Add a metadata-injection mutation to the CLI challenge

[Victoria824]

Why

The web playground demonstrates untrusted glossary text attempting to disable policy checks. The CLI mutation suite should produce the same release-blocking evidence.

Acceptance criteria

• Add a deterministic metadata-injection mutation in src/queryassure/challenge.py
• Treat retrieved descriptions and glossary text as data, never policy instructions
• Emit a named failure gate and sanitized evidence
• Add a focused regression test
• Update the challenge documentation and expected mutation count

The test must not call a live model provider.