feat(log): v2 is the LIVE write format — envelope records with minted ints, genesis, sector seals; v1 readable forever

The cutover: new tail segments write format v2 (per-record [type, version,
cipherFlag, keyId] envelope; noun/verb after-images carry dense ints
MINTED AT APPEND from the id mapper — a rebuilt mapper reproduces
assignments exactly; log.genesis opens every new log with the id-space
width + a minted brain id; sync() seals to the header-declared sector
boundary with reader-invisible pad frames). Existing v1 segments are
never rewritten — per-segment decoder dispatch reads both formats and v2
facts map to the exact CommitFact shape all consumers already read.
Cutover on a live v1 log: an empty v1 tail re-heads in place; a non-empty
one is sealed by rotation, byte-identical. Records reserve the encryption
fields (cipherFlag 0 / keyId nil are the only legal values; anything else
refuses typed naming the needed newer reader) — crypto-ready with no
future bump on the compat surface. Empty-records facts are legal (an
all-deduped batch is a real generation — v1 semantics preserved; the
refusal there tore a column-store flush mid-commit in the full suite, the
consistency guard caught it loudly, and the root is fixed).

Golden byte vectors pinned for the second (native) reader implementation.
Pins: cutover 5/5 · codec 54 · kill-matrix stays 11/11.
This commit is contained in:
David Snelling 2026-08-10 10:55:11 -07:00
parent 73eb88d481
commit 26c6025158
6 changed files with 1372 additions and 135 deletions

View file

@ -41,10 +41,57 @@
* terminal-readable) is the single source of truth for the segment SET;
* rotation flips it atomically (write-new fsync rename) BEFORE the new
* tail's first byte exists, so no segment file is ever unaccounted for.
*
* ## Mixed-version logs (the v2 live-write cutover)
*
* The segment header's `formatVersion` selects the decoder PER SEGMENT:
* v1 segments (ops-shaped facts, the format above) stay readable forever and
* are NEVER rewritten; a NEW tail segment writes the v2 format
* (`src/db/factLogFormat.ts` record envelope, minted dense ints, genesis,
* sector seals) whenever the int minter is installed ({@link FactLog.setIntMinter}
* the brain wires it from the metadata index's id mapper right after init).
* A bare `FactLog` with no minter keeps writing v1 (there is no authority
* that could reproduce int assignments, and 0 is never written). Cutover
* mechanics on an existing v1 log: an EMPTY v1 tail is re-headed to v2 in
* place; a non-empty v1 tail is sealed by an immediate rotation and the new
* tail is v2. Decoded v2 facts map back to the SAME {@link CommitFact} shape
* v1 consumers read (noun/verb ops with `{metadata, vector} | null` records)
* the vector wrapper object is reconstructed from the record's metadata leg
* through the reserved-field hydration law (see `commitFactFromV2`).
*
* V2 tails additionally: write the `log.genesis` record (id-space width 64 +
* the brain id, minted once into the manifest's additive `brainId` field) as
* the first record of the FIRST fact of a brand-new log, and seal every
* `sync()` to the header-declared sector size with pad frames that are
* invisible to readers (torn-page defense at group-commit boundaries).
*/
import { encode as defaultEncode, decode as defaultDecode } from '@msgpack/msgpack'
import { crc32c } from '../utils/crc32c.js'
import { prodLog } from '../utils/logger.js'
import {
FACT_LOG_FORMAT_V1,
FACT_LOG_FORMAT_V2,
DEFAULT_SEAL_SIZE,
parseSegmentHeader,
encodeSegmentHeaderV2,
encodeFactV2,
decodeFact as decodeFormatFact,
decodeGroupV2,
encodePadFrame,
minPadFrameBytes,
type CommitFactV2,
type LogRecord,
type EmbedPendingRecord,
type EmbedLandedRecord,
type BlobManifestRecord,
type BootstrapBaselineRecord,
type ProjectionNoteRecord
} from './factLogFormat.js'
import {
splitNounMetadataRecord
} from '../types/reservedFields.js'
import { NounType } from '../types/graphTypes.js'
import { v4 as uuidv4 } from '../universal/uuid.js'
// Swappable msgpack implementation — defaults to the JS codec; a native
// provider (registered via the plugin registry's 'msgpack' key) may replace
@ -65,7 +112,12 @@ export function setFactCodec(impl: {
export const FACTS_PREFIX = '_generations/facts'
/** The facts manifest path (JSON). */
export const FACTS_MANIFEST_PATH = `${FACTS_PREFIX}/manifest.json`
/** Current segment format version (header field; additive-only within a major). */
/**
* The v1 segment format version the MANIFEST's formatVersion gate and the
* header value of v1 (minter-less) tails. NOT the live-write ceiling: new
* tails write `FACT_LOG_FORMAT_V2` (src/db/factLogFormat.ts) whenever the
* int minter is installed; both versions are read forever, per segment.
*/
export const FACTS_FORMAT_VERSION = 1
/** Rotation threshold: seal the tail segment once it exceeds this many bytes. */
const SEGMENT_ROTATE_BYTES = 8 * 1024 * 1024
@ -83,6 +135,29 @@ export interface FactOp {
record: { metadata: unknown | null; vector: unknown | null } | null
}
/**
* V2-native records beyond noun/verb ops that a fact may carry through the
* ENCODER (types 6/7/8/9/10 of the v2 registry: embed markers, blob
* manifests, projection notes, bootstrap baselines). Encoder-ready by
* design; nothing produces them yet the deferred-embed sidecar and blob
* lifecycle remodel onto these records in a later leg.
*/
export type FactMarkerRecord =
| EmbedPendingRecord
| EmbedLandedRecord
| BlobManifestRecord
| ProjectionNoteRecord
| BootstrapBaselineRecord
/**
* Mints the dense integer handle for an entity/verb id at fact-append time
* REQUIRED to be reproducible: a rebuilt id mapper must reproduce the same
* assignments exactly, so the only legal implementation delegates to the
* metadata index's id mapper (`getOrAssign`). Returns a POSITIVE bigint; a
* minter that cannot resolve its mapper throws an int of 0 is never written.
*/
export type FactIntMinter = (kind: 'noun' | 'verb', id: string) => bigint
/** One committed generation, as scanned back out of the log. */
export interface CommitFact {
generation: number
@ -90,6 +165,12 @@ export interface CommitFact {
ops: FactOp[]
meta?: Record<string, unknown>
blobHashes?: string[]
/**
* V2-native marker records riding this fact (see {@link FactMarkerRecord}).
* Optional and additive: absent on every v1 fact and on every fact the
* current writers produce; requires a v2 tail to encode.
*/
records?: FactMarkerRecord[]
}
/** The telemetry a scan batch carries (frozen shape). */
@ -143,6 +224,12 @@ interface FactsManifest {
/** The append target. Its true content is established by scanning (crash tolerance). */
tailSegment: string | null
updatedAt: string
/**
* This brain's stable id (additive, v2 cutover): minted as a uuid at the
* first v2 tail creation and never changed; the `log.genesis` record
* carries it. Absent on logs that have never had a v2 tail.
*/
brainId?: string
}
/** The narrow byte-level storage surface the fact log rides. */
@ -251,40 +338,339 @@ function decodeFact(payload: Uint8Array): CommitFact {
}
}
/**
* Deep-normalize a decoded v2 JSON position (metadata legs, meta maps,
* notes) back to plain-JSON values: the v2 codec decodes msgpack int64/uint64
* as `bigint` (its u64 wire discipline), but canonical records are JSON a
* metadata timestamp like `createdAt: 1786…` must come back as the NUMBER it
* was encoded from. Safe-range bigints narrow exactly; anything beyond the
* safe-integer range in a JSON position refuses loudly (it cannot have come
* from a JSON write).
