Fast path: pdf-inspector for text-based PDFs, Marker only when needed

[odfalik]

Adds a CPU-only fast conversion path so agents (and laptops) can ingest papers without loading Marker's ML models.

What it does

• fast_convert.py: classifies the PDF with pdf-inspector (pure Rust, no models). Text-based/mixed PDFs with clean encodings convert to markdown in ~100ms; figures are dumped with PyMuPDF into the usual images/ folder.
• convert_pdf_auto() routes: fast path first, automatic fallback to Marker for scanned/image-based PDFs, encoding issues, or any failure. process_paper now calls it, so behavior is unchanged for PDFs that need OCR.
• Output layout, metadata and hashes are identical; metadata gains engine and pdf_type.

Why
On the opendataloader benchmark pdf-inspector scores 0.875 overall / 0.814 TEDS on tables, ahead of pymupdf4llm and markitdown, at ~35x the speed. Verified locally: 'Attention Is All You Need' classified text_based (confidence 1.0), converted in 111ms, 3 figures extracted, on an 8-CPU box with no GPU.

Note: pdf-inspector does not extract figures itself — that's why PyMuPDF handles images on this path.

[odfalik]

Ran the head-to-head against the existing library before merging this — results argue for re-scoping the PR.

Compared 5 papers already in R2 (their paper.md was produced by Marker on GPU) against pdf-inspector on the same source PDFs, on an 8-CPU box:

Paper	Pages	pdf-inspector	Old chars	New chars	Old table rows	New table rows	Old math	New math
CLAM_2004.09666	35	0.13s	91,789	88,586	0	9	42	0
CONCH_arxiv	57	0.32s	145,688	129,908	292	255	22	0
CellViT	23	0.28s	159,147	110,591	141	114	458	0
DNABERT-2	23	0.09s	110,733	78,701	317	222	58	0
GigaTIME_Cell_2025	35	1.93s	106,214	99,435	32	16	68	0

Speed is 100-1000x better, but: no LaTeX at all (equation fragments bleed into prose in two-column papers), fewer table rows on 4/5 papers with complex tables merged, and PyMuPDF figure extraction is unreliable (155 fragments vs 9 real figures on CellViT, 0 on DNABERT-2 whose figures are vector).

Proposal: don't make this the default path for process_paper. Keep Marker canonical, expose pdf-inspector as an explicit quick_convert + use classify_pdf as an OCR router. I'll push that revision unless someone prefers the current shape.