TPP: rebuild the interactive panel around a…

Details

Complete makeover of the scaling-laws panel (split into tpp-view.ts +
tpp-widget.ts). Fixes the core "nothing changes when I switch" problem and adds
the motion + free navigation the old one lacked.
Root cause of the bug: the (N,D) grid is ~identical across corpora — only the
LOSSES differ. The old plot encoded N·D position, so switching corpus/law/epoch
barely moved anything. The new plot encodes LOSS: measured loss (y) vs training
tokens D (x, log), each dot a run coloured by model size N, on FIXED axes so a
switch reads as points sliding to new heights.
- Every control now visibly changes the picture (verified by canvas pixel-diff):
  corpus → losses slide; epoch → points shift along D; law → the fitted iso-N
  curves change; design (new Collinear/Non-collinear chip) → the whole cloud +
  fit swap, and the held-out error whiskers + R² change (e.g. RedPajama/Kaplan
  CO 0.51 → NC 0.91).
- Animated transitions: points are keyed by physical run (design|N|D) and tween
  between states over ~0.7s, so corpus/epoch/design/phase changes MOVE the dots.
  Snaps under prefers-reduced-motion (canvas stays interactive, no auto-play).
- Fit & extrapolate phase draws the law as faint per-size curves with held-out
  rings + red prediction-error whiskers; coverage phase keeps the live in-browser
  refit + V_K/τ_K regime lamp + draggable window.
- Any combination reachable at any time without losing your place; choices
  persist to localStorage and restore on reload (returning visitors land on
  their state, not the tour).
Node tests 16/16, tsc clean. Old tpp-explainer.ts removed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014c7dbCGbQxtWedKRSx4sJi

Files changed (6)