MCP backend

Use when: Landmark validation, feature-point deltas, control-point repair, face/eye/smile alignment, leaf tip alignment, aura ring fitting, iterative recipe parameter updates.

Before the first Blender call, read ../text-to-blender/references/mcp-backends.md. Prefer official MCP for shared capabilities; preserve the selected backend for the whole task. Tool names in recipes are client/legacy examples: map them through that reference and its agent-host guide.

Landmark Fit Repair

Bbox and IoU are too coarse for final mascot/logo matching. This skill turns named feature drift into recipe edits.

Landmark classes

• structural_tip: outer part tips and extrema;
• shell_corner: face/shell/rim corners;
• feature: eyes, smile, brows, decals;
• ring: aura center/radius/dots;
• depth_marker: side/back/top thickness markers.

Workflow

1. Define a landmark JSON for the source reference.
2. Detect or render corresponding product landmarks.
3. Compare by name and view.
4. Convert deltas into repair actions: move boundary control point, scale one component, shift feature curve, tune depth, or adjust aura radius.
5. Rebuild from recipe; do not hand-edit final geometry without updating the recipe/manifest.

Hard gates

• No final export when named structural landmarks exceed tolerance.
• Face features must be validated separately from body silhouette.
• Aura/context landmarks must not affect base GLB acceptance.

Script

• scripts/landmark_fit_report.py compares two named-landmark JSON files and emits repair deltas.

Landmark JSON schema

{
  "schema": "landmarks.v1",
  "image_size": [width, height],
  "landmarks": [
    {"view":"front", "name":"leaf_top_tip", "class":"structural_tip", "x":627, "y":130}
  ]
}

For part masks, generate initial landmarks from extrema: top/bottom/left/right, centroid, and contour inflection points. Then rename them to semantic names before repair.