forward(retptr, this.__wbg_ptr, ptr0, len0, ptr1, len1, ptr2, len2, dim);
+            var r0 = getDataViewMemory0().getFloat32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            if (r2) {
+                throw takeObject(r1);
+            }
+            return r0;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Compute loss for a single triplet
+     * @param {Float32Array} anchor
+     * @param {Float32Array} positive
+     * @param {Float32Array} negative
+     * @returns {number}
+     */
+    forward_single(anchor, positive, negative) {
+        try {
+            const retptr = wasm.__wbindgen_add_to_stack_pointer(-16);
+            const ptr0 = passArrayF32ToWasm0(anchor, wasm.__wbindgen_export2);
+            const len0 = WASM_VECTOR_LEN;
+            const ptr1 = passArrayF32ToWasm0(positive, wasm.__wbindgen_export2);
+            const len1 = WASM_VECTOR_LEN;
+            const ptr2 = passArrayF32ToWasm0(negative, wasm.__wbindgen_export2);
+            const len2 = WASM_VECTOR_LEN;
+            wasm.wasmtripletloss_forward_single(retptr, this.__wbg_ptr, ptr0, len0, ptr1, len1, ptr2, len2);
+            var r0 = getDataViewMemory0().getFloat32(retptr + 4 * 0, true);
+            var r1 = getDataViewMemory0().getInt32(retptr + 4 * 1, true);
+            var r2 = getDataViewMemory0().getInt32(retptr + 4 * 2, true);
+            if (r2) {
+                throw takeObject(r1);
+            }
+            return r0;
+        } finally {
+            wasm.__wbindgen_add_to_stack_pointer(16);
+        }
+    }
+    /**
+     * Get the margin parameter
+     * @returns {number}
+     */
+    get margin() {
+        const ret = wasm.wasmtripletloss_margin(this.__wbg_ptr);
+        return ret;
+    }
+    /**
+     * Create new triplet loss with margin
+     * @param {number} margin
+     */
+    constructor(margin) {
+        const ret = wasm.wasmtripletloss_new(margin);
+        this.__wbg_ptr = ret >>> 0;
+        WasmTripletLossFinalization.register(this, this.__wbg_ptr, this);
+        return this;
+    }
+}
+if (Symbol.dispose) WasmTripletLoss.prototype[Symbol.dispose] = WasmTripletLoss.prototype.free;
+
+/**
+ * Initialize panic hook for better error messages
+ */
+export function init() {
+    wasm.init();
+}
+
+function __wbg_get_imports() {
+    const import0 = {
+        __proto__: null,
+        __wbg___wbindgen_throw_39bc967c0e5a9b58: function(arg0, arg1) {
+            throw new Error(getStringFromWasm0(arg0, arg1));
+        },
+        __wbg_error_a6fa202b58aa1cd3: function(arg0, arg1) {
+            let deferred0_0;
+            let deferred0_1;
+            try {
+                deferred0_0 = arg0;
+                deferred0_1 = arg1;
+                console.error(getStringFromWasm0(arg0, arg1));
+            } finally {
+                wasm.__wbindgen_export(deferred0_0, deferred0_1, 1);
+            }
+        },
+        __wbg_new_227d7c05414eb861: function() {
+            const ret = new Error();
+            return addHeapObject(ret);
+        },
+        __wbg_stack_3b0d974bbf31e44f: function(arg0, arg1) {
+            const ret = getObject(arg1).stack;
+            const ptr1 = passStringToWasm0(ret, wasm.__wbindgen_export2, wasm.__wbindgen_export3);
+            const len1 = WASM_VECTOR_LEN;
+            getDataViewMemory0().setInt32(arg0 + 4 * 1, len1, true);
+            getDataViewMemory0().setInt32(arg0 + 4 * 0, ptr1, true);
+        },
+        __wbindgen_cast_0000000000000001: function(arg0, arg1) {
+            // Cast intrinsic for `Ref(String) -> Externref`.
