feat: adapter — run InternLM/WildClawBench (60 in-the-wild OpenClaw tasks) under the ClawBench harness

Goal

Let the ClawBench harness load and execute tasks from InternLM/WildClawBench (60 original in-the-wild tasks; agents run in a live OpenClaw personal-assistant environment; multi-step tool orchestration + multimodal evidence; arXiv 2605.10912; HF internlm/WildClawBench; strongest model ~62%) so users can:

1. Run a long-horizon, tool-heavy agent eval through our 5-layer trace pipeline + interception schema, without standing up the OpenClaw stack separately.
2. Share traces in our existing *Trace HF dataset format so comparisons across benchmarks live in one place.

Scope

• Environment bridge — package the OpenClaw runtime as an optional container layer (Dockerfile.openclaw-style) the harness can boot per-batch. Reuse our existing harness-isolation pattern.
• Task loader — clawbench.corpus.adapters.wildclawbench ingests their tasks/ directory + HF metadata; maps their agency vs multimodal axes onto our category enum.
• Tool-call accounting — their tasks chain 10–60+ tool calls; our actions.jsonl already captures these but we need a counter exposed in run-meta.json to match their reported metrics.
• Scoring passthrough — wire to their checker code in eval/ so leaderboard-comparable scores come out of the same run.
• CLI — clawbench run --corpus wildclawbench --model <m> end-to-end.

Why now

Different axis from ours (long-horizon agency + multimodal vs. our single-request interception). Demonstrating that our harness can host both makes the case for a common evaluation infrastructure across the agent-benchmark ecosystem.

Acceptance

• clawbench run --corpus wildclawbench --limit 5 --model <m> boots OpenClaw, runs 5 tasks, produces 5-layer trace bundles.
• Scores within ±3pp of the upstream eval/ pipeline on a sampled subset.
• Docs: eval/adapters/wildclawbench.md + OpenClaw container README.

Start by reading the existing harness-isolation pattern and corpus adapter structure, then inspect the WildClawBench tasks/ and eval/ code referenced in the issue. Trace how actions.jsonl and run-meta.json are produced before planning the OpenClaw container bridge. Done means the --limit 5 command runs five tasks, emits trace bundles and comparable scores, with the two named documentation files added.