*/
function normalizeWireJson(value: unknown): unknown {
if (typeof value === 'bigint') {
if (value > BigInt(Number.MAX_SAFE_INTEGER) || value < -BigInt(Number.MAX_SAFE_INTEGER)) {
throw new Error(
`fact log v2: decoded integer ${value} exceeds the JS safe-integer range in a JSON position`
)
}
return Number(value)
}
if (Array.isArray(value)) return value.map(normalizeWireJson)
if (value && typeof value === 'object' && !(value instanceof Uint8Array)) {
const out: Record<string, unknown> = {}
for (const [k, v] of Object.entries(value)) out[k] = normalizeWireJson(v)
return out
}
return value
}
/**
* JSON-serialization equivalence for a v2 ENCODE-side JSON position: drop
* undefined-valued object keys and map undefined array elements to null
* exactly what `JSON.stringify` does when canonical records are persisted.
* Commit facts are built from write-cache-WARM objects that may still carry
* undefined-valued engine keys (`service: undefined`, ) which the durable
* JSON never had; msgpack would preserve them as nil (the v1 capture's known
* wart), so the v2 capture the future storage authority sanitizes to the
* DURABLE truth instead.
*/
function toJsonSafe(value: unknown): unknown {
if (value === undefined) return null
if (Array.isArray(value)) return value.map((v) => (v === undefined ? null : toJsonSafe(v)))
if (value && typeof value === 'object' && !(value instanceof Uint8Array)) {
const out: Record<string, unknown> = {}
for (const [k, v] of Object.entries(value)) {
if (v === undefined) continue
out[k] = toJsonSafe(v)
}
return out
}
return value
}
/** Mirror of the storage layer's stored-timestamp normalization, minus its
* `Date.now()` fallback (a DECODER must be deterministic an unreadable
* timestamp is omitted, and the divergence surfaces via the oracle). */
function reconstructTimestamp(value: unknown): number | undefined {
if (typeof value === 'number' && value > 0) return value
if (
value !== null &&
typeof value === 'object' &&
typeof (value as { seconds?: unknown }).seconds === 'number'
) {
return (value as { seconds: number }).seconds * 1000
}
return undefined
}
/**
* Rebuild a noun's canonical VECTOR-FILE wrapper from a v2 after-image
* the read-side of the hydration law. Canonical noun vector files hold the
* denormalized enumerable entity (`{id, vector, connections, level, type,
* reserved fields, metadata}` — the write path's composition); the v2
* record deliberately carries only the ENTITY state (metadata leg + embedding
* floats), because connections/level are derived HNSW residue with their own
* rebuild paths (empty in every 8.x write) and the denormalized top-level
* fields are projections of the metadata leg. This reconstruction applies
* the SAME split/hydrate law the storage layer uses
* (`splitNounMetadataRecord` the single source of truth in
* src/types/reservedFields.ts; field map mirrors
* `BaseStorage.hydrateNounWithMetadata`, undefined keys omitted exactly as
* JSON serialization omits them), so in the no-drift case the reconstructed
* wrapper digests byte-equal to canonical. A drifted denormalized copy
* surfaces as an oracle `state-differs` named, never silently absorbed.
*/
function reconstructNounWrapper(
id: string,
metadataLeg: unknown,
floats: number[]
): Record<string, unknown> {
const { reserved, custom } = splitNounMetadataRecord(
(metadataLeg ?? null) as Record<string, unknown> | null
)
const wrapper: Record<string, unknown> = {
id,
vector: floats,
connections: {},
level: 0,
type: (reserved.noun as string) || NounType.Thing
}
if (reserved.subtype !== undefined) wrapper.subtype = reserved.subtype
if (reserved.visibility !== undefined) wrapper.visibility = reserved.visibility
const createdAt = reconstructTimestamp(reserved.createdAt)
if (createdAt !== undefined) wrapper.createdAt = createdAt
const updatedAt = reconstructTimestamp(reserved.updatedAt)
if (updatedAt !== undefined) wrapper.updatedAt = updatedAt
if (reserved.confidence !== undefined) wrapper.confidence = reserved.confidence
if (reserved.weight !== undefined) wrapper.weight = reserved.weight
if (reserved.service !== undefined) wrapper.service = reserved.service
if (reserved.data !== undefined) wrapper.data = reserved.data
if (reserved.createdBy !== undefined) wrapper.createdBy = reserved.createdBy
wrapper._rev = typeof reserved._rev === 'number' ? reserved._rev : 1
wrapper.metadata = custom
return wrapper
}
/** Coerce a candidate embedding to `number[]`: plain arrays pass through
* (element-checked); numeric typed arrays (the JS HNSW rebuild path stores
* `Float32Array` vectors on the memory adapter) widen via `Array.from`. */
function floatsOf(candidate: unknown, context: string): number[] | undefined {
if (Array.isArray(candidate)) {
for (const el of candidate) {
if (typeof el !== 'number') {
throw new Error(`fact log v2: ${context} vector carries a non-number element`)
}
}
return candidate as number[]
}
if (ArrayBuffer.isView(candidate) && !(candidate instanceof DataView)) {
return Array.from(candidate as unknown as ArrayLike<number>)
}
return undefined
}
/** Extract the embedding float array from a canonical vector value: a bare
* float array (or numeric typed array) passes through; a wrapper object
* yields its `vector` floats; `null` stays `null`; anything else refuses
* loudly. */
function embeddingLegOf(value: unknown, context: string): number[] | null {
if (value === null || value === undefined) return null
const direct = floatsOf(value, context)
if (direct !== undefined) return direct
if (typeof value === 'object') {
const nested = floatsOf((value as { vector?: unknown }).vector, context)
if (nested !== undefined) return nested
}
throw new Error(
`fact log v2: ${context} has a canonical vector record with no float vector — ` +
`cannot encode its after-image`
)
}
/**
* Map one decoded v2 fact to the {@link CommitFact} shape every consumer
* already reads: noun/verb after-images and tombstones become ops (vector
* wrappers reconstructed see {@link reconstructNounWrapper}); a
* `batch.meta` record becomes `meta` when the fact position carries none;
* `log.genesis` is log-level metadata (its width was verified at decode) and
* is not an op; marker records surface on the additive `records` field so
* nothing is silently dropped. Decoded JSON positions are normalized back
* from the codec's bigint discipline ({@link normalizeWireJson}).
