Reframe AgentKernelArena for A/B experiments and improve agent setup

[irvineoy]

Summary

• Reframe AgentKernelArena as a controlled A/B experimentation platform with RL-ready compilation, correctness, and performance signals.
• Refresh the README, Sphinx documentation, task/agent guides, security notes, held-out workflow, and visualization copy to match the current repository.
• Fix onboarding issues with the real clone URL, Node.js/npm prerequisites, native and npm Claude Code installations, config-aware agent checks, and explicit physical-GPU matching guidance.
• Align task-validator/result-schema guidance and remove stale or inaccurate task and HIP optimization instructions.
• Add Docker-runner regression coverage for config-selected agent checks, npm-installed Claude Code, explicit AGENTS=all, and internal argument forwarding.

Validation

• git diff --check
• bash -n src/scripts/docker_benchmark.sh tests/test_docker_benchmark.sh visualization/setup.sh
• make check-evaluator (9 tests)
• make check-perf-helpers
• Python compile check for modified Python modules
• YAML parsing for modified configuration and result-template files
• Local Markdown-link validation across all modified documentation files
• Make target dry-runs for default and explicit docker-check-agents selection

[sinarafati-amd]

The new global default run config is a 60-task Cursor benchmark. A first-time user who runs make docker-run with no CONFIG will kick off a large Cursor run. maybe lets try to change default