Add opt-in drawing PDF views and Markdown reports

[Çağdaş Yürekli (cagdasyurekli)]

Vector drawing PDFs can return successful Markdown containing only isolated labels, while SHX annotation text and all drawing lines are omitted. Add an explicit, local --pdf-drawing-mode / pdf_drawing_mode=True path that returns located native/SHX labels, raster page views, and an English Markdown limitations report. Existing conversion defaults remain unchanged. This draft is a concrete proposal for review; core/plugin placement and API/report shape have not been accepted by maintainers.

CLI: markitdown drawing.pdf --pdf-drawing-mode -o drawing.md produces drawing.md, drawing.report.md, and drawing.assets/page-1.png, etc. New destinations only; collisions, symlinks, and source overwrites are refused. Output is staged, published through exclusive hard links, and rolled back on failure. Destination filesystems must support hard links. --conversion-report selects a different report path. Stdout requires explicit --keep-data-uris plus a report path.

The API exposes result.assets and result.conversion_report without filesystem writes. ZIP members retain distinct image names and reports; failed drawing rendering is not silently skipped or downgraded to successful text-only conversion. The optional MCP argument returns inline views and the report, with no file writes. Drawing mode runs before PDF plugin/cloud fallback selection; CLI cloud-mode combinations are rejected.

This is not lossless or semantic drawing reconstruction. PNG views use the CropBox at up to 200 DPI with an 8192-pixel longest-edge cap; fine detail can be lost. Reports give MediaBox/CropBox bounds and identify unresolved information without claiming exhaustive detection. No OCR, dimension associations, unit/scale inference, or external model calls. Angled lettering can have imperfect word grouping. The complete assets map is held in memory, and large drawings can be resource-heavy. Clients must provide image pixels to visual consumers explicitly.

Validation (Python 3.12):

• Initial new drawing-mode regression set: 8 failed / 5 passed on the previous implementation; the completed set has 25 passing cases. A separate stream-position regression failed before its correction and passed afterward.
• Final core suite: 1,075 passed, 57 skipped. OCR: 108 passed. MCP: 22 passed. Black 23.7.0 and git diff --check passed. OCR/MCP suites run separately because their test package names collide in a combined invocation.
• Covers native and SHX text, geometry-only pages, 90/180/270-degree text, rotated/cropped multipage files, decodable images, escaping, plugin bypass, missing dependencies, rendering failure, file/report/source collisions, symlinks, staging rollback, stream position, and archive assets/reports/failures.
• Tested seven public PDFs / all 56 pages: five direct CAD exports and two processed drawing documents. No third-party PDFs or assets are committed. Page-view output was visually reviewed on representative drawings and a synthetic fixture.
• Image sizing preceded the default choice: 23-page AutoCAD export at 200 DPI = 75,335,775 PNG bytes / 100,447,732 base64 bytes; dense Revit page = 3,696,835 / 4,929,116 bytes. External assets therefore remain default. Full conversion produced approximately 0.89 MB AutoCAD Markdown and 5 KB Revit Markdown; process high-water mark across the corpus was about 1.3 GB.

Scope: net +338 production lines, +342 test lines, +59 documentation lines. Growth adds drawing extraction, safe output publication, and integration/regression coverage. No new runtime/test dependencies. Independent from the opt-in rejection/warning PR #2668. Related overlapping proposals: #1398 (bbox JSON/OCR) and #2055 (embedded image extraction).

Refs #2688. This is a separate RFC/draft contribution; it does not imply maintainer approval of core placement or sidecar behavior.