fix(security): resolve CodeQL findings

[FenjuFu]

Summary

• parse uploaded SKILL.md frontmatter with a bounded SnakeYAML SafeConstructor
• replace the polynomial placeholder regex with linear marker matching
• restrict public label security matching to GET and retain default CSRF protection
• validate CLI callbacks as credential-free HTTP loopback URL objects, discard supplied fragments, and remove sensitive debug logging
• use cryptographic UUIDs for E2E account identifiers
• add YAML tag, label security-chain, and CLI callback regression tests

Code scanning coverage

Addresses alerts #1, #2, #4, #5, #6, #7, #8, and #9. The CLI callback is intentionally limited to HTTP localhost, 127.0.0.1, and IPv6 loopback; the remote CodeQL result will confirm whether the analyzer recognizes the returned validated URL object.

Validation

• backend app reactor: 773 tests passed, 1 skipped
• frontend unit tests: 695 passed
• frontend typecheck: passed
• frontend lint: passed
• frontend production build: passed

Full staging and Playwright E2E were not run locally because they require the complete service stack; PR CI provides the real-services E2E job.