User request ("show me a ridgeline plot of scores by domain")
  |
  v
1. COMPOSE — Generate D3/Plotly/matplotlib code from primitives
  |         (marks: area, line, rect; scales: linear, band, time;
  |          encodings: x, y, color, opacity)
  |
  v
2. RENDER — Execute in isolated sandbox, capture output
  |         Did it produce valid SVG/HTML without errors?
  |         Is the output non-empty and reasonable dimensions?
  |
  v
3. EVALUATE — Score the result:
  |   a) Render success (0/1) — did D3 throw?
  |   b) Visual fidelity — does it have axes, labels, data marks?
  |   c) Intent match — does it answer what the user asked?
  |   d) Distance-aware — are fonts 18px+, strokes 2px+?
  |
  v
4. ITERATE — If score < threshold:
  |   - Diagnose: missing scale? wrong mark type? data shape mismatch?
  |   - Adjust: swap marks, fix encodings, add missing axes
  |   - Re-render (max 3 iterations)
  |
  v
5. PROMOTE — If score >= 0.85 after iteration:
     - Save to gallery/ as reusable preset
     - Update d3_catalog.py: NOT_YET -> D3_INLINE (or appropriate backend)
     - Generate /create-figure CLI command registration
     - Learn to /memory for future recall

/memory recall "visualization composition for <type>"

# From a user description
./run.sh compose "ridgeline plot of scores by domain" --data sample.json

# From a d3_catalog type name
./run.sh compose --type ridgeline --data sample.json

# Interactive iteration
./run.sh compose --type streamgraph --data sample.json --iterate --max-rounds 3

# Score an existing composition
./run.sh evaluate gallery/ridgeline_v1.html --intent "show distribution overlap"

# Batch evaluate all gallery items
./run.sh evaluate-all --gallery ./gallery/

# Promote a tested composition to a /create-figure command
./run.sh promote gallery/ridgeline_v1.html --name ridgeline --family distribution

# Dry-run to see what would change
./run.sh promote gallery/ridgeline_v1.html --name ridgeline --dry-run

# List all compositions in the gallery
./run.sh gallery

# Show details of a composition
./run.sh gallery show ridgeline_v1

# Delete a failed experiment
./run.sh gallery delete ridgeline_v0

{
  "name": "ridgeline",
  "version": 1,
  "family": "distribution",
  "description": "Overlapping density plots showing distribution differences across groups",
  "data_shapes": ["distribution", "categorical"],
  "min_data_points": 20,
  "max_data_points": 10000,
  "min_dimensions": 2,
  "max_dimensions": 3,
  "keywords": ["ridgeline", "joy plot", "density", "distribution", "overlap"],
  "backend": "d3_inline",
  "d3_modules": ["d3-shape", "d3-scale", "d3-axis"],
  "canvas_compatible": true,
  "composition": {
    "marks": ["area"],
    "scales": {"x": "scaleLinear", "y": "scaleBand", "color": "scaleSequential"},
    "encodings": {"x": "value", "y": "group", "opacity": 0.7}
  },
  "template_html": "<!-- self-contained D3 code -->",
  "test_data": [{"group": "A", "value": 42}, {"group": "B", "value": 37}],
  "scores": {
    "render_success": 1.0,
    "data_marks": 1.0,
    "axes_labels": 0.9,
    "intent_match": 0.95,
    "distance_aware": 1.0,
    "overall": 0.97
  },
  "promoted": true,
  "promoted_at": "2026-02-25T12:00:00Z",
  "iterations": 2,
  "created_at": "2026-02-25T11:30:00Z"
}

gallery/
├── ridgeline_v1.json       # Metadata + scores
├── ridgeline_v1.html       # Self-contained D3 output
├── streamgraph_v1.json
├── streamgraph_v1.html
├── beeswarm_v2.json        # v2 = iterated improvement
├── beeswarm_v2.html
└── _failed/                # Failed experiments (for learning)
    ├── ridgeline_v0.json
    └── ridgeline_v0.html