# Full-sync shared contracts

This package establishes the boundary for later queue accounting, archive,
replay, and HTTP work. It does not change running service behavior. Production
code in this package does not import the queue package; compatibility tests
compare its models and keys with the existing queue implementation.

## Compatibility and counters

- `RunMeta` retains all existing JSON fields and adds optional fields. Existing
  `zset.RunMeta`, `zset.Task`, and `zset.FailureRecord` remain unchanged; adapters
  are a later integration step. The new `RunStatus` is a named string type.
- `Validate` accepts missing extension fields on old records.
  `ValidateAccounting` is opt-in for runs with known extended counters and
  enforces `enqueuedCount = successCount + failedCount + activeCount + nextRoundCount`.
  `inflightCount` is a subset of `activeCount`.
- Source cleanup changes `pendingFailureCount`, never historical `failedCount`.
- Round statistics have no task/card/channel dimensions. `ErrorCounts`
  normalizes both JSON input and output to eight canonical categories, merging
  unknown names into `other` and rejecting negative/overflowing totals.
- Existing failure payloads retain their free-form classification and message.
  The HTTP `Failure` projection normalizes categories and bounds messages to
  2,048 UTF-8 bytes. Historical blocked tasks may have round zero.

## Status domains

| Domain | Wire values |
| --- | --- |
| Run | `producing`, `running`, `completed`, `completed_with_failures` |
| Sync type | `normal`, `replay` |
| Archive | `disabled`, `pending`, `processing`, `succeeded`, `failed` |
| Source replay disposition | `none`, `pending`, `processing`, `completed`, `needs_attention` |
| Replay action | `accepted`, `processing`, `cleanup_pending`, `receipt_pending`, `completed`, `needs_attention` |
| Error category | `timeout`, `rate_limited`, `authentication`, `validation`, `upstream`, `transport`, `internal`, `other` |

The source field is `replayDisposition`, distinct from the replay action's
`status`. Lease tokens, record versions, action revision, idempotency key, and
request fingerprint are internal and excluded from HTTP JSON.

## Interfaces and persistence obligations

HTTP consumers depend on `RunReader`, `FailurePager`, `ReplayRequester`, and
`ActionReader`. Separate live/archive readers support fallback only after a live
`ErrNotFound`; storage failures must remain visible. An archived summary cannot
provide per-card details: use `ErrFailureDetailsUnavailable` when live details
are gone. Repositories and HTTP both apply the default page limit 100 and
maximum 500. Cursors remain opaque and are bound to the original query. An empty
scan page with a next cursor does not signal completion.

Archive discovery, scheduling, snapshot loading, and result registration use
separate narrow interfaces. Registration is conditional on an opaque work
version. Archives contain run and round aggregates only. Archive receipts use
a schema version, storage-owned path, lowercase SHA-256 checksum of stored
bytes, and publication time. Archive schema version starts at 1.

Replay acceptance atomically reserves `(sourceRunId, idempotencyKey)` with a
SHA-256 fingerprint of the source and exact requestedBy body value. Repeating
the request returns its original action/target; changing the request returns
`ErrConflict`. The action fixes its target ID before target creation/enqueue.

Replay work discovery, lease management, progress updates, source pagination,
target creation/enqueue, and source cleanup are separate interfaces. Progress
updates are fenced and use a revision check. Target enqueue distinguishes
durably present duplicates from rejected writes; a zero insert count by itself
does not authorize source cleanup. Cleanup checks exact source record versions
and target acceptance evidence. Implementations must persist enough evidence
to recover partial batches; a scan cursor by itself is insufficient. Source
pagination must remain correct while already-enqueued entries are deleted.

`ErrNotFound`, `ErrConflict`, `ErrInvalidArgument`,
`ErrFailureDetailsUnavailable`, `ErrLeaseLost`, and `ErrBusy` support
`errors.Is` through wrapping. HTTP error envelopes and status mappings remain
the HTTP implementation's responsibility.

## Redis naming

All names are prefixed with the configured namespace (`flow:` by default;
`sim-status:` for status sync). Existing run/round/task-state/failure names are
preserved. New keys include:

| Suffix after namespace | Scope |
| --- | --- |
| `{runId}:round:N:stats` | Run slot |
| `{runId}:archive` | Run slot |
| `runs:all` | Global index |
| `archive:pending` | Global index |
| `{replay}:action:actionId:meta` | Replay slot |
| `{replay}:action:actionId:lease` | Replay slot |
| `{replay}:idempotency:SHA256(sourceRunId,key)` | Replay slot; tuple encoded as a JSON array before hashing |
| `{replay}:pending` | Replay slot |

Replay action creation, idempotency, and discovery can be atomic within their
shared slot. Source/target runs and global indexes can occupy different slots;
their coordination requires recoverable registrations/receipts rather than a
single cross-slot Lua transaction. New identifier validation rejects hash-tag
or path delimiters without rewriting legacy JSON on decode.

## HTTP surface

| Method | Route | Response |
| --- | --- | --- |
| GET | `/internal/v1/full-sync-runs?syncKind=&limit=` | `RunList` merged from active Redis Runs and `run-<runId>.json` archives |
| GET | `/internal/v1/full-sync-runs/{runId}` | `RunView` with `source`, `run`, optional `rounds` |
| GET | `/internal/v1/full-sync-runs/{runId}/failures?cursor=&limit=` | `FailurePage` with `items`, `nextCursor` |
| POST | `/internal/v1/full-sync-runs/{runId}/replay` | HTTP 202 `ReplayAccepted` |
| GET | `/internal/v1/full-sync-actions/{actionId}` | `ReplayAction` |

POST requires `Idempotency-Key` and JSON `{"requestedBy":"..."}`. Authentication,
HTTP handlers, storage, replay execution, archive files, and service wiring are
intentionally left to subsequent implementations.

## Verification

```text
gofmt -l internal/fullsync/contract
go test ./internal/fullsync/contract
go test ./internal/queue/zset ./internal/consumer/zset ./internal/producer/simstatus ./internal/producer/monthlyusage ./internal/service/simstatus ./internal/service/monthlyusage
```

Tests cover legacy schema and JSON compatibility, existing key compatibility,
cluster key boundaries, bounded categories/messages, overflow, run accounting,
pagination limits, status domains, replay identity, internal field exclusion,
and receipt/progress validation. Regression tests use the repository's isolated
Redis helper and require `redis-server` or `REDIS_SERVER_BINARY`.