feat(sensing-server): per-node CSI separation + dynamic classifier classes

[taylorjdawson]

Summary

• Track each ESP32 node independently instead of merging all CSI frames into a single buffer
• Make adaptive classifier classes dynamic — users add classes via filename convention, no code changes needed
• Add per-node status UI with colored markers and signal features per node
• Fix RSSI sign bug and XSS vulnerability in sensing UI

Motivation

The sensing server merges CSI frames from all ESP32 nodes into one frame_history buffer, discarding node_id after parsing. This means:

• Temporal features (variance, motion) compare frames from different physical nodes
• No spatial information — can't tell which node is seeing activity
• UI shows "1 ESP32" despite multiple nodes connected
• Classification accuracy is degraded by mixed-node data

Addresses #237 (multi-node display identical for all states), #276 (only one detected), #51 (amplitude detection fragile).

Implements server-side per-node tracking from the ADR-029 (RuvSense multistatic sensing) architecture.

Changes

Per-node CSI separation (sensing-server/src/main.rs)

• NodeState struct — per-node frame_history, RSSI history, features, classification, smoothing state
• smooth_and_classify_node() — per-node motion classification with EMA/debounce
• compute_fused_features() — weighted aggregation across active nodes; max-boosted for presence-sensitive features (variance, motion_band_power) so single-node strong signals aren't diluted
• build_per_node_features() — sorted per-node feature list for WebSocket broadcast
• nodes_endpoint() — new GET /api/v1/nodes endpoint returns per-node health, frame rate, features, classification
• RSSI sign fix — saturating_neg() for correct negative dBm values
• Signal field uses fused features instead of single-node
• Node timeout — stale after 5s, removed after 30s
• SensingUpdate.node_features — optional field, backward compatible via skip_serializing_if
• Default impls for FeatureInfo and ClassificationInfo

Dynamic classifier classes (adaptive_classifier.rs)

• Removed hardcoded CLASSES array and N_CLASSES constant
• classify_recording_name returns Option<String> — discovers classes from filenames
• Convention: train_<class>_<description>.jsonl (e.g., train_cooking_kitchen.jsonl)
• Common patterns still recognized for backward compat: *absent*, *still*, *walking*, *active*
• Unknown patterns extract class from filename structure as fallback
• AdaptiveModel.class_names: Vec<String> — dynamic, serialized in model JSON
• AdaptiveModel.weights: Vec<Vec<f64>> — dynamic class count instead of fixed array
• Backward compatible: old 4-class models load via #[serde(default)]

UI changes

• Dynamic node count (was hardcoded "1 ESP32")
• Per-node status cards with RSSI, variance, classification (DOM createElement, no innerHTML — XSS safe)
• Color-coded node markers in 3D gaussian splat view (8-color palette)
• Per-node RSSI history tracking in sensing service

Backward Compatibility

• SensingUpdate.features still populated with fused aggregate — existing consumers unchanged
• SensingUpdate.nodes now contains ALL active nodes (was single node per message) — existing code reading nodes[0] still works
• node_features field is Option with skip_serializing_if — old clients don't receive it
• Old 4-class adaptive models load correctly via serde defaults
• Global frame_history still maintained alongside per-node histories
• Supports any number of nodes (1 to 256) — single-node deployments work identically to before

How to test

# Build
cargo build -p wifi-densepose-sensing-server

# Run with ESP32 nodes
cargo run -p wifi-densepose-sensing-server -- --http-port 3000 --source esp32

# Verify per-node data
curl http://localhost:3000/api/v1/nodes

# Verify backward compat (fused features still present)
curl http://localhost:3000/api/v1/sensing/latest | jq '.features'

# Open UI — should show per-node cards and colored markers
open http://localhost:3000/ui/index.html

# Test dynamic classes — add any class by filename
echo '{"features":{...}}' > data/recordings/train_cooking_kitchen.jsonl
curl -X POST http://localhost:3000/api/v1/adaptive/train
# → model now includes "cooking" class

🤖 Generated with Claude Code