v2.0.0 — the post-benchmark milestone

The release shaped by the project's own preregistered studies. Six phases, merged as #61–#66:

• Token diet (bands A+B): ~56 KB cut across the corpus with zero obligations lost; the always-loaded core dropped ~600 tokens/session.
• Three Level A holes closed (Page Title, Language, Motion Actuation) + the ABNT NBR 17225 typography floor (80ch line cap, typographic rhythm, start-aligned text).
• The Brazil layer: NBR regular/plena ↔ Standard/Shield mapping, the real Annex A critical-items list, the functional-performance section (Annex B / EN 301 549 / VPAT), a new sign-language guide (Hand Talk/VLibras mechanics, the Libras window, the avatar line), and guide-navigation rebuilt as Navigation & Document Structure.
• The agentic web, in the right order: built for people — agents benefit as a corollary. New guide, ARIA-for-agents named inside ARIA Soup, and the Machine-Only Content Doors anti-pattern.
• What the benchmark ordered: tools/contrast-check.py (deterministic contrast — computed, never estimated), the Palette Definition Protocol with its §2.1 trigger, the Half-Climbed ARIA Ladders anti-pattern with a declared kill criterion, double governance named in the Definition of Done, and the interactive-states inventory in REPORT.md.

Net: obligations 62 → 71 per edition; cost per task within +1.5% of the 1.8.0 floor. Full detail in the CHANGELOG.

This tree is what the round-2 benchmark will freeze by hash.