feat(ui): 3D stick figure visualization (COCO-17 + smooth animations)

[schwarztim]

Motivation

Spatial dashboards currently show voxel "blobs" from CSI tomography — non-intuitive for end users. This PR adds a COCO-17 stick figure renderer to ui/spatial.html that maps body cluster centroids to anatomical poses, with smooth lerp animations between ticks and posture-aware idle motion (breathing, gentle sway).

Changes

• ui/spatial.html — Three.js / Canvas-based 3D renderer with:
  • COCO-17 joint definitions inline (17 keypoints, 18 bone pairs)
  • Three pose variants driven by pose_hint: standing (green), sitting (yellow), lying (blue)
  • FigureState persistent figure map — no recreate-on-every-frame
  • Position lerp (factor 0.1) for smooth interpolation between cluster ticks
  • Posture-transition keypoint blending over ~1 s (poseBlend 0→1 at +0.03/frame)
  • Breathing animation: torso Y-scale at detected BR rate from /api/v1/vital-signs
  • Activity-based sway: still=0.01 m, moderate=0.03 m, active=0.08 m
  • Opacity fadeout for disappearing bodies (quality decrements at −0.05/tick)
  • Fallback: single standing figure from voxel centroid when no cluster data
• CHANGELOG.md — entry under [Unreleased]

Backend dependency

This UI consumes:

• GET /api/v1/spatial/clusters — body cluster centroids + bbox + pose hints (provided by body tracking PR feat(sensing-server): body tracking platform with persistent IDs and zones #487)
• GET /api/v1/tracking/bodies — persistent body IDs (provided by feat(sensing-server): body tracking platform with persistent IDs and zones #487)
• GET /api/v1/vital-signs — breathing rate for animation pacing (existing endpoint)

For best results merge this AFTER #487. Filing now to surface for parallel review.

WebSocket schema

The renderer subscribes to the existing /ws/sensing endpoint and handles spatial_update messages that include a clusters array:

{
  "type": "spatial_update",
  "data": {
    "clusters": [
      {
        "id": 0,
        "centroid": [x, y, z],
        "bbox_min": [x, y, z],
        "bbox_max": [x, y, z],
        "pose_hint": "standing | sitting | lying"
      }
    ]
  }
}

Falls back to polling /api/v1/spatial/clusters every 2 s if WebSocket is unavailable.

COCO-17 joint definitions

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

18 bone pairs connecting adjacent joints per the COCO topology. Definitions are documented inline in spatial.html as STANDING_KEYPOINTS and BONES constants.

Visual notes

• Dark theme: background #050508, overlays rgba(10,10,15,0.85) — MD3-compatible dark surface
• Stick figures render at 60 fps via requestAnimationFrame; lerp keeps motion smooth at 2 s poll intervals
• No external 3D dependencies beyond Three.js (loaded via importmap from unpkg CDN, same pattern as other UI pages)
• Legend updated with pose-colored entries (standing/sitting/lying swatches)

Test results

• cargo check --workspace --no-default-features (v2/): PASS (1 pre-existing warning, 0 errors)
• No hardcoded IPs, hostnames, or credentials in the HTML file (verified via grep)
• Manual review: all four activity states (absent/gentle/moderate/active) produce distinct animation amplitudes

Notes

• Pure frontend addition — zero impact on sensing pipeline, signal processing, or Rust crates
• COCO-17 joint definitions documented inline in HTML for future maintenance
• Activity-state animations tuned for adult human bodies; non-human subjects would render unnaturally (acceptable trade-off, noted in subtitle disclaimer)