*/
function commitFactFromV2(f: CommitFactV2): CommitFact {
const ops: FactOp[] = []
const markers: FactMarkerRecord[] = []
let batchMeta: Record<string, unknown> | undefined
for (const r of f.records) {
switch (r.type) {
case 'noun.afterImage': {
const metadata = normalizeWireJson(r.metadata) ?? null
let vector: unknown | null = null
if (r.vectorLeg !== null) {
if (!Array.isArray(r.vectorLeg)) {
throw new Error(
`fact log v2: noun.afterImage ${r.id} carries a vector ref — this reader ` +
`resolves inline vectors only (refs are a later leg); refusing`
)
}
vector = reconstructNounWrapper(r.id, metadata, r.vectorLeg)
}
ops.push({ kind: 'noun', id: r.id, record: { metadata, vector } })
break
}
case 'noun.tombstone':
ops.push({ kind: 'noun', id: r.id, record: null })
break
case 'verb.afterImage': {
const metadata = normalizeWireJson(r.metadata) ?? null
if (r.vectorLeg !== null && !Array.isArray(r.vectorLeg)) {
throw new Error(
`fact log v2: verb.afterImage ${r.id} carries a vector ref — this reader ` +
`resolves inline vectors only (refs are a later leg); refusing`
)
}
// The canonical verb vector-file wrapper: endpoints + verb name ride
// as first-class v2 wire fields precisely so this reconstruction is
// exact ({id, vector, connections:{}, verb, sourceId, targetId} —
// verbs carry no `level`).
const vector: Record<string, unknown> = {
id: r.id,
vector: r.vectorLeg ?? [],
connections: {},
verb: r.verb,
sourceId: r.sourceId,
targetId: r.targetId
}
ops.push({ kind: 'verb', id: r.id, record: { metadata, vector } })
break
}
case 'verb.tombstone':
ops.push({ kind: 'verb', id: r.id, record: null })
break
case 'batch.meta':
batchMeta = normalizeWireJson(r.meta) as Record<string, unknown>
break
case 'log.genesis':
break // the log's birth certificate — log-level metadata, not an op
case 'projection.note':
markers.push({ ...r, note: normalizeWireJson(r.note) as Record<string, unknown> })
break
case 'bootstrap.baseline':
markers.push({ ...r, metadata: normalizeWireJson(r.metadata) })
break
default:
// embed.pending / embed.landed / blob.manifest carry no loose JSON maps.
markers.push(r)
break
}
}
const meta = f.meta ? (normalizeWireJson(f.meta) as Record<string, unknown>) : batchMeta
return {
generation: f.generation,
timestamp: f.timestamp,
ops,
...(meta ? { meta } : {}),
...(f.blobHashes && f.blobHashes.length > 0 ? { blobHashes: f.blobHashes } : {}),
...(markers.length > 0 ? { records: markers } : {})
}
}
/** One intact v2 frame's extent inside a segment (byte-slicing support). */
interface V2FrameExtent {
/** Byte offset just past this frame. */
end: number
/** The frame's generation (0 for pad filler). */
generation: number
/** True when the frame is a pad (invisible filler). */
isPad: boolean
}
/** Walk a v2 segment's intact frames (torn-tail terminated), returning each
* frame's extent the byte-level view truncation slices against, so kept
* frames are never re-encoded (byte-immutability of CRC-covered frames). */
function walkV2Frames(bytes: Uint8Array): V2FrameExtent[] {
const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength)
const extents: V2FrameExtent[] = []
let offset = HEADER_BYTES
while (offset + FRAME_PREFIX_BYTES <= bytes.length) {
const length = view.getUint32(offset, true)
const expectedCrc = view.getUint32(offset + 4, true)
const start = offset + FRAME_PREFIX_BYTES
const end = start + length
if (end > bytes.length) break // torn tail
const payload = bytes.subarray(start, end)
if (crc32c(payload) !== expectedCrc) break // torn tail
const fact = decodeFormatFact(payload, FACT_LOG_FORMAT_V2, {
expectedIdSpaceWidth: 64
}) as CommitFactV2
extents.push({ end, generation: fact.generation, isPad: fact.records.length === 0 })
offset = end
}
return extents
}
/**
* The byte offset a v2 segment is cut at to keep exactly the facts with
* `generation ≤ keepThrough`: the end of the last kept FACT frame (pads
* between kept facts sit inside the retained span; pads after the cut are
* dropped and re-sealed at the next sync). When nothing is dropped the cut
* lands after the last intact frame trailing pads retained, only a torn
* suffix (if any) removed.
*/
function v2CutOffset(extents: V2FrameExtent[], keepThrough: number): number {
let cut = HEADER_BYTES
let lastIntactEnd = HEADER_BYTES
for (const e of extents) {
lastIntactEnd = e.end
if (e.isPad) continue
if (e.generation <= keepThrough) {
cut = e.end
} else {
return cut // first beyond-keep fact: everything from here (pads included) goes
}
}
return lastIntactEnd
}
/**
* Parse a segment's bytes: verify the header, then walk frames until the end
* or a torn tail (length overrun / CRC mismatch), which terminates the walk
* everything before it is intact. Returns the decoded facts plus the byte
* length of the VALID prefix (header + intact frames), which reconciliation
* uses to cut a torn tail without re-encoding.
* everything before it is intact. The header's formatVersion selects the
* decoder: the v1 walk below is byte-identical to the original v1 reader;
* v2 segments decode through the reference codec (`decodeGroupV2`, pads
* invisible, id-space width verified at 64 a disagreeing genesis throws
* the codec's typed `GenesisWidthMismatchError`). Returns the decoded facts
* plus the byte length of the VALID prefix (header + intact frames), which
* reconciliation uses to cut a torn tail without re-encoding.
*/
function parseSegment(
file: string,
bytes: Uint8Array
): { facts: CommitFact[]; validBytes: number } {
): { facts: CommitFact[]; validBytes: number; formatVersion: number; sealSize?: number } {
if (bytes.length < HEADER_BYTES) {
prodLog.warn(`[FactLog] segment ${file} shorter than its header — treating as empty`)
return { facts: [], validBytes: 0 }
return { facts: [], validBytes: 0, formatVersion: 0 }
}
for (let i = 0; i < MAGIC.length; i++) {
if (bytes[i] !== MAGIC[i]) {
throw new Error(`fact log: segment ${file} has a bad magic — not a fact segment`)
}
let header: { formatVersion: number; sealSize?: number }
try {
header = parseSegmentHeader(bytes.subarray(0, HEADER_BYTES))
} catch (err) {
throw new Error(`fact log: segment ${file}: ${(err as Error).message}`)
}
const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength)
const version = view.getUint32(8, true)
if (version !== FACTS_FORMAT_VERSION) {
throw new Error(
`fact log: segment ${file} has formatVersion ${version}; this build reads ${FACTS_FORMAT_VERSION}`
)
}
for (let i = 20; i < HEADER_BYTES; i++) {
if (bytes[i] !== 0) {
// Non-zero reserved bytes = a future format this build cannot verify.
throw new Error(`fact log: segment ${file} has non-zero reserved header bytes — unverifiable`)
if (header.formatVersion === FACT_LOG_FORMAT_V2) {
const group = decodeGroupV2(bytes.subarray(HEADER_BYTES), { expectedIdSpaceWidth: 64 })
return {
facts: group.facts.map(commitFactFromV2),
validBytes: HEADER_BYTES + group.validBytes,
formatVersion: FACT_LOG_FORMAT_V2,
sealSize: header.sealSize
}
}
// v1 walk — byte-identical to the original reader.
const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength)
const facts: CommitFact[] = []
let offset = HEADER_BYTES
while (offset + FRAME_PREFIX_BYTES <= bytes.length) {
@ -298,7 +684,7 @@ function parseSegment(
facts.push(decodeFact(payload))
offset = end
}
return { facts, validBytes: offset }
return { facts, validBytes: offset, formatVersion: FACT_LOG_FORMAT_V1 }
}
/**
@ -318,18 +704,38 @@ export class FactLog {
}
/** Decoded facts of the TAIL segment (bounded by the rotation threshold). */
private tailFacts: CommitFact[] = []
/** Byte size of the tail segment file (valid prefix). */
/** Byte size of the tail segment file (valid prefix, pads included
* pads count toward bytes but NEVER toward facts). */
private tailBytes = 0
/** Highest generation in the log (0 = empty). */
private head = 0
/** Segment paths appended since the last sync (the fsync batch). */
private readonly dirtySegments = new Set<string>()
/** The TAIL segment's on-disk format version (selects the live encoder). */
private tailVersion: number = FACT_LOG_FORMAT_V1
/** The tail's sector-seal size (v2 tails; from its header on reopen). */
private tailSealSize: number = DEFAULT_SEAL_SIZE
/** The v2 int minter (see {@link FactIntMinter}); null = v1 live writes. */
private intMinter: FactIntMinter | null = null
constructor(storage: FactLogStorage, options?: { rotateBytes?: number }) {
this.storage = storage
this.rotateBytes = options?.rotateBytes ?? SEGMENT_ROTATE_BYTES
}
/**
* Install the v2 int minter the capability gate for v2 LIVE WRITES.
