Add human-reviewed model-backed behavioral probes

[askmy-stack]

Summary

Extend offline probes with an optional model-backed execution path, while retaining deterministic offline checks as the default.

Acceptance criteria

• Human-reviewed probe format and approval workflow are documented
• Model-backed execution is explicitly opt-in
• Probe results record selected tool, arguments, outcome, and errors
• Secrets and untrusted tool metadata are handled safely
• Offline probe behavior remains backward compatible
• Fixtures cover positive, negative, and ambiguous probes