Add interactive scaling-laws explainer section +…

Details

A new v3 section "Interactive · scaling laws" placed right after the EDCR
explainer. The canvas widget (tpp-explainer.ts) is driven by the real fitter:
- Phase 0 shows the collinear-fan vs non-collinear-grid designs on a log-log
  (N, D) plot with live counts.
- Phases 1-2 run the actual multi-start L-BFGS-B + DE search in-browser (restart
  counter + best-SSE ticking) and draw a measured-vs-predicted parity plot;
  captions report the live collinear vs non-collinear holdout R² (reproducing
  the paper's Fig-1 pair, e.g. RedPajama/Kaplan 0.51 vs 0.91).
- Phase 3 is a draggable tokens-per-parameter coverage brush: the V_K >= tau_K
  regime lamp recomputes live on every drag and the window refits in-browser —
  a single ray reads Regime A with negative holdout R², a wide window reads
  Regime B and extrapolates. The "aha", computed not scripted.
- Phase 4 contrasts the two designs and attributes the paper's headline (97.3%
  win rate, CI) plus per-corpus honest-limit notes.
Follows the EDCR pattern: global v3tpp* scope, theme-reactive canvas, DPR clamp,
hover-inspect, keyboard, reduced-motion fallback, IntersectionObserver lazy
start, dataset/law/mode/seed/window chips, fetch of the shipped JSONs. Wired via
index.md, v3.html (optim -> core -> explainer), v3.scss, and the backdrop-filter
fallback list. Validated across dark/light x desktop/~390px x reduced-motion
with zero page errors.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014c7dbCGbQxtWedKRSx4sJi

Files changed (7)