[Desktop] Add advisory mode for Claude Desktop

[teseo]

Kind

implement

Goal

Provide safe advisory processing of human.md for Claude Desktop without claiming hard enforcement.

Non-goals

• Pretending Claude Desktop is Claude Code.
• Blocking desktop actions without a verified hook surface.
• Capturing or transmitting private conversation content.

Invariants

• Advisory messages identify that enforcement is not guaranteed.
• No credentials or conversation content are persisted.
• The policy engine remains the single source of policy truth.
• Upgrading to hard enforcement requires a separate verified adapter.

Behavioral contract / States

When Claude Desktop can load project guidance, expose the human.md policy and a clear advisory status. When it cannot be loaded, report unsupported/manual mode. Never report a deny as if it prevented an action.

Scope surface

Documentation/integration guidance, optional helper command, diagnostics, and manual smoke tests.

Acceptance

• Test project guidance discovery and missing-guidance behavior.
• Verify every advisory message includes the enforcement limitation.
• Run: python3 -m pytest tests/test_guard.py -v.
• Add a support matrix entry marking Claude Desktop as advisory/manual.

Split triggers

• Claude Desktop exposes a supported deterministic hook or plugin API.
• A user request requires hard prevention rather than advisory guidance.

Dependencies / Parent

Blocked by: #1
Parent Epic: #15