feat: harden format-aware scanning and Skill-Inject benchmarking

[charliechenye]

Summary

• Add bounded logical source spans for wrapped and malformed content.
• Preserve raw source hashes, evidence, line locations, and existing output contracts.
• Resolve wrapped script references safely within known directories.
• Add opt-in --format-aware scanning for scan, check, and diff.
• Add local-only Skill-Inject differential benchmarking with controls and regression gates.
• Add mal-format fixtures and documentation.

Validation

• 40 focused tests passed.
• Ruff checks passed.
• Formatting checks passed.
• Snapshot checks passed.
• npm tests passed.
• Skill-Inject benchmark remains at 77/84 cases with any signal and 14/84 with high/critical signal.

Safety and compatibility

• No payloads, agents, Docker containers, or network connections are used by the benchmark path.
• Existing physical line scanning remains the default.