Add caption-free visual-body crops and review gate

[917Dhj]

Summary

• generate caption-free Figure/Table visual-body crops while retaining nearby internal scientific labels and a small safety margin
• require model-led review_pending resolution with 300 dpi crop, full-page preview, exact SHA-256 binding, and one bounded normalized-bbox repair
• fail grounding and Formal Save closed for unresolved, malformed, failed, stale, or byte-mismatched insert approvals

Closes #17.

Validation

• /opt/anaconda3/bin/python3 -m pytest -q -ra — 820 passed, 1 skipped
• focused extraction/decision/bundle/grounding/save suites — 211 passed
• ruff check --select F,I on all changed Python scripts — passed
• git diff --check origin/main...HEAD — passed
• independent Standards and Spec reviews — no remaining actionable findings

Real-paper visual smoke (uncommitted)

M3-Agent Figure 1

• before: the crop included the external Figure 1 ... explanation block
• after: the complete memorization/control architecture, long-term-memory graph, internal labels, and arrows remain; the external caption and surrounding prose are absent
• 300 dpi crop bbox: [66.560, 434.138, 545.135, 639.463]
• caption starts at y=644.528, leaving a visible separation margin
• final gate: insert / pass, zero omitted scientific elements

Attention Is All You Need Table 2

• after: all model rows, BLEU headers/values, training-cost headers/values, and table rules remain; the external caption is absent
• 300 dpi crop bbox: [130.671, 93.071, 481.324, 246.217]
• caption ends at y=92.071, leaving a 1 pt separation margin
• final gate: insert / pass, zero omitted scientific elements

The two source PDFs, page previews, and generated screenshots remain outside the repository.