Optimize the homepage — design + clarity pass (ref: ProgramBench, PaperBench, AutoLab, MLSBench)

[reacher-z]

Optimize the claw-bench.com homepage (design + clarity pass)

Bring the homepage up to the bar set by peer benchmark sites. Reference designs to draw from:

• ProgramBench — leaderboard-first hero, tight header CTAs (Leaderboard / Paper / GitHub / Team), task cards, inline metric tooltips.
• PaperBench (OpenAI) — clean academic framing, crisp "what/why/how" above the fold, rubric transparency.
• AutoLab — clear task taxonomy + reproducibility surfacing.
• MLSBench — scannable results presentation and methodology section.

Concrete asks (consolidated from the competitive audit)

1. Responsive/mobile layout — the stylesheet currently has essentially one @media (reduced-motion) and no width breakpoints; the 8-col leaderboard, hero stat string, and CTA cards don't degrade on narrow screens. Add 768/600/480px breakpoints with leaderboard column-priority hiding (mirror ProgramBench's .th-full/.th-mobile).
2. One source of truth for headline numbers — task/site counts and top score currently disagree across hero, README, HF card, and OG image. Pick one canonical triple (e.g. "V1: 153 · V2: 130 · N live sites") + one top-score-with-metric, and propagate everywhere. (Ties to Data integrity: published V2 leaderboard top rows lack backing runs (README opus-4-7 44.6% vs META.md glm-5.1 18.5%) #221 / Shall we default all the datasets and traces on HF to v2? #216.)
3. Above-the-fold messaging — lead with a one-line "what ClawBench is" + the version badge (V2 default) before the leaderboard, like PaperBench/ProgramBench.
4. Leaderboard presentation — drill-down per model row + a tasks×models view (see feat(leaderboard): 2-axis view — fixed-harness (compare models) + fixed-model (compare harnesses) #157, Leaderboard clarity: add tasks×models accuracy chart + per-model drill-down on row click #217); inline tooltips explaining intercepted vs reward/judged.
5. Hero demo — a short trace-replay gif/carousel of an agent run (we have real trace screenshots ready).
6. Methodology + reproducibility sections surfaced (judge design, two-stage scoring) — match MLSBench/PaperBench transparency.

Full competitive analysis (ClawBench vs ProgramBench/deep-swe across 10 dimensions) is available on request.

cc @Perry2004 — assigning for the homepage revisions.