fix: add --dangerously-skip-permissions for CI/CD environments

[melanie531]

Problem

In CI/CD environments (CodeBuild, GitHub Actions, etc.), Claude Code has no user configuration (~/.claude/settings.json). Without permission bypass, Claude Code prompts for confirmation before executing Bash commands.

In -p (print) mode, this causes tool calls to be silently skipped — the agent never executes Bash commands like python3 scripts/check.py, resulting in:

• Agent cannot run skill scripts → functional score is artificially low
• With-skill and without-skill runs produce nearly identical results
• Functional scores stuck around 28-36% regardless of skill quality

Evidence

Same skill (acme-compliance) tested in two environments:

Environment	Permission Config	Functional Score
EC2 (has ~/.claude/settings.json)	skipDangerousModePermissionPrompt: true	90%
CodeBuild (fresh install, no config)	Default (prompts for permission)	31%

The 59-point gap is entirely explained by Claude Code's inability to execute Bash commands in the CI/CD environment.

Fix

Add --dangerously-skip-permissions to both _build_cmd_with_skill() and _build_cmd_without_skill() in ClaudeRunner.

This is safe because:

1. Eval runs in isolated temp workspaces (no access to real user data)
2. The agent is already sandboxed by --allowedTools
3. CI/CD environments are ephemeral containers
4. This matches Claude Code docs: "Recommended only for sandboxes with no internet access"

Testing

• 652 tests pass (1 updated to verify the flag is present)
• No changes to public API or CLI interface