* With a minter installed, every NEW tail segment writes the v2 format and
* after-image records carry minted dense ints; without one, live writes
* stay v1 (no authority could reproduce int assignments, and 0 is never
* written). The brain wires this from the metadata index's id mapper right
* after the index is ready; an existing v1 tail cuts over on the next
* append (empty tail: re-headed in place; non-empty: sealed by rotation).
*/
setIntMinter(mint: FactIntMinter): void {
this.intMinter = mint
}
/** The highest committed generation the log holds (0 = empty). */
headGeneration(): number {
return this.head
@ -412,11 +818,17 @@ export class FactLog {
const tailPath = `${FACTS_PREFIX}/${this.manifest.tailSegment}`
const bytes = await this.storage.readRawBytes(tailPath)
if (bytes === null) {
// Manifest named a tail whose first byte never landed — an empty tail.
// Manifest named a tail whose first byte never landed — an empty
// tail. Its header (and format version) is established at the next
// append (see the tail-provisioning ladder there).
this.tailFacts = []
this.tailBytes = 0
} else {
const { facts, validBytes } = parseSegment(this.manifest.tailSegment, bytes)
const parsed = parseSegment(this.manifest.tailSegment, bytes)
const { facts, validBytes } = parsed
this.tailVersion =
parsed.formatVersion === FACT_LOG_FORMAT_V2 ? FACT_LOG_FORMAT_V2 : FACT_LOG_FORMAT_V1
this.tailSealSize = parsed.sealSize ?? DEFAULT_SEAL_SIZE
const kept = facts.filter((f) => f.generation <= committedGeneration)
if (kept.length !== facts.length || validBytes !== bytes.length) {
const dropped = facts.length - kept.length
@ -426,7 +838,16 @@ export class FactLog {
`${committedGeneration} from the tail (never committed)`
)
}
await this.rewriteTail(kept)
if (this.tailVersion === FACT_LOG_FORMAT_V2) {
// V2: byte-slice at frame boundaries — CRC-covered frames are
// byte-immutable; a truncation never re-encodes what it keeps.
const cut = v2CutOffset(walkV2Frames(bytes), committedGeneration)
await this.storage.writeRawBytes(tailPath, bytes.subarray(0, cut))
this.tailFacts = kept
this.tailBytes = cut
} else {
await this.rewriteTail(kept)
}
} else {
this.tailFacts = facts
this.tailBytes = validBytes
@ -441,6 +862,15 @@ export class FactLog {
* Append one committed generation's fact. NOT durable until {@link sync}
* the caller batches durability at its commit barrier (transact syncs in
* the same call; Model-B group-commit syncs at flush).
*
* Tail provisioning (in order): a missing tail starts one; a named tail
* whose header never landed (manifest-first crash) gets its header now; an
* existing V1 tail cuts over to v2 once the minter is installed (empty:
* re-headed in place, non-empty: sealed by rotation v1 segments are never
* rewritten); a full tail rotates. The frame then encodes in the TAIL's
* format: v2 tails carry after-image records with minted ints (and the
* genesis record on the very first fact of a brand-new log); v1 tails keep
* the v1 wire format byte-identically.
*/
async append(fact: CommitFact): Promise<void> {
if (fact.generation <= this.head) {
@ -450,10 +880,42 @@ export class FactLog {
}
if (this.manifest.tailSegment === null) {
await this.startTail(fact.generation)
} else if (this.tailBytes === 0) {
await this.reinitializeTailHeader()
} else if (this.intMinter !== null && this.tailVersion === FACT_LOG_FORMAT_V1) {
if (this.tailFacts.length === 0 && this.tailBytes <= HEADER_BYTES) {
await this.upgradeEmptyTailToV2()
} else {
await this.rotate(fact.generation)
}
} else if (this.tailBytes >= this.rotateBytes) {
await this.rotate(fact.generation)
}
const frame = encodeFrame(fact)
let frame: Uint8Array
if (this.tailVersion === FACT_LOG_FORMAT_V2) {
const records = this.buildV2Records(fact)
if (this.needsGenesis()) {
if (this.ensureBrainId()) await this.persistManifest()
records.unshift(this.genesisRecord())
}
frame = encodeFactV2({
generation: fact.generation,
timestamp: fact.timestamp,
records,
...(fact.meta ? { meta: toJsonSafe(fact.meta) as Record<string, unknown> } : {}),
...(fact.blobHashes && fact.blobHashes.length > 0 ? { blobHashes: fact.blobHashes } : {})
})
} else {
if (fact.records && fact.records.length > 0) {
throw new Error(
`fact log: marker records (${fact.records.map((r) => r.type).join(', ')}) require a ` +
`v2 tail segment — this log's tail is v1 (no int minter installed); refusing rather ` +
`than silently dropping them`
)
}
frame = encodeFrame(fact)
}
const tailPath = `${FACTS_PREFIX}/${this.manifest.tailSegment}`
await this.storage.appendRawBytes(tailPath, frame)
this.tailFacts.push(fact)
@ -462,8 +924,16 @@ export class FactLog {
this.dirtySegments.add(tailPath)
}
/** Fsync every segment appended since the last sync. */
/**
* Fsync every segment appended since the last sync. SEALS AT SYNC: a v2
* tail is first padded to its sector-seal boundary (one pad frame,
* invisible to readers; a gap smaller than the smallest constructible pad
* frame pads through one extra sector the codec's rule), so every
* durability barrier leaves the tail sector-aligned: a torn page can only
* tear INSIDE the group being written, never a previously-sealed one.
*/
async sync(): Promise<void> {
await this.padTailToSealBoundary()
if (this.dirtySegments.size === 0) return
const paths = [...this.dirtySegments]
this.dirtySegments.clear()
@ -671,7 +1141,25 @@ export class FactLog {
`(head ${this.head}) — the fact to drop was already sealed; the log needs reopen`
)
}
await this.rewriteTail(kept)
if (this.tailVersion === FACT_LOG_FORMAT_V2) {
// V2: byte-slice at frame boundaries (kept frames stay byte-identical;
// pads between kept facts are retained inside the prefix, trailing pads
// go and the next sync re-seals). The dropped frames may be unsynced —
// readRawBytes is read-after-write coherent over the append path.
