Visual Deck Builder

Create high-quality visual PPT decks from whatever the user supplies: a topic, a document, a paper, a folder of materials, a style reference, or existing slide images. The main workflow is spec first, page-description driven, full-slide image first, then image-only PPTX packaging:

input -> intent/source resolver
      -> generate_slide_spec.py: LLM outline (title + page_role + banana 9-field layout_profile per slide,
                                cover = title/subtitle/presenter only)
                                 -> LLM per-page description (density-floor body_text + layout_profile + extra_fields)
      -> visual_style_contract.json from the preset library (exact hex/fonts/percentages)
      -> independent Stylist pass refines extra_fields against the style contract (aesthetics only, no semantic change)
      -> banana-style per-slide image prompt (演讲者备注 excluded, anti-duplication/anti-floating-text guards) -> full-slide image
      -> slide 1 image locked as shared style-anchor reference for slides 2..N
      -> deck.json -> image-only PPTX
      -> preview + render manifest + QA report (content density/page-role diversity, style-contract compliance, formal-PPT-feel, template-overfit axes)

The deck is not a fixed "paper mode" or "topic mode". Route by the user input, but the default final slide is a complete full-slide image inside PPTX. Use layered editable reconstruction only when the user explicitly asks for editable PPT objects and accepts the extra QA loop.