Track each ESP32 node independently instead of merging all CSI frames
into a single buffer. This enables per-node feature computation,
spatial awareness, and proper multi-node visualization.
Per-node CSI separation:
- Add NodeState struct with per-node frame_history, RSSI history,
features, classification, and smoothing state
- Compute features per-node using each node's own temporal history
- Add compute_fused_features() for backward-compatible aggregate
- Add smooth_and_classify_node() for per-node motion classification
- Add GET /api/v1/nodes endpoint for per-node health/status
- Add PerNodeFeatureInfo to WebSocket SensingUpdate messages
- Fix RSSI sign (use saturating_neg for correct negative dBm values)
- Node timeout: stale after 5s, removed after 30s
Dynamic classifier classes:
- Remove hardcoded CLASSES array and N_CLASSES constant
- Discover classes automatically from training data filenames
- Convention: train_<class>_<description>.jsonl
- Users can add any class by recording with appropriate filename
- Backward compatible with existing 4-class models via serde default
- AdaptiveModel now stores class_names as Vec<String>
UI changes:
- Dynamic node count display (was hardcoded "1 ESP32")
- Per-node status cards showing RSSI, variance, classification
- Color-coded node markers in 3D gaussian splat view
- Per-node RSSI history tracking in sensing service
- XSS-safe DOM element creation (no innerHTML with server data)
Addresses #237, #276, #51
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
11a413d7ef