const file = this.manifest.tailSegment
if (!file) return
const tailPath = `${FACTS_PREFIX}/${file}`
const bytes = await this.storage.readRawBytes(tailPath)
if (bytes === null) {
throw new Error(`fact log: dropAbove(${keepThrough}) cannot read the tail segment ${file}`)
}
const cut = v2CutOffset(walkV2Frames(bytes), keepThrough)
await this.storage.writeRawBytes(tailPath, bytes.subarray(0, cut))
this.tailFacts = kept
this.tailBytes = cut
} else {
await this.rewriteTail(kept)
}
this.head = this.computeHead()
}
@ -684,25 +1172,42 @@ export class FactLog {
return 0
}
/** The header bytes for a NEW tail: v2 whenever the minter is installed. */
private newTailHeader(firstGeneration: number): Uint8Array {
return this.intMinter !== null
? encodeSegmentHeaderV2(firstGeneration, DEFAULT_SEAL_SIZE)
: buildHeader(firstGeneration)
}
/** Record the just-created tail's format in memory (mirrors its header). */
private noteFreshTail(): void {
this.tailVersion = this.intMinter !== null ? FACT_LOG_FORMAT_V2 : FACT_LOG_FORMAT_V1
this.tailSealSize = DEFAULT_SEAL_SIZE
}
/** Create the very first tail segment (manifest-first, then header bytes). */
private async startTail(firstGeneration: number): Promise<void> {
const file = segmentFileName(firstGeneration)
this.manifest.tailSegment = file
if (this.intMinter !== null) this.ensureBrainId()
await this.persistManifest()
await this.storage.appendRawBytes(`${FACTS_PREFIX}/${file}`, buildHeader(firstGeneration))
await this.storage.appendRawBytes(`${FACTS_PREFIX}/${file}`, this.newTailHeader(firstGeneration))
this.tailFacts = []
this.tailBytes = HEADER_BYTES
this.noteFreshTail()
}
/**
* Seal the tail into the manifest and start a new one. Manifest-first: the
* flip both seals the old tail AND names the new one atomically, so no
* segment file ever exists unaccounted for.
* segment file ever exists unaccounted for. The NEW tail's format follows
* the minter gate ({@link newTailHeader}) this is also the v1v2 cutover
* seam for a non-empty v1 tail (sealed as-is, never rewritten).
*/
private async rotate(nextGeneration: number): Promise<void> {
const sealedFile = this.manifest.tailSegment
if (!sealedFile) return
// Seal what the tail actually holds.
// Seal what the tail actually holds (sync() also sector-seals a v2 tail).
await this.sync() // sealed segments are always fully durable
const entry: SegmentEntry = {
file: sealedFile,
@ -714,10 +1219,181 @@ export class FactLog {
const newFile = segmentFileName(nextGeneration)
this.manifest.segments.push(entry)
this.manifest.tailSegment = newFile
if (this.intMinter !== null) this.ensureBrainId()
await this.persistManifest()
await this.storage.appendRawBytes(`${FACTS_PREFIX}/${newFile}`, buildHeader(nextGeneration))
await this.storage.appendRawBytes(`${FACTS_PREFIX}/${newFile}`, this.newTailHeader(nextGeneration))
this.tailFacts = []
this.tailBytes = HEADER_BYTES
this.noteFreshTail()
}
/**
* The v1v2 cutover for an EMPTY v1 tail: re-head it in place (nothing but
* the 32-byte header exists, so no v1 frame is ever rewritten). Also the
* cheapest cutover shape: brand-new brains whose first tail predates the
* minter installation converge here on their first post-install append.
*/
private async upgradeEmptyTailToV2(): Promise<void> {
const file = this.manifest.tailSegment
if (!file) return
if (this.ensureBrainId()) await this.persistManifest()
const first = this.segmentFirstGenerationFromName(file)
const path = `${FACTS_PREFIX}/${file}`
await this.storage.writeRawBytes(path, encodeSegmentHeaderV2(first, DEFAULT_SEAL_SIZE))
this.tailBytes = HEADER_BYTES
this.tailVersion = FACT_LOG_FORMAT_V2
this.tailSealSize = DEFAULT_SEAL_SIZE
this.dirtySegments.add(path)
}
/**
* A manifest-named tail whose header never landed (crash between the
* manifest flip and the first header byte previously this appended
* frames into a headerless file the next open could not parse): write the
* header now, in the CURRENT format gate.
*/
private async reinitializeTailHeader(): Promise<void> {
const file = this.manifest.tailSegment
if (!file) return
if (this.intMinter !== null && this.ensureBrainId()) await this.persistManifest()
const first = this.segmentFirstGenerationFromName(file)
const path = `${FACTS_PREFIX}/${file}`
await this.storage.writeRawBytes(path, this.newTailHeader(first))
this.tailBytes = HEADER_BYTES
this.noteFreshTail()
this.dirtySegments.add(path)
}
/** True when the NEXT appended fact is the first fact of a brand-new v2
* log the one that must open with the log.genesis record. */
private needsGenesis(): boolean {
return (
this.tailVersion === FACT_LOG_FORMAT_V2 &&
this.manifest.segments.length === 0 &&
this.tailFacts.length === 0
)
}
/** Mint the brain id into the manifest if absent; true when it changed. */
private ensureBrainId(): boolean {
if (this.manifest.brainId) return false
this.manifest.brainId = uuidv4()
return true
}
/** The log's birth certificate (id-space width 64 the only width this
* writer mints; a reader expecting another width refuses at decode). */
private genesisRecord(): LogRecord {
const brainId = this.manifest.brainId
if (!brainId) {
throw new Error(
'fact log v2: genesis requires a brainId in the facts manifest — invariant violated'
)
}
return { type: 'log.genesis', idSpaceWidth: 64, brainId, createdAt: Date.now() }
}
/**
* Convert one CommitFact's ops (+ optional marker records) to v2 wire
* records, MINTING ints at append time: entity/verb ints come from the
* injected minter (the metadata index's id mapper the one authority a
* rebuild reproduces exactly). Verb endpoints and the verb name ride as
* first-class wire fields, lifted from the canonical verb vector wrapper.
* Every refusal here is loud an after-image without a mintable int, a
* verb without endpoints, or a vector record without floats fails the
* WRITE, never writes a 0.
*/
private buildV2Records(fact: CommitFact): LogRecord[] {
const mint = (kind: 'noun' | 'verb', id: string): bigint => {
if (this.intMinter === null) {
throw new Error(
`fact log v2: no int minter is installed — cannot mint the ${kind} int for ${id}; ` +
`refusing to write a v2 after-image (an int of 0 is never written)`
)
}
const minted = this.intMinter(kind, id)
if (typeof minted !== 'bigint' || minted <= 0n) {
throw new Error(
`fact log v2: the int minter returned ${String(minted)} for ${kind} ${id}` +
`minted ints are positive bigints; refusing to write`
)
}
return minted
}
const records: LogRecord[] = []
for (const op of fact.ops) {
if (op.kind === 'noun') {
if (op.record === null) {
records.push({ type: 'noun.tombstone', id: op.id })
continue
}
records.push({
type: 'noun.afterImage',
id: op.id,
entityInt: mint('noun', op.id),
metadata: toJsonSafe(op.record.metadata ?? null),
vectorLeg: embeddingLegOf(op.record.vector, `noun ${op.id}`)
})
} else {
if (op.record === null) {
records.push({ type: 'verb.tombstone', id: op.id })
continue
}
const wrapper = op.record.vector as Record<string, unknown> | null
const verbName = wrapper?.verb
const sourceId = wrapper?.sourceId
const targetId = wrapper?.targetId
if (
typeof verbName !== 'string' ||
typeof sourceId !== 'string' ||
typeof targetId !== 'string'
) {
throw new Error(
`fact log v2: verb ${op.id} has no canonical endpoints (verb/sourceId/targetId ` +
`live in its vector record, which is missing or torn) — refusing to write an ` +
`after-image that could not be replayed`
)
}
const floats = floatsOf(wrapper?.vector, `verb ${op.id}`) ?? []
records.push({
type: 'verb.afterImage',
id: op.id,
verbInt: mint('verb', op.id),
metadata: toJsonSafe(op.record.metadata ?? null),
vectorLeg: floats,
verb: verbName,
sourceId,
sourceInt: mint('noun', sourceId),
targetId,
targetInt: mint('noun', targetId)
})
}
}
for (const marker of fact.records ?? []) records.push(marker)
return records
}
/**
* Pad a v2 tail to its next sector-seal boundary with ONE pad frame
* called from {@link sync} so alignment holds at every durability barrier.
