fix(convert): make PDF→PPTX produce editable slides

[nelsonduarte]

Problem

User report: ao converter PDF→PPTX, os slides ficam como uma única imagem, não como texto editável.

The previous _convert_pptx rasterised each PDF page at 2× DPI and embedded the resulting PNG filling the whole slide. PowerPoint opened the file but every slide was a flat picture — no selectable / editable text.

Fix

Replace the rasterised path with a structured extraction over page.get_text("dict").blocks:

• Image blocks (type=1) → slide.shapes.add_picture at the block's bbox (so scanned/image-only PDFs still work).
• Text blocks (type=0) → slide.shapes.add_textbox at the block's bbox; each line becomes a paragraph; each span becomes a run carrying its own font size, bold, italic, and RGB color (extracted from PDF span attributes: size, flags & 16, flags & 2, color).
• Internal textbox margins zeroed so the textbox bbox matches the PDF bbox more faithfully.
• Reuse the auto-created empty run on the first paragraph for the first span; add new runs for subsequent spans (avoids the stale empty run that paragraph.text = "" leaves behind).

PDF and PowerPoint both use top-left origin so no Y flip; PDF points → EMU is × 12700.

Test plan

• Smoke test on Ubuntu 26.04 + Py3.14.4: a 3-page PDF with a heading + two body lines per page now produces a 30 KB PPTX with 9 textframes (3 per slide) and 0 image shapes. Before the fix, the same input produced ~60 KB of rasterised PNGs per slide. Sample run output:

  [info] textframes total: 9
  [info] images total: 0
  [slide 1] ['Heading 1', 'Body paragraph on page 1.', 'Line two of the body.']
  [slide 2] ['Heading 2', 'Body paragraph on page 2.', 'Line two of the body.']
  [slide 3] ['Heading 3', 'Body paragraph on page 3.', 'Line two of the body.']
  

• Live test in PowerPoint: convert a real text-rich PDF, verify text is selectable and editable, fonts respect bold/italic, no rasterised page background remains.
• Live test with a scanned/image-only PDF: image blocks should still embed as picture shapes; the slide won't be editable text but at least the existing image fidelity is preserved.

🤖 Generated with Claude Code