[M6] Human-escalation stub + safe abstain/stop policy

## Why
High-risk paths should escalate or stop rather than silent retry.

## Scope
- Stub escalation event in traces
- Stop/abstain action with risk gate

## Acceptance criteria
- [ ] Stub + tests
- [ ] Documented in `docs/recovery-policy.md`

## References
- Roadmap: `docs/roadmap.md`
- Related vertical slice: `experiments/results/tool-semantic-drift-001.md`
