Fix content_scanner.py: only require SKILL.md frontmatter on SKILL.md itself

[ankitsingh015]

Summary

Directly implements the review I flagged earlier ("Unreviewed external content" — content_scanner.py exists exactly for this, so 150+ external skill files shouldn't be trusted without actually scanning them). Cloned uphiago/recon-skills and ran content_scanner.py against its full 145-SKILL.md tree.

Found a real bug in the scanner itself: scan_skill_file() required YAML frontmatter on every .md file it was pointed at, not just files literally named SKILL.md. That produced 27 false positives (all references/*.md and top-level docs like README.md/AGENTS.md, which legitimately have no frontmatter — only SKILL.md is the frontmatter contract, matching this repo's own .claude/skills/<name>/{SKILL.md, references/} layout).

Fixed by scoping the frontmatter-specific checks to os.path.basename(path) == "SKILL.md" — the content-safety checks (prompt-injection phrasing, hidden Unicode, base64 blobs) still apply to every .md file regardless of filename.

Scan result after the fix: 5 HIGH + 1 MEDIUM, all manually verified as legitimate — every "prompt-injection" hit is inside a skill that specifically hunts for prompt injection (hunt-llm-ai, llm-prompt-injection, hunt-mcp-security, etc.), so it necessarily documents the injection phrasing as an example test payload — the exact same pattern as this repo's own known emerging-surfaces false positive. The one base64 blob is a synthetic ASP.NET __VIEWSTATE example. No genuine malicious content found.

Test plan

• 2 new regression tests: a references/*.md without frontmatter is no longer flagged; the same file still gets flagged for genuine content issues (prompt-injection phrasing)
• All 8 existing content_scanner tests still pass (all use SKILL.md as the filename, so the fix doesn't change their behavior)
• Full 205-test suite passes
• Re-ran the real external-repo scan before/after to confirm the false-positive count dropped from 27→0 and the 6 remaining findings are genuine (verified by hand, cited above)