feat(orchestration): stranded-task operator notify and manual resume

**Context:** ROADMAP.md → Stuck-task reconciliation (operator notify/resume)

---

## Component

API or orchestration

## Describe the feature

Extend shipped stranded-task reconciler: **operator notification** before forced failure and **manual resume** option for recoverable stuck states.

## Use case

Reconciler today fails stuck tasks automatically—correct for hygiene but harsh when root cause was transient platform outage. Operators want warn-then-act and resume.

## Proposed solution

1. Grace period: emit `stranded_warning` event + notify before `failTask`.
2. `POST /tasks/{id}/resume` for operator-forced continuation from last checkpoint (when safe).
3. Blueprint config: `stranded_grace_minutes`, `notify_channel`.
4. Audit all operator resume actions.
5. Tests for warning → fail and warning → resume paths.

## Other information

- Base reconciler shipped; this is **operator UX** on top.
- Design context: `docs/design/ORCHESTRATOR.md`.

- [ ] This might be a breaking change

Start with the “Stuck-task reconciliation” section in ROADMAP.md and read docs/design/ORCHESTRATOR.md for the existing stranded-task reconciler. Trace the current failTask path and the task API entry points before defining the warning, resume, configuration, and audit flows. Done means the warning→fail and warning→resume paths have tests and the proposed operator behavior is covered.

Written by the indexing model from the issue text.