Blame for docs/FORMAT.md

# Format 5 reference

This document describes the persistent format used by the NYARION page engine.
The standalone release number and the image format number are separate: release
7 uses format 5 images. The engine is a POSIX regular-file storage engine, with
all recovered logical data held in RAM. It is not an NTFS image or a mounted
filesystem.

## Image layout

All offsets are byte offsets from the start of one regular file. Multi-byte
binary header fields use big-endian order.

| Region | Location or size |
|---|---|
| Four catalogue seal copies | offsets `0`, `4096`, `8192`, `12288`; each uses one 512-byte unit |
| Reserved seal region | bytes `0` through `16383` |
| Mirrored capsule window | starts at `16384`; two arrays of 64 slots, each slot 65,536 bytes |
| Mirrored object arena | starts at `8404992`; `pool_blocks` physical blocks of 4,096 bytes |
| Exact apparent image length | `8404992 + pool_blocks * 4096` |

The arena size is configured at creation. `pool_blocks` must be an integer from
16 through 1,048,576. The default is 16,384. The value counts physical blocks,
including both replicas of each object; it does not count logical objects or
file-data bytes. The apparent file size does not establish filesystem-allocated
space or physical SSD write traffic.

For logical capsule epoch `e` and replica `r` in `{0, 1}`, the slot offset is:

```text
16384 + (r * 64 + e % 64) * 65536
```

For object block index `i`, the offset is `8404992 + i * 4096`.

## Record headers and identity

The magic is the eight-byte string `NYPAGES5`. Every record starts with a
104-byte header. The first 72 bytes form the checksummed prefix:

| Field | Size | Meaning |
|---|---|---|
| magic | 8 bytes | `NYPAGES5` |
| format | 2 bytes | integer `5` |
| kind | 2 bytes | typed record identity |
| payload length | 4 bytes | bytes immediately following the header |
| epoch / birth generation | 8 bytes | meaning depends on kind |
| volume UUID | 16 bytes | image identity |
| payload SHA-256 | 32 bytes | digest of the exact payload bytes |
| prefix SHA-256 | 32 bytes | digest of the preceding 72 bytes |

Kinds are seal `1`, catalogue `2`, trie `3`, blob manifest `4`, raw data `5`, and
capsule `6`. Seal headers carry a catalogue generation. Arena objects carry
their birth generation. Capsule headers carry a logical transaction epoch.
Epochs and generations are unsigned 64-bit integers, with exhaustion rejected.

An arena object has at most 3,992 payload bytes and is written as one padded
4,096-byte block per replica. A seal is padded to 512 bytes, leaving at most 408
payload bytes. A capsule payload is bounded by 65,432 bytes; capsule writes need
not overwrite the unused tail of the reserved slot. The parser follows the
validated length rather than treating old tail bytes as part of the record.

SHA-256 is used for accidental-corruption detection and content binding under
the stated collision-resistance/ideal-detection assumption. This is not an
authenticated storage format against an attacker who can rewrite records and
their hashes. Both replicas reside in the same image; this alone does not
provide whole-drive failure independence.

## Canonical structured payloads

Seals, catalogues, tries, blob manifests and capsules use compact ASCII JSON:
sorted keys, no insignificant spaces, comma/colon separators, ASCII escaping,
and no non-finite numbers. Duplicate object fields and noncanonical encodings
are rejected on parsing. Raw data pages contain bytes rather than JSON.

Binary values in JSON use standard padded base64. Decoding is strict and must
round-trip to the same canonical encoding. Volume identifiers are 32 lowercase
hexadecimal characters; digests are 64 lowercase hexadecimal characters.

## Object descriptors

A descriptor is a six-element JSON array:

```text
[replica_a_index, replica_b_index, birth_generation, kind, payload_length, sha256]
```

Replica indices must be distinct and within the configured arena. Following a
descriptor checks the expected kind, volume, birth generation, payload length
and digest. A current object may refer to an older immutable child, but not a
child born after the selected catalogue generation. The catalogue itself must
have the selected seal's generation as its birth generation.

This identity distinguishes a reused address from the object that formerly
occupied it. It does not authorize early reuse: blocks reachable from a root
that remains protected cannot be allocated merely because a later pointer
would detect that overwrite.

## Seal and catalogue

A seal has exactly these fields:

```json
{
  "generation": 7,
  "volume": "0123456789abcdef0123456789abcdef",
  "pool_blocks": 16384,
  "root": [20, 21, 7, 2, 123, "<64 lowercase hex characters>"]
}
```

The root shown above is explanatory; the length and digest must match an actual
catalogue for a valid image. A catalogue contains `format`, `volume`, `epoch`,
`history`, `files_root`, and `tokens_root`. An empty trie root is JSON `null`.
The catalogue's logical epoch identifies a complete global checkpoint.

Seal copies are grouped into banks `(0, 1)` and `(2, 3)`. New generation `g`
publishes to bank `g % 2`; creation initially writes all four seal positions.
Recovery elects the greatest complete valid seal generation, rejects conflicting
same-generation identities, and validates that generation's dependency graph.
Missing dependencies for that selected root stop recovery. The engine does not
silently try an older generation to hide that failure.

