vet: add --dangerous-accept and sandboxed sub-agent vetting
Three changes to vit vet:
1. --dangerous-accept (human only): permanent project-wide vet bypass.
Writes .vit/dangerous-accept flag file. Two-step with --confirm.
Writes .vit/.gitignore to prevent committing the flag.
2. Agent error: when an agent runs plain vit vet <ref>, new error
explains the sandboxed sub-agent pattern and suggests --trust --confirm.
3. --trust --confirm: bypasses requireNotAgent() for sandboxed sub-agents.
--confirm alone does nothing. Must be paired with --trust.
Ref argument changed from <ref> to [ref] (not needed for --dangerous-accept).
Tests cover: dangerous-accept with/without confirm, agent blocking,
duplicate gitignore, --trust --confirm agent/human, --confirm without --trust.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>