feat(ingest): describe PDF figures with vision model

[scaleborg]

Summary

• Enable include_image_base64=True in Mistral OCR and describe significant figures using Gemini 2.5 Flash
• Figures (>2% page area) get text descriptions appended as [Figure description – page N] blocks before chunking
• Parallel vision calls (4 workers), graceful degradation if no Gemini key, non-PDF paths unaffected

Verified on

Test	Result
PDF with figures (IRPAPERS, 24pp, 7 figures)	91 chunks, 6 chunks contain figure descriptions, 8 concepts + 13 tools extracted
PDF without figures (SVD_Notes)	27 chunks, zero vision calls, no overhead
Markdown file (AGENTS.md)	8 chunks, no regression
Before/after text size	91,064 → 95,339 chars (+4.7%)
Vision cost	~$0.01 per PDF (~1.3s/image, Gemini Flash)

Example stored chunk

[Source: 2602.17687v1]

[Figure description – page 2]
This diagram illustrates a multi-modal approach for "Retrieval and QA Analysis"
of "IRPAPERS". It processes input papers through three distinct pathways:
"Page Image" leading to "Multi-Vector Image Embeddings", "OCR Transcription"
leading to "BM25" and "Single-Vector Text Embeddings". All these embedding and
retrieval methods converge into the final "Retrieval and QA Analysis" stage.

Test plan

• API ingest of PDF with figures — chunks include descriptions
• API ingest of PDF without figures — no vision calls, no overhead
• API ingest of non-PDF file — no regression
• Learnings extraction works with figure-enriched text
• Ruff lint passes (no new violations)
• Pre-commit hooks pass