Apply the two-panel review to the EDCR explainer

Details

Paper-fidelity panel (verdict was NO_SHIP, all copy-level):
- Recover caption/comment no longer credit the paper with "100% precision;
  recall grows with data" — that figure is from the released PyEDCR
  example (ImageNet50); the paper claims majority-recovery with graceful
  degradation under noise. Both now say exactly that.
- "held-out labels" → "training labels" (rules train on the same split as
  the base model, per the paper), and the CIKM method is named f-EDR
  (EDCR is the original framework, whose correction rules the widget
  doesn't show) in the lede, captions and fallback.
- Disclose that the demo uses only cross-granularity conditions (the full
  method adds secondary-/binary-model conditions).
- "rebuild the taxonomy" → "recover the constraints of a taxonomy";
  honest-limit annotation scoped to "no cross-grain rule can see it".
Design-caliber panel (verdict was NO_SHIP):
- Touch was dead: pointerleave cleared hover before click on tap. The
  click handler now hit-tests its own coordinates; on touch, first tap
  inspects and a second tap flips (Learn phase) — desktop keeps
  hover-inspect / click-flip.
- Autoplay plays once and lands on the finale instead of looping (and
  Predict/Learn durations grew ~20% so captions can be read).
- Figure/caption integrity: wide stages now show ALL 16 training cards
  (two rows of eight), so "16 samples, 6 wrong" matches what's on screen;
  narrow stages keep 6 and the caption says so.
- Each rule's F1 renders next to the rule head, not right-flushed across
  the canvas; long rules wrap at the conjunction on narrow stages.
- Caption discipline: dropped the self-praising "The widget doesn't
  pretend otherwise" and the "ratio of two submodular functions"
  name-drop; added the missing hover-to-inspect hints; Recover caption
  now calls out live when a flipped label produces an exclusion that
  contradicts the true tree.
- Legend trimmed to the three colours actually on stage; dead neural-blue
  token removed; caption reserves height so controls stop jumping;
  reduced-motion fallback constrained to 60ch; mobile Recover graph
  cleared from under the sticky header and type raised to 10.5px.
22/22 core tests still pass; all five phases re-screenshotted across
dark/light, desktop/mobile, and reduced-motion.

Files changed (5)