# TASK 2.2: Add Idempotency, Checkpoint, and Full-Refresh Behavior

## Summary

Add the operational behavior that makes the synthetic feed safe to rerun.

This task adds minimal checkpointing and/or deterministic full-refresh support. It does not implement production replay, DLQ, or backfill tooling.

## Dependencies

- TASK 2.1 (tenant-resolved Redshift curated writer)

## Files to Create

### 1. Checkpoint/full-refresh support

Create services or methods needed to:

- read synthetic feed checkpoints,
- update checkpoints only after successful writes,
- run a deterministic full refresh for the configured tenant/store.

Exact placement depends on the metadata abstractions created in Redshift Foundation.

## Files to Modify

### 2. Feed service orchestration

Extend the synthetic feed orchestration to:

- create/update ingestion batch skeleton records where available,
- use checkpoint values for incremental sync,
- support full-refresh mode,
- update checkpoints only after success,
- include row counts in the feed result.

## Required Behavior

Minimum acceptable behavior:

- deterministic keys:
  - `order_id` for orders,
  - `order_item_id` for order items,
- rerunning the same source window does not duplicate rows,
- updated source rows are reflected in Redshift,
- failed feed runs do not advance checkpoints,
- full-refresh mode is tenant-scoped and source-store-scoped,
- full-refresh mode does not affect unrelated tenants or stores.

Recommended checkpoint identity:

- `source_type = synthetic-kds`
- `checkpoint_key = store:{store_guid}:orders`
- `checkpoint_key = store:{store_guid}:order_items`

Recommended batch metadata:

- `ingestion_mode = synthetic-feed`
- `source_type = synthetic-kds`
- `tenant_id = configured target tenant`
- `tenant_database = resolved target database`
- `raw_s3_path = null`
- `load_s3_path = null`

## Acceptance Criteria

- [x] Feed can run safely more than once for the same source rows
- [x] Feed can update Redshift rows when source rows change
- [x] Feed supports deterministic full-refresh mode
- [x] Full refresh is scoped to the selected tenant/store
- [x] Incremental checkpoint behavior exists or a clear full-refresh-only mode is documented
- [x] Checkpoints advance only after successful Redshift writes
- [x] Failed runs do not advance checkpoints
- [x] Batch metadata is recorded where the slice-01 skeleton is available
- [x] Batch metadata does not require S3 paths
- [x] Solution builds without errors

## Notes

- This task is intentionally minimal operational hardening.
- Do not build production replay, backfill, DLQ, or S3 raw retention here.
- Idempotency matters more than sophisticated scheduling in this slice.