* Pads count toward {@link tailBytes} but never toward facts (they are
* invisible to every reader); a gap smaller than the smallest constructible
* pad frame pads through one extra sector (the codec's rule). No-op for v1
* tails, empty tails, and already-aligned tails.
*/
private async padTailToSealBoundary(): Promise<void> {
if (this.tailVersion !== FACT_LOG_FORMAT_V2) return
const file = this.manifest.tailSegment
if (!file || this.tailBytes <= HEADER_BYTES) return
const remainder = this.tailBytes % this.tailSealSize
if (remainder === 0) return
let padBytes = this.tailSealSize - remainder
if (padBytes < minPadFrameBytes()) padBytes += this.tailSealSize
const tailPath = `${FACTS_PREFIX}/${file}`
await this.storage.appendRawBytes(tailPath, encodePadFrame(padBytes))
this.tailBytes += padBytes
this.dirtySegments.add(tailPath)
}
/** Atomically persist the manifest (write-new → fsync → rename downstream). */
@ -746,17 +1422,24 @@ export class FactLog {
this.tailBytes = total
}
/** Cut a SEALED segment back to `committedGeneration` (atomic replace). */
/** Cut a SEALED segment back to `committedGeneration` (atomic replace).
* v2 segments byte-slice at frame boundaries (kept frames pads
* included are never re-encoded); the v1 re-encode path is unchanged. */
private async truncateSegmentTo(file: string, committedGeneration: number): Promise<void> {
const path = `${FACTS_PREFIX}/${file}`
const bytes = await this.storage.readRawBytes(path)
if (bytes === null) return
const { facts } = parseSegment(file, bytes)
const { facts, formatVersion } = parseSegment(file, bytes)
const kept = facts.filter((f) => f.generation <= committedGeneration)
prodLog.warn(
`[FactLog] truncating sealed segment ${file} to generation ${committedGeneration} ` +
`(${facts.length - kept.length} uncommitted fact(s) dropped)`
)
if (formatVersion === FACT_LOG_FORMAT_V2) {
const cut = v2CutOffset(walkV2Frames(bytes), committedGeneration)
await this.storage.writeRawBytes(path, bytes.subarray(0, cut))
return
}
const first = kept[0]?.generation ?? this.segmentFirstGenerationFromName(file)
const parts: Uint8Array[] = [buildHeader(first)]
for (const f of kept) parts.push(encodeFrame(f))

View file

@ -26,9 +26,21 @@
* position 2 is `records`, not v1's `ops`)
*
* fact := [ generation:u64, timestamp:u64, records, meta|nil, blobHashes|nil ]
* record := [ recordType:u8, recordVersion:u8, ...type-specific fields ]
* record := [ recordType:u8, recordVersion:u8, cipherFlag:u8, keyId:bin16|nil,
* ...type-specific fields ]
*
* Record type registry (all recordVersion = 1):
* `cipherFlag`/`keyId` are RESERVED crypto envelope fields: `0`/`nil` (a
* plaintext record) is the ONLY legal combination this release writes or
* reads. Any nonzero cipherFlag or non-nil keyId refuses with the typed
* {@link UnknownLogRecordError} ("encrypted records need a newer reader")
* so record-level encryption can land later without a format-version bump on
* the one compat surface. No crypto logic exists here; the bytes are reserved
* only. Pad records (type 0) are exempt: they are skipped WHOLESALE as
* length-only filler, so their fields beyond [type, version] are never
* inspected (this keeps pad frames byte-stable across the envelope change).
*
* Record type registry (all recordVersion = 1; type-specific fields listed
* every record carries the 4-field envelope above first):
*
* 0 pad [] length-only filler; readers SKIP; crc-covered
* 1 noun.afterImage [id bin16, entityInt u64, metadata, vectorLeg]
@ -109,6 +121,13 @@ export const DEFAULT_SEAL_SIZE = 4096
/** The record version this reader knows (all registry types are version 1). */
export const LOG_RECORD_VERSION = 1
/**
* The only legal `cipherFlag` value this release: plaintext. The encoder
* always writes it (with a nil keyId); the decoder refuses anything else
* with {@link UnknownLogRecordError} encrypted records need a newer reader.
*/
export const LOG_RECORD_CIPHER_PLAINTEXT = 0
/** The v2 record-type registry — wire codes for every record type. */
export const LOG_RECORD_TYPES = {
PAD: 0,
@ -648,22 +667,31 @@ function decodeVectorLeg(wire: unknown, context: string): VectorLeg {
// Record encode/decode
// ---------------------------------------------------------------------------
/** Encode one record into its positional wire array. */
/**
* Encode one record into its positional wire array. Every record leads with
* the 4-field envelope [type, version, cipherFlag, keyId]; this release
* writes cipherFlag {@link LOG_RECORD_CIPHER_PLAINTEXT} and a nil keyId
* always (the fields are crypto-RESERVED, carrying no logic yet).
*/
function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefined): unknown[] {
const T = LOG_RECORD_TYPES
const V = LOG_RECORD_VERSION
const C = LOG_RECORD_CIPHER_PLAINTEXT
const K = null // keyId: nil until record-level encryption exists
switch (record.type) {
case 'noun.afterImage':
return [
T.NOUN_AFTER_IMAGE,
V,
C,
K,
uuidToBytes(record.id),
toWireU64(record.entityInt, 'entityInt'),
record.metadata ?? null,
encodeVectorLeg(record.vectorLeg, options, `noun.afterImage ${record.id}`)
]
case 'noun.tombstone':
return [T.NOUN_TOMBSTONE, V, uuidToBytes(record.id)]
return [T.NOUN_TOMBSTONE, V, C, K, uuidToBytes(record.id)]
case 'verb.afterImage': {
if (typeof record.verb !== 'string' || record.verb.length === 0) {
throw new Error(`fact log v2: verb.afterImage ${record.id} needs a non-empty verb name`)
@ -671,6 +699,8 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
return [
T.VERB_AFTER_IMAGE,
V,
C,
K,
uuidToBytes(record.id),
toWireU64(record.verbInt, 'verbInt'),
record.metadata ?? null,
@ -683,16 +713,18 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
]
}
case 'verb.tombstone':
return [T.VERB_TOMBSTONE, V, uuidToBytes(record.id)]
return [T.VERB_TOMBSTONE, V, C, K, uuidToBytes(record.id)]
case 'batch.meta':
if (!isPlainMap(record.meta)) {
throw new Error('fact log v2: batch.meta requires a map')
}
return [T.BATCH_META, V, record.meta]
return [T.BATCH_META, V, C, K, record.meta]
case 'embed.pending':
return [
T.EMBED_PENDING,
V,
C,
K,
uuidToBytes(record.id),
toWireU64(record.enqueuedAt, 'enqueuedAt')
]
