feat: implement predictive location modeling with Kalman filtering (closes #263)

[levoski1]

Summary

Implements predictive location modeling engine with Kalman filtering, trajectory extrapolation, and anomaly detection for sub-second ETA and geofence pre-alerts.

Changes

New Module: src/predictor.js

• PredictiveEngine class with per-client Kalman filter (Constant Velocity model)
• ENU coordinate transformation (GPS lat/lon <-> local tangent plane meters)
• Trajectory extrapolation API - getTrajectory(clientId, horizons) returns predicted positions with confidence ellipses from filter covariance
• Geofence pre-alerts - checkPreAlerts(clientId) evaluates predicted trajectory against geofences, emits geofence_pre_alert events 60s before predicted entry (debounced per client/fence)
• Anomaly detection:
  • GPS anomaly: innovation (residual) exceeds 3σ -> gps_anomaly
  • Kinematic anomaly: impossible acceleration (>10 m/s²) or heading rate (>π/2 rad/s) -> kinematic_anomaly
  • Rate limited to 1/min per client per anomaly type
• ETA computation - getETA(clientId, targetLat, targetLon) returns mean, stddev, and arrival probability function
• Session persistence - filter state saved/restored via session resumption (issue 18)

Configuration (.env.example)

• PREDICTOR_ENABLE=true
• PREDICTOR_MODEL=CV
• PREDICTOR_PROCESS_NOISE=0.1
• PREDICTOR_MEASUREMENT_NOISE=5.0
• PREDICTOR_PRE_ALERT_HORIZON_S=60
• PREDICTOR_MAX_HORIZON_S=120
• Additional anomaly detection and TTL settings

Integration (src/server.js)

• Initialize PredictiveEngine with roomManager
• Call predictor.update() on each location_update
• Broadcast anomaly events (gps_anomaly, kinematic_anomaly)
• Call predictor.checkPreAlerts() for geofence pre-alerts
• Cleanup filter on client disconnect

Tests (tests/predictor.test.js)

• 19 comprehensive tests covering:
  • ENU coordinate transformations
  • Kalman filter convergence (stationary vehicle uncertainty decreases, moving vehicle 30m/s prediction error < 50m)
  • Trajectory extrapolation with confidence ellipses
  • Geofence pre-alerts with debouncing
  • Anomaly detection (GPS, kinematic) with rate limiting
  • ETA computation with Monte Carlo validation
  • Session persistence/restoration
  • Cleanup and disabled mode

Bug Fixes

• Fixed pre-existing duplicate tokenRefreshSchema declaration in validator.js
• Fixed pre-existing undefined sessions variable references in server.js (renamed to sessionManager)

Acceptance Criteria Met

• Kalman filter converges: stationary vehicle -> position uncertainty decreases over time
• Moving vehicle at 30 m/s -> 30s prediction error < 50m (simulated GPS noise σ=5m)
• getTrajectory returns correct ENU->lat/lon transform with confidence ellipses
• Geofence pre-alert emitted 60s before predicted entry
• GPS anomaly detected when innovation > 3σ
• Kinematic anomaly detected for impossible acceleration/heading rate
• ETA distribution mean ± stddev matches Monte Carlo simulation
• Filter state saved/restored via session resumption
• npm run lint passes
• All existing tests pass (core modules)
• New test file with comprehensive coverage

Closes #263