v2 phase 3: 4 figures via figures:scientific-figure composer

[neuromechanist]

Summary

v2 Phase 3: build 4 publication-ready figures by actually invoking the figure-pipeline skills. Closes #11.

Skill invocations (in order)

Skill	What it did	What it didn't do
/figures:scientific-figure	Loaded the 5-step pipeline (Plan -> Build -> Compose -> Validate -> Export). Provided the Cell journal spec (174mm two-column, 6pt minimum body font). Provided compose.py (svgutils-based composer at exact mm coords), validate_fonts.py, export.py	Doesn't author panels; that's delegated to sub-skills
/figures:svg-figure	Authored 4 hand-authored SVG panels under panels/: fig1_matrix.svg (4-perspective x 15-theme matrix), fig2_plot.svg (naturalness gradient plot), fig3_gap_matrix.svg (8 x 4 gap table), fig4_predictions.svg (4 perspectives x 5 columns)	None
/figures:transparent-icons	Generated 10 transparent PNG icons via Codex gpt-image-2: 6 stimulus thumbnails (gratings, photographs, Heider-Simmel triangles, abstract animation, character animation, live-action film) + 4 brain topography icons (occipital, central rolandic, left frontotemporal, occipital-and-frontal). Each at 256x256 with a shared theme.json for consistency	None
/figures:figure-qa (agents)	Couldn't get Bash permission; QA run inline via check_svg.py + visual Read; reports under qa/	Programmatic QA via agent blocked by sandbox; same checks performed inline
Composer	compose.py configs/figN.json -o figN.svg for each. Single panel each, scale = 174/493 = 0.3529 to map pt-viewBox to mm-canvas	None
Validator	validate_fonts.py figN.svg --journal cell returns 0 issues across all 4 figures (25 + 30 + 85 + 71 = 211 text elements; all >= 6pt)	None

Pipeline lesson (the demo observation)

The skill's example uses mm-viewBox with the claim "<text font-size=9> is 9 pt", but in practice rsvg-convert treats unit-less font-size in an mm-viewBox SVG as user units (= mm). Workaround applied: use pt-viewBox (1 unit = 1 pt) with width/height in mm; the validator's pt-unit assumption then matches the renderer.

A second pipeline lesson: rsvg-convert silently drops external image references (CVE-2023-38633 mitigation), even with absolute file:// URIs. Workaround applied: inline all icons as base64 data URIs in the panel SVGs. Adds ~130 KB per figure but renders portably across rsvg/cairosvg/Inkscape.

Programmatic QA results

After in-phase fixes (off-palette #e7f0fa and #fef0e7 lightened tints in Fig 3 swapped to Okabe-Ito hex with fill-opacity="0.10"):

Fig	Text elements >= 6pt	Off-palette	Bbox overlaps	Status
1 strand map	25/25	0	0	ship
2 naturalness gradient	30/30	0	0	ship
3 gap matrix	85/85	0 (after fix)	0	ship
4 predictions table	71/71	0	0	ship

QA reports saved at manuscript/narrative-review/figures/qa/figN_qa.md.

Output

• figures/fig1-4.svg (final composed SVGs, 174mm wide; Figs 1 and 3 are 130mm tall, Figs 2 and 4 are 100mm tall)
• figures/fig1-4.png (300 dpi PNG exports)
• figures/panels/figN_*.svg (panel sources)
• figures/icons/*.png (10 transparent icons, 256x256)
• figures/configs/figN.json (composer configs)
• figures/qa/figN_qa.md (QA reports)

Test plan

• All 4 skills loaded and applied via the Skill tool
• Panel SVGs validated by validate_fonts.py at Cell 6pt minimum (211 of 211 text elements pass)
• Programmatic checks (check_svg.py) all pass: 0 off-palette, 0 bbox overlaps, 0 arrow-tip issues
• Visual judgment: all 4 read cleanly at 174mm two-column Cell Press width
• Composer (scientific-figure compose.py) produces SVGs at the exact mm dimensions
• Pre-commit (markdownlint, typos, trailing-whitespace, end-of-file, json check) passes
• Phase 4 (paper-review + humanizer) will review prose
• Phase 5 (final assembly) compiles and opens final PR