PPT Rebuild Skill v15

Purpose

Convert image-only PPT pages into editable .pptx files. v15 favors a practical high-fidelity route:

1. Extract or provide a scene graph with text bboxes and visual objects.
2. If an original slide image is available, create a textless backdrop by inpainting text regions.
3. Crop detected photos/icons from the original image by bbox and reuse them as PPT image assets.
4. Overlay all detected text as editable PPT text.
5. For structured pages, render native editable shapes/tables/KPI cards from the scene graph.
6. Run QA for editable text count, text overflow, and text-to-text overlap.

Key Difference From v14

v15 keeps v14's placement-plan renderer and adds stricter delivery rules learned from rendered badcases:

• default macOS Chinese font is PingFang SC for PPT output;
• text measurement is more conservative for CJK line height and width;
• the paradigm_shift_1_vs_2 comparison page is a supported reference scene;
• a rendered preview must be visually checked before delivery, even when JSON QA passes.