protectskills/MaliciousAgentSkillsBench: v0.5

Dynamic-execution portability.

• The run's exit code now comes from Claude rather than the syscall tracer, so best-effort tracing never determines the run's success.
• Syscall tracing defaults off on macOS (in-container ptrace is unreliable in the Docker Desktop VM); Linux keeps it on. Override with USE_STRACE.
• README documents host deps: Claude Code install/login, GNU timeout (coreutils), and jq (optional triage).
• analysis/verify_dataset.py: stdlib check that prints the released dataset counts.