v2.1.0 — visual upgrade (cover restyle + figures)

v2.1.0 — visual upgrade over v2.0.0.

Same methodology, same findings, same paper text. This release adds:

• Cover-page restyle — RESEARCH ARTICLE eyebrow, subtitle, version line, license footer. Removes email and personal website from page 1 for cleaner citation aesthetics. Matches the format used by my companion rPPG cardiac preprint for consistent author identity.
• 3 new figures embedded in §3.4 (multi-seed cross-lingual audit):
  • Figure 1 — Per-seed UAR forest plot with pooled-prediction 95% bootstrap CIs and the 4-class chance baseline. Shows at a glance that Urdu → Urdu sits below chance, Urdu → Punjabi sits at chance with a tight CI, and only Pun → Urdu clears chance.
  • Figure 2 — Per-class recall heatmap across the four conditions. Visualises the arousal-biased transfer pattern hypothesised in Observation 7.
  • Figure 3 — Pooled confusion matrices (2×2 grid) for the four conditions. The Urdu → Punjabi panel makes the single-class collapse claim of Observation 6 immediately visible.
• Appendix A reproducibility statement — replaces the [TODO-ZENODO-DOI] placeholder with the actual Concept DOI (10.5281/zenodo.20709692) and v2.0.0 version DOI (10.5281/zenodo.20709693).
• scripts/generate_figures.py — reusable figure-builder, reads existing results JSONs, regenerates all three figures in ~10 seconds. Fully reproducible.

The four findings (acoustic feature collapse, multilingual transformer dominance on English, multi-seed transformer cross-lingual audit with three structural failure modes) are unchanged from v2.0.0. Cite the Concept DOI (10.5281/zenodo.20709692) — it always resolves to the latest version.

Artifacts in this release

• paper_v2.pdf — the rebuilt preprint with figures
• figures/*.png and figures/*.pdf — the three new figures as standalone files

Reproducibility

python scripts/generate_figures.py regenerates all figures from the existing per-seed result JSONs in results/multiseed_cross_lingual/. No new runs required.