@ -703,7 +735,7 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
`refs and nil are not allowed here`
)
}
return [T.EMBED_LANDED, V, uuidToBytes(record.id), record.vector]
return [T.EMBED_LANDED, V, C, K, uuidToBytes(record.id), record.vector]
}
case 'blob.manifest': {
if (typeof record.mimeType !== 'string') {
@ -715,6 +747,8 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
return [
T.BLOB_MANIFEST,
V,
C,
K,
hashToBytes(record.hash),
toWireU64(record.size, 'blob size'),
record.mimeType,
@ -725,7 +759,7 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
if (!isPlainMap(record.note)) {
throw new Error('fact log v2: projection.note requires a map')
}
return [T.PROJECTION_NOTE, V, record.note]
return [T.PROJECTION_NOTE, V, C, K, record.note]
case 'bootstrap.baseline': {
if (record.kind !== 'noun' && record.kind !== 'verb') {
throw new Error(`fact log v2: bootstrap.baseline kind must be 'noun' or 'verb'`)
@ -733,6 +767,8 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
return [
T.BOOTSTRAP_BASELINE,
V,
C,
K,
uuidToBytes(record.id),
record.kind === 'noun' ? 0 : 1,
record.metadata ?? null,
@ -748,6 +784,8 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
return [
T.LOG_GENESIS,
V,
C,
K,
record.idSpaceWidth,
uuidToBytes(record.brainId),
toWireU64(record.createdAt, 'createdAt')
@ -763,35 +801,39 @@ function encodeRecord(record: LogRecord, options: EncodeFactV2Options | undefine
}
}
/** Exact wire arity per record type (envelope of 2 + type-specific fields). */
/** Exact wire arity per record type (envelope of 4 + type-specific fields). */
const RECORD_ARITY: Record<number, number> = {
[LOG_RECORD_TYPES.NOUN_AFTER_IMAGE]: 6,
[LOG_RECORD_TYPES.NOUN_TOMBSTONE]: 3,
[LOG_RECORD_TYPES.VERB_AFTER_IMAGE]: 11,
[LOG_RECORD_TYPES.VERB_TOMBSTONE]: 3,
[LOG_RECORD_TYPES.BATCH_META]: 3,
[LOG_RECORD_TYPES.EMBED_PENDING]: 4,
[LOG_RECORD_TYPES.EMBED_LANDED]: 4,
[LOG_RECORD_TYPES.BLOB_MANIFEST]: 6,
[LOG_RECORD_TYPES.PROJECTION_NOTE]: 3,
[LOG_RECORD_TYPES.BOOTSTRAP_BASELINE]: 6,
[LOG_RECORD_TYPES.LOG_GENESIS]: 5
[LOG_RECORD_TYPES.NOUN_AFTER_IMAGE]: 8,
[LOG_RECORD_TYPES.NOUN_TOMBSTONE]: 5,
[LOG_RECORD_TYPES.VERB_AFTER_IMAGE]: 13,
[LOG_RECORD_TYPES.VERB_TOMBSTONE]: 5,
[LOG_RECORD_TYPES.BATCH_META]: 5,
[LOG_RECORD_TYPES.EMBED_PENDING]: 6,
[LOG_RECORD_TYPES.EMBED_LANDED]: 6,
[LOG_RECORD_TYPES.BLOB_MANIFEST]: 8,
[LOG_RECORD_TYPES.PROJECTION_NOTE]: 5,
[LOG_RECORD_TYPES.BOOTSTRAP_BASELINE]: 8,
[LOG_RECORD_TYPES.LOG_GENESIS]: 7
}
/**
* Decode one wire record. Returns `null` for pads (skipped by definition).
* Unknown type / newer version throw {@link UnknownLogRecordError} never
* skip-and-continue.
* skip-and-continue. The reserved crypto envelope is verified BEFORE the
* arity check (an encrypted record's field layout is a newer reader's
* business, not a malformed-record error): any nonzero cipherFlag or non-nil
* keyId refuses with the same typed error class.
*/
function decodeRecord(raw: unknown): LogRecord | null {
if (!Array.isArray(raw) || raw.length < 2) {
throw new Error('fact log v2: malformed record envelope (need [type, version, ...])')
throw new Error('fact log v2: malformed record envelope (need [type, version, cipherFlag, keyId, ...])')
}
const recordType = wireToU8(raw[0], 'recordType')
const recordVersion = wireToU8(raw[1], 'recordVersion')
if (recordType === LOG_RECORD_TYPES.PAD) {
// Length-only filler: skipped wholesale, filler fields never inspected.
// Length-only filler: skipped wholesale, filler fields never inspected
// (pads therefore carry no crypto envelope — by definition, not omission).
return null
}
const arity = RECORD_ARITY[recordType]
@ -814,6 +856,20 @@ function decodeRecord(raw: unknown): LogRecord | null {
if (recordVersion !== LOG_RECORD_VERSION) {
throw new Error(`fact log v2: record type ${recordType} has invalid record version ${recordVersion}`)
}
if (raw.length < 4) {
throw new Error('fact log v2: malformed record envelope (need [type, version, cipherFlag, keyId, ...])')
}
const cipherFlag = wireToU8(raw[2], 'cipherFlag')
const keyId = raw[3]
if (cipherFlag !== LOG_RECORD_CIPHER_PLAINTEXT || (keyId !== null && keyId !== undefined)) {
throw new UnknownLogRecordError(
recordType,
recordVersion,
`fact log v2: record type ${recordType} carries cipherFlag ${cipherFlag}` +
`${keyId !== null && keyId !== undefined ? ' and a keyId' : ''}` +
`encrypted records need a newer reader`
)
}
if (raw.length !== arity) {
throw new Error(
`fact log v2: record type ${recordType} expects ${arity} wire fields; got ${raw.length}`
@ -824,94 +880,94 @@ function decodeRecord(raw: unknown): LogRecord | null {
case LOG_RECORD_TYPES.NOUN_AFTER_IMAGE:
return {
type: 'noun.afterImage',
id: bytesToUuid(raw[2], 'noun.afterImage id'),
entityInt: wireToBigint(raw[3], 'entityInt'),
metadata: raw[4] ?? null,
vectorLeg: decodeVectorLeg(raw[5], 'noun.afterImage')
id: bytesToUuid(raw[4], 'noun.afterImage id'),
entityInt: wireToBigint(raw[5], 'entityInt'),
metadata: raw[6] ?? null,
vectorLeg: decodeVectorLeg(raw[7], 'noun.afterImage')
}
case LOG_RECORD_TYPES.NOUN_TOMBSTONE:
return { type: 'noun.tombstone', id: bytesToUuid(raw[2], 'noun.tombstone id') }
return { type: 'noun.tombstone', id: bytesToUuid(raw[4], 'noun.tombstone id') }
case LOG_RECORD_TYPES.VERB_AFTER_IMAGE: {
if (typeof raw[6] !== 'string') {
if (typeof raw[8] !== 'string') {
throw new Error('fact log v2: verb.afterImage verb name is not a string')
}
return {
type: 'verb.afterImage',
id: bytesToUuid(raw[2], 'verb.afterImage id'),
verbInt: wireToBigint(raw[3], 'verbInt'),
metadata: raw[4] ?? null,
vectorLeg: decodeVectorLeg(raw[5], 'verb.afterImage'),
verb: raw[6],
sourceId: bytesToUuid(raw[7], 'verb.afterImage sourceId'),
sourceInt: wireToBigint(raw[8], 'sourceInt'),
targetId: bytesToUuid(raw[9], 'verb.afterImage targetId'),
targetInt: wireToBigint(raw[10], 'targetInt')
id: bytesToUuid(raw[4], 'verb.afterImage id'),
verbInt: wireToBigint(raw[5], 'verbInt'),
metadata: raw[6] ?? null,
vectorLeg: decodeVectorLeg(raw[7], 'verb.afterImage'),
verb: raw[8],
sourceId: bytesToUuid(raw[9], 'verb.afterImage sourceId'),
sourceInt: wireToBigint(raw[10], 'sourceInt'),
targetId: bytesToUuid(raw[11], 'verb.afterImage targetId'),
