feat(memory): memory extraction replay operator API

**Context:** ROADMAP.md → Memory security → Memory extraction replay
**Related:** #27

---

## Component

API or orchestration

## Describe the feature

Operator API (e.g. `POST /memory/extraction-jobs` or `start_memory_extraction_job`) to **re-run failed PR-review extraction** after webhook or Lambda errors.

## Use case

Transient failures during review-comment extraction leave gaps in learning. Operators should replay without re-running the full agent task.

## Proposed solution

1. Durable job record: PR URL, task_id, last error, status.
2. Idempotent replay from stored webhook payload or GitHub API fetch.
3. CLI: `bgagent memory replay-extraction `.
4. Link outcomes to #27 memory loop metrics.

## Other information

- Design context: `docs/design/MEMORY.md`, `docs/design/ORCHESTRATOR.md`.

- [ ] This might be a breaking change

Commencez par docs/design/MEMORY.md et docs/design/ORCHESTRATOR.md, puis suivez les points d’entrée existants de l’extraction de mémoire et de l’orchestration. Définissez l’enregistrement de job durable, le chemin de rejeu idempotent, l’API opérateur et la CLI `bgagent memory replay-extraction`, en reliant les résultats aux métriques de #27.

Rédigé par le modèle d'indexation à partir du texte de l'issue.