## Trie and file representations

Files and retry receipts occupy two separate immutable adaptive 16-way tries.
Routing uses successive hexadecimal digits of SHA-256 over the UTF-8 key.
Leaves keep complete keys. A leaf includes `family`, `type: "leaf"`, `depth`,
and an `entries` map. A branch includes `family`, `type: "branch"`, `depth`,
and a `children` map whose single-character keys are hexadecimal digits and
whose values are descriptors. Depth cannot exceed 64.

Leaves split when their encoded payload would exceed one object page. An
oversized bucket at exhausted routing depth is rejected. Deletion removes
entries and empty children; globally minimal tree shape is not guaranteed.

A file entry is either `{"inline": "<base64>"}` for at most 256 bytes, or
`{"blob": <descriptor>}`. A blob manifest contains `length`, `digest`, and a
nonempty `chunks` list. Its raw data chunks have the expected bounded lengths,
and their concatenation must match the manifest's total length and digest.

A token entry is `[logical_epoch, intent_sha256]`. Each distinct acknowledged
token remains live logical data; the format has no implicit token expiration.
Its intent digest binds a sorted list of paths and either a deletion marker or
the replacement value's length and digest.

## Capsule and recovery tail

A capsule contains exactly `format`, `volume`, `epoch`, `parent`, `token`,
`intent`, and `changes`. `changes` is a nonempty sorted list of
`[path, base64_value_or_null]` pairs, with no duplicate paths. `null` denotes a
deletion; an empty base64 string denotes an empty byte value. The parent is the
preceding global history digest.

After loading a global checkpoint, recovery requires newer complete capsules
to form a contiguous, history-linked tail within the 64-slot window. It checks
all evidence from valid newer capsule headers. A checksum-valid newer header
without a complete valid payload blocks recovery, even if that interrupted
operation was never acknowledged. Fully checkpointed slots may be reused.

## Publication and allocator state

An ordinary fitting transaction writes both complete capsule copies and waits
for one successful `fsync` before returning its epoch. A transaction that must
publish a checkpoint also includes the checkpoint's persistence operations.

A checkpoint writes and validates new dependencies while keeping the old root
protected, forces those dependencies, writes both new seals, and forces their
publication. Only then may it release blocks absent from the successor graph.
Writable recovery repairs selected reachable mirrors and normalizes the newest
seal bank before another publication.

The free heap, descriptor ownership, reference counts, verified edge cache and
temporary reservations are RAM state. They are rebuilt from the selected disk
graph after recovery. Positive reference counts represent the retained roots'
reachable graph; changed roots activate and release edges at zero/one
transitions. The checkpoint path avoids a whole-arena rescan. Invalid reachable
graphs, cycles and conflicting block ownership stop recovery before allocation.

Release 8 uses volatile capacity admission evidence before acknowledging a
new capsule. An incrementally maintained conservative bound covers changed
published-leaf cells, shared dirty ancestors, blob pages and the new catalogue.
Each changed cell uses a weighted persistent symbolic trie. Leaves maintain
canonical encoded-entry weight sums and counts; branches retain subtree page
counts. Offers copy changed symbolic paths and bounded leaf maps, and unchanged
symbolic subtrees are shared. Splitting follows the hash partition; retaining
an existing branch after a shrink can conservatively charge extra pages.
Empty symbolic branches are removed.

Symbolic new descriptors use maximum format widths. If the bound fits, exact
compilation can be deferred until checkpoint; if it cannot certify fit, the
complete exact planner is used before accepting or rejecting the offer.
An exact plan binds actual free addresses and encoded dependency/seal bytes
to the published roots, generation and logical history. Planning uses private
overlays and a nonmutating free-heap cursor. Publication consumes that exact
allocator prefix, forces dependencies and then forces the new root seals.

Reservation cells, weighted symbolic nodes, bound totals, route caches and exact plans are process-time
admission state, not new persistent fields. Writable restart recompiles an exact
plan for the recovered tail before accepting more changes. Read-only recovery
can expose an older tail whose checkpoint cannot fit. These rules guarantee
conditional engine-arena headroom, not host filesystem allocation, available
RAM or successful device I/O. An older compatible writer need not enforce the
release 8 admission policy merely because it writes the same image format.

The bounded fault argument assumes truthful persistence barriers, one serialized
publisher, disjoint replicas, detectable loss of at most one protected sector
in addition to pending torn writes, and repair/drain conditions before a later
writer. Two durable newest seal witnesses keep a stale bank from being elected
within that loss budget. Older bank graphs may already contain reclaimed pages.
Losing every newest witness is outside that promise. Finite model tests do not
qualify a physical device's power-loss behavior.

## Compatibility boundary

Release 8 retains format 5 rather than adding a new disk-format version for its
volatile admission and planning rules. Compatibility tests and their exact
source hashes are provided with the release. Format 3 and format 5 are different
formats; no in-place format-3 migration is implied. The on-disk format describes
data representation and recovery; process-time admission guarantees also depend
on which engine version is allowed to write the image.