feat(agent): S3-backed Claude SDK session store

**Context:** ROADMAP.md → S3-backed SDK session store (portable transcripts)

---

## Component

Agent (Python runtime)

## Describe the feature

Plumb the Claude Agent SDK **`SessionStore`** to S3 (dedicated bucket or prefix) with:

- Eager flush, IAM-scoped access, conditional part creates, checksums, adaptive retries, structured logging
- Metrics/alarms on transcript mirror failures
- Graceful shutdown (`disconnect` on SIGTERM/cancel) so in-flight frames flush
- Persist **`task_id` ↔ Claude session UUID** (from first `ResultMessage`) for resume on another worker
- Keep agent **`cwd` stable** so SDK-derived `project_key` paths stay predictable
- Plan **compaction** when part count threatens resume latency
- Optional **S3 Express One Zone** when fleet is single-AZ

Complements shipped **Persistent session storage** (`/mnt/workspace` FUSE caches) and end-of-task **trace** upload to `traces/...jsonl.gz`.

## Use case

In-VM caches do not survive worker migration or long pauses. Portable transcripts enable resume, debugging, and compliance without relying solely on end-of-task trace upload.

## Proposed solution

1. S3 prefix per `task_id` + SDK `project_key`; SessionRole scoped writes.
2. Flush on frame boundary; `disconnect` on shutdown.
3. Alarms on mirror failures; structured logging.
4. Compaction when part count threatens resume latency.
5. Optional S3 Express One Zone when fleet is single-AZ.
6. Map first `ResultMessage` session UUID to `task_id` in DynamoDB for cross-worker resume.

## Other information

- Shipped: persistent session storage on `/mnt/workspace`; execution tracing to `traces/...jsonl.gz`.
- Design context: `docs/design/COMPUTE.md`, `docs/design/OBSERVABILITY.md`.

- [ ] This might be a breaking change

Commencez par ROADMAP.md et le contexte de conception dans docs/design/COMPUTE.md et docs/design/OBSERVABILITY.md. Localisez ensuite les points d’entrée d’upload fournis pour /mnt/workspace session storage et traces/...jsonl.gz. C’est terminé lorsqu’une implémentation approuvée couvre la mise en miroir portable de SessionStore, le mappage des sessions entre workers, le flush à l’arrêt, les alarmes de défaillance, les tentatives de nouvelle exécution et la compaction.

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