+            const ret = getStringFromWasm0(arg0, arg1);
+            return addHeapObject(ret);
+        },
+        __wbindgen_object_drop_ref: function(arg0) {
+            takeObject(arg0);
+        },
+    };
+    return {
+        __proto__: null,
+        "./ruvector_cnn_wasm_bg.js": import0,
+    };
+}
+
+const EmbedderConfigFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_embedderconfig_free(ptr >>> 0, 1));
+const LayerOpsFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_layerops_free(ptr >>> 0, 1));
+const SimdOpsFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_simdops_free(ptr >>> 0, 1));
+const WasmCnnEmbedderFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_wasmcnnembedder_free(ptr >>> 0, 1));
+const WasmInfoNCELossFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_wasminfonceloss_free(ptr >>> 0, 1));
+const WasmTripletLossFinalization = (typeof FinalizationRegistry === 'undefined')
+    ? { register: () => {}, unregister: () => {} }
+    : new FinalizationRegistry(ptr => wasm.__wbg_wasmtripletloss_free(ptr >>> 0, 1));
+
+function addHeapObject(obj) {
+    if (heap_next === heap.length) heap.push(heap.length + 1);
+    const idx = heap_next;
+    heap_next = heap[idx];
+
+    heap[idx] = obj;
+    return idx;
+}
+
+function _assertClass(instance, klass) {
+    if (!(instance instanceof klass)) {
+        throw new Error(`expected instance of ${klass.name}`);
+    }
+}
+
+function dropObject(idx) {
+    if (idx < 1028) return;
+    heap[idx] = heap_next;
+    heap_next = idx;
+}
+
+function getArrayF32FromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return getFloat32ArrayMemory0().subarray(ptr / 4, ptr / 4 + len);
+}
+
+let cachedDataViewMemory0 = null;
+function getDataViewMemory0() {
+    if (cachedDataViewMemory0 === null || cachedDataViewMemory0.buffer.detached === true || (cachedDataViewMemory0.buffer.detached === undefined && cachedDataViewMemory0.buffer !== wasm.memory.buffer)) {
+        cachedDataViewMemory0 = new DataView(wasm.memory.buffer);
+    }
+    return cachedDataViewMemory0;
+}
+
+let cachedFloat32ArrayMemory0 = null;
+function getFloat32ArrayMemory0() {
+    if (cachedFloat32ArrayMemory0 === null || cachedFloat32ArrayMemory0.byteLength === 0) {
+        cachedFloat32ArrayMemory0 = new Float32Array(wasm.memory.buffer);
+    }
+    return cachedFloat32ArrayMemory0;
+}
+
+function getStringFromWasm0(ptr, len) {
+    ptr = ptr >>> 0;
+    return decodeText(ptr, len);
+}
+
+let cachedUint8ArrayMemory0 = null;
+function getUint8ArrayMemory0() {
+    if (cachedUint8ArrayMemory0 === null || cachedUint8ArrayMemory0.byteLength === 0) {
+        cachedUint8ArrayMemory0 = new Uint8Array(wasm.memory.buffer);
+    }
+    return cachedUint8ArrayMemory0;
+}
+
+function getObject(idx) { return heap[idx]; }
+
+let heap = new Array(1024).fill(undefined);
+heap.push(undefined, null, true, false);
+
+let heap_next = heap.length;
+
+function isLikeNone(x) {
+    return x === undefined || x === null;
+}
+
+function passArray8ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 1, 1) >>> 0;
+    getUint8ArrayMemory0().set(arg, ptr / 1);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passArrayF32ToWasm0(arg, malloc) {
+    const ptr = malloc(arg.length * 4, 4) >>> 0;
+    getFloat32ArrayMemory0().set(arg, ptr / 4);
+    WASM_VECTOR_LEN = arg.length;
+    return ptr;
+}
+
+function passStringToWasm0(arg, malloc, realloc) {
+    if (realloc === undefined) {
+        const buf = cachedTextEncoder.encode(arg);
+        const ptr = malloc(buf.length, 1) >>> 0;
+        getUint8ArrayMemory0().subarray(ptr, ptr + buf.length).set(buf);
+        WASM_VECTOR_LEN = buf.length;
+        return ptr;
+    }
+
+    let len = arg.length;
+    let ptr = malloc(len, 1) >>> 0;
+
+    const mem = getUint8ArrayMemory0();
+
+    let offset = 0;
+
+    for (; offset < len; offset++) {
+        const code = arg.charCodeAt(offset);
+        if (code > 0x7F) break;
+        mem[ptr + offset] = code;
+    }
+    if (offset !== len) {
+        if (offset !== 0) {
+            arg = arg.slice(offset);
+        }
+        ptr = realloc(ptr, len, len = offset + arg.length * 3, 1) >>> 0;
+        const view = getUint8ArrayMemory0().subarray(ptr + offset, ptr + len);
+        const ret = cachedTextEncoder.encodeInto(arg, view);
+
+        offset += ret.written;
