Review round 1 fixes: sharpen the "aha", fill the…

Details

Addresses all 8 findings from the 4-lens adversarial review (0 high, 3 medium,
5 low — every lens already voted SHIP; these are polish to the bar):
- [design] Reframe the coverage regime copy: V_K >= tau_K reads as IDENTIFIABILITY
  ("enough diversity to identify the exponents; the fit is well-posed"), not a
  blanket "extrapolation holds". Surface the non-collinear grid's R^2 as a
  reference so the collinear number is consistent between Phase 3 and Phase 4
  (a well-posed collinear fit still trails the grid — collinearity is the cap).
- [design] Rebuild Phase 4: larger centred R^2 bars, the 97.3% seed-paired
  win-rate rendered as a proportion strip (was caption-only prose), and enlarged
  parity plots — stacked right on desktop, side-by-side below on mobile. Fills
  the previously-empty climax panel; responsive title/labels (no clipping @390).
- [design] Drop the blue "fitted law" legend entry (no persistent on-canvas
  referent — it was only a transient progress bar); fix the SCSS comment.
- [a11y] Keyboard brush can now narrow/retract each edge ({ }) as well as widen
  ([ ]), reaching every window the pointer can; add aria-label + aria-keyshortcuts
  on the card; stop rewriting the aria-live caption on every pointermove (canvas
  tracks live via rAF; caption updates on drag end only).
- [paper] Correct the BigTPP note: paper caps holdouts at TPP 20 (k in {10..15}),
  not "~1000"; regenerate wikipedia_bigtpp.json.
- [tests] Widen the procedure fidelity test to an ill-scaled law (redpajama
  Kaplan, params 1e3..1e14) so an optimizer regression there is caught.
Node tests still 16/16 green; tsc clean; re-validated the changed states across
desktop/mobile/light 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)