Update: ClawBench entry — V2 corpus (283 tasks, 163 platforms)

[reacher-z]

ClawBench is already in the Benchmarks & Research section, but the blurb
is from the V1 release. This PR refreshes it to reflect the V2 corpus
that shipped last week.

What changed since the entry was added:

• Task count: 153 → 283 (V1 153 + V2 130)
• Platforms covered: 144 → 163
• Scoring: was just HTTP-request interception; now a two-stage
  pipeline — Stage 1 final-request interception, Stage 2 LLM-judge
  reviews the intercepted body. This addresses the "agent reached the
  endpoint but submitted the wrong thing" failure mode.
• Public leaderboard at https://claw-bench.com with 6 tabs by corpus ×
  harness pairing.
• All eval runs come with a 5-layer execution-trace bundle (recording,
  action log, request log, agent messages, interception JSON) — useful
  for anyone who wants to audit, debug, or build on top of the runs.

No other entries touched.

[nibzard]

Merged — thanks for the V2 update! Two tiny follow-ups whenever you have a minute (no rush, happy to take them as a separate PR):

1. The repo now lives at TIGER-AI-Lab/ClawBench — the old reacher-z/ClawBench URL still redirects, but updating the README link (and the stars badge path) would make it canonical.
2. Small mismatch between the entry ("163 platforms") and the upstream README ("144 websites"). Whichever number is right, would be great to have them agree.