+        ptr = realloc(ptr, len, offset, 1) >>> 0;
+    }
+
+    WASM_VECTOR_LEN = offset;
+    return ptr;
+}
+
+function takeObject(idx) {
+    const ret = getObject(idx);
+    dropObject(idx);
+    return ret;
+}
+
+let cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+cachedTextDecoder.decode();
+const MAX_SAFARI_DECODE_BYTES = 2146435072;
+let numBytesDecoded = 0;
+function decodeText(ptr, len) {
+    numBytesDecoded += len;
+    if (numBytesDecoded >= MAX_SAFARI_DECODE_BYTES) {
+        cachedTextDecoder = new TextDecoder('utf-8', { ignoreBOM: true, fatal: true });
+        cachedTextDecoder.decode();
+        numBytesDecoded = len;
+    }
+    return cachedTextDecoder.decode(getUint8ArrayMemory0().subarray(ptr, ptr + len));
+}
+
+const cachedTextEncoder = new TextEncoder();
+
+if (!('encodeInto' in cachedTextEncoder)) {
+    cachedTextEncoder.encodeInto = function (arg, view) {
+        const buf = cachedTextEncoder.encode(arg);
+        view.set(buf);
+        return {
+            read: arg.length,
+            written: buf.length
+        };
+    };
+}
+
+let WASM_VECTOR_LEN = 0;
+
+let wasmModule, wasm;
+function __wbg_finalize_init(instance, module) {
+    wasm = instance.exports;
+    wasmModule = module;
+    cachedDataViewMemory0 = null;
+    cachedFloat32ArrayMemory0 = null;
+    cachedUint8ArrayMemory0 = null;
+    wasm.__wbindgen_start();
+    return wasm;
+}
+
+async function __wbg_load(module, imports) {
+    if (typeof Response === 'function' && module instanceof Response) {
+        if (typeof WebAssembly.instantiateStreaming === 'function') {
+            try {
+                return await WebAssembly.instantiateStreaming(module, imports);
+            } catch (e) {
+                const validResponse = module.ok && expectedResponseType(module.type);
+
+                if (validResponse && module.headers.get('Content-Type') !== 'application/wasm') {
+                    console.warn("`WebAssembly.instantiateStreaming` failed because your server does not serve Wasm with `application/wasm` MIME type. Falling back to `WebAssembly.instantiate` which is slower. Original error:\n", e);
+
+                } else { throw e; }
+            }
+        }
+
+        const bytes = await module.arrayBuffer();
+        return await WebAssembly.instantiate(bytes, imports);
+    } else {
+        const instance = await WebAssembly.instantiate(module, imports);
+
+        if (instance instanceof WebAssembly.Instance) {
+            return { instance, module };
+        } else {
+            return instance;
+        }
+    }
+
+    function expectedResponseType(type) {
+        switch (type) {
+            case 'basic': case 'cors': case 'default': return true;
+        }
+        return false;
+    }
+}
+
+function initSync(module) {
+    if (wasm !== undefined) return wasm;
+
+
+    if (module !== undefined) {
+        if (Object.getPrototypeOf(module) === Object.prototype) {
+            ({module} = module)
+        } else {
+            console.warn('using deprecated parameters for `initSync()`; pass a single object instead')
+        }
+    }
+
+    const imports = __wbg_get_imports();
+    if (!(module instanceof WebAssembly.Module)) {
+        module = new WebAssembly.Module(module);
+    }
+    const instance = new WebAssembly.Instance(module, imports);
+    return __wbg_finalize_init(instance, module);
+}
+
+async function __wbg_init(module_or_path) {
+    if (wasm !== undefined) return wasm;
+
+
+    if (module_or_path !== undefined) {
+        if (Object.getPrototypeOf(module_or_path) === Object.prototype) {
+            ({module_or_path} = module_or_path)
+        } else {
+            console.warn('using deprecated parameters for the initialization function; pass a single object instead')
+        }
+    }
+
+    if (module_or_path === undefined) {
+        module_or_path = new URL('ruvector_cnn_wasm_bg.wasm', import.meta.url);
+    }
+    const imports = __wbg_get_imports();
+
+    if (typeof module_or_path === 'string' || (typeof Request === 'function' && module_or_path instanceof Request) || (typeof URL === 'function' && module_or_path instanceof URL)) {
+        module_or_path = fetch(module_or_path);
+    }
+
+    const { instance, module } = await __wbg_load(await module_or_path, imports);
+
+    return __wbg_finalize_init(instance, module);
+}
+
+export { initSync, __wbg_init as default };