2.4 KiB
Local world manifest v1
This is an M5.7 single-host publication record, not a production/world-server manifest. The encoding is fixed-width little-endian and exactly 88 bytes:
| Offset | Width | Field |
|---|---|---|
| 0 | 8 | magic 0x0031464d574f5645 |
| 8 | 4 | version 1 |
| 12 | 4 | total length 88 |
| 16 | 8 | world identity |
| 24 | 8 | frame identity |
| 32 | 8 | generation, nonzero |
| 40 | 8 | FNV-1a checksum of complete snapshot payload |
| 48 | 8 | snapshot command high-water |
| 56 | 1 | journal-present flag, 0 or 1 |
| 57 | 7 | reserved zero bytes |
| 64 | 8 | FNV-1a checksum of complete journal payload, or zero |
| 72 | 8 | journal final command, or zero |
| 80 | 8 | FNV-1a checksum of bytes 0 through 79 |
FNV detects accidental corruption only. It is not authentication. Payload names are derived from their checksum and complete bytes are compared when a name already exists. A manifest reference is accepted only after the M5.6 snapshot restore and optional all-or-nothing journal replay succeed.
Publication and recovery
The writer holds the per-world OS lock. It writes a same-directory temporary payload and flushes it, installs a new immutable payload name, then flushes the directory. It writes and flushes the inactive manifest temporary and atomically replaces that manifest slot. The replacement is the commit point. Slots alternate by generation. Recovery parses both slots and selects the greatest generation; any malformed present slot or invalid selected payload fails closed. It never silently falls back from a valid published pointer to an older generation.
stateDiagram-v2
[*] --> PreparePayload
PreparePayload --> PayloadSynced
PayloadSynced --> PayloadInstalled
PayloadInstalled --> PrepareInactiveManifest
PrepareInactiveManifest --> ManifestSynced
ManifestSynced --> Published: atomic replace (commit)
ManifestSynced --> PreviousGeneration: failure before replace
Published --> ReopenRequired: ambiguous/post-publication error
Published --> Recoverable: durable slot + referenced payload
An interruption before replacement recovers the old generation. An interruption after replacement recovers the new generation. Unknown files and orphan payloads are ignored; this version does not delete them. OS locks, fsync/write-through and same-directory rename semantics depend on the local filesystem and storage stack.