targetInt: wireToBigint(raw[12], 'targetInt')
}
}
case LOG_RECORD_TYPES.VERB_TOMBSTONE:
return { type: 'verb.tombstone', id: bytesToUuid(raw[2], 'verb.tombstone id') }
return { type: 'verb.tombstone', id: bytesToUuid(raw[4], 'verb.tombstone id') }
case LOG_RECORD_TYPES.BATCH_META: {
if (!isPlainMap(raw[2])) throw new Error('fact log v2: batch.meta payload is not a map')
return { type: 'batch.meta', meta: raw[2] }
if (!isPlainMap(raw[4])) throw new Error('fact log v2: batch.meta payload is not a map')
return { type: 'batch.meta', meta: raw[4] }
}
case LOG_RECORD_TYPES.EMBED_PENDING:
return {
type: 'embed.pending',
id: bytesToUuid(raw[2], 'embed.pending id'),
enqueuedAt: wireToNumber(raw[3], 'enqueuedAt')
id: bytesToUuid(raw[4], 'embed.pending id'),
enqueuedAt: wireToNumber(raw[5], 'enqueuedAt')
}
case LOG_RECORD_TYPES.EMBED_LANDED: {
const leg = decodeVectorLeg(raw[3], 'embed.landed')
const leg = decodeVectorLeg(raw[5], 'embed.landed')
if (!Array.isArray(leg)) {
throw new Error(
'fact log v2: embed.landed must carry an INLINE float vector — refs and nil are not allowed here'
)
}
return { type: 'embed.landed', id: bytesToUuid(raw[2], 'embed.landed id'), vector: leg }
return { type: 'embed.landed', id: bytesToUuid(raw[4], 'embed.landed id'), vector: leg }
}
case LOG_RECORD_TYPES.BLOB_MANIFEST: {
if (typeof raw[4] !== 'string') {
if (typeof raw[6] !== 'string') {
throw new Error('fact log v2: blob.manifest mimeType is not a string')
}
const refOp = wireToU8(raw[5], 'refOp')
const refOp = wireToU8(raw[7], 'refOp')
if (refOp !== 0 && refOp !== 1) {
throw new Error(`fact log v2: blob.manifest refOp must be 0 (add) or 1 (release); got ${refOp}`)
}
return {
type: 'blob.manifest',
hash: bytesToHash(raw[2]),
size: wireToNumber(raw[3], 'blob size'),
mimeType: raw[4],
hash: bytesToHash(raw[4]),
size: wireToNumber(raw[5], 'blob size'),
mimeType: raw[6],
refOp: refOp === 0 ? 'add' : 'release'
}
}
case LOG_RECORD_TYPES.PROJECTION_NOTE: {
if (!isPlainMap(raw[2])) throw new Error('fact log v2: projection.note payload is not a map')
return { type: 'projection.note', note: raw[2] }
if (!isPlainMap(raw[4])) throw new Error('fact log v2: projection.note payload is not a map')
return { type: 'projection.note', note: raw[4] }
}
case LOG_RECORD_TYPES.BOOTSTRAP_BASELINE: {
const kind = wireToU8(raw[3], 'bootstrap.baseline kind')
const kind = wireToU8(raw[5], 'bootstrap.baseline kind')
if (kind !== 0 && kind !== 1) {
throw new Error(`fact log v2: bootstrap.baseline kind must be 0 (noun) or 1 (verb); got ${kind}`)
}
return {
type: 'bootstrap.baseline',
id: bytesToUuid(raw[2], 'bootstrap.baseline id'),
id: bytesToUuid(raw[4], 'bootstrap.baseline id'),
kind: kind === 0 ? 'noun' : 'verb',
metadata: raw[4] ?? null,
vectorLeg: decodeVectorLeg(raw[5], 'bootstrap.baseline')
metadata: raw[6] ?? null,
vectorLeg: decodeVectorLeg(raw[7], 'bootstrap.baseline')
}
}
case LOG_RECORD_TYPES.LOG_GENESIS: {
const width = wireToU8(raw[2], 'idSpaceWidth')
const width = wireToU8(raw[4], 'idSpaceWidth')
if (width !== 32 && width !== 64) {
throw new Error(`fact log v2: log.genesis idSpaceWidth must be 32 or 64; got ${width}`)
}
return {
type: 'log.genesis',
idSpaceWidth: width,
brainId: bytesToUuid(raw[3], 'log.genesis brainId'),
createdAt: wireToNumber(raw[4], 'createdAt')
brainId: bytesToUuid(raw[5], 'log.genesis brainId'),
createdAt: wireToNumber(raw[6], 'createdAt')
}
}
default:
@ -944,8 +1000,12 @@ export function encodeFactV2(fact: CommitFactV2, options?: EncodeFactV2Options):
if (!Number.isSafeInteger(fact.timestamp) || fact.timestamp < 0) {
throw new Error(`fact log v2: timestamp must be a non-negative integer; got ${fact.timestamp}`)
}
if (!Array.isArray(fact.records) || fact.records.length === 0) {
throw new Error('fact log v2: a fact must carry at least one record')
// records MAY be empty: a committed generation whose ops all collapsed
// (e.g. a batch whose relates deduped to no-ops) is still a real
// generation — v1 encoded empty ops the same way; refusing here would
// fork the two formats' commit semantics.
if (!Array.isArray(fact.records)) {
throw new Error('fact log v2: records must be an array')
}
if (fact.meta !== undefined && !isPlainMap(fact.meta)) {
throw new Error('fact log v2: fact meta must be a map when present')
@ -1092,9 +1152,15 @@ function decodeFactV2(payload: Uint8Array, options?: DecodeFactV2Options): Commi
// Sector seals
// ---------------------------------------------------------------------------
/** Smallest constructible pad frame (envelope + bare pad record), memoized. */
/**
* Smallest constructible pad frame in bytes (frame prefix + the bare pad
* record fact), memoized. Exported for streaming writers that pad an
* append-only tail to a seal boundary: a gap smaller than this cannot hold
* any frame, so the writer pads through one extra sector (the same rule
* {@link sealGroup} applies).
*/
let minPadFrameBytesMemo: number | null = null
function minPadFrameBytes(): number {
export function minPadFrameBytes(): number {
if (minPadFrameBytesMemo === null) {
minPadFrameBytesMemo =
FRAME_PREFIX_BYTES +
@ -1145,6 +1211,25 @@ function buildPadFrame(totalBytes: number): Uint8Array {
return buildFrame(payload)
}
/**
* Build a pad frame of EXACTLY `totalBytes` the streaming-append counterpart
* of {@link sealGroup} for writers that append pads directly to a live tail
* instead of sealing an in-memory group. Refuses sizes smaller than the
* smallest constructible pad frame ({@link minPadFrameBytes}); readers skip
* the result by definition (a type-0 record is length-only filler).
*
* @param totalBytes - The exact frame size to construct (prefix included).
* @returns The complete pad frame bytes.
*/
export function encodePadFrame(totalBytes: number): Uint8Array {
if (!Number.isInteger(totalBytes) || totalBytes < minPadFrameBytes()) {
throw new Error(
`fact log v2: a pad frame must be at least ${minPadFrameBytes()} bytes; got ${totalBytes}`
)
}
return buildPadFrame(totalBytes)
}
/**
* Seal a group of frames to a sector boundary: concatenate the frames and pad
* to the next `sealSize` multiple with ONE pad frame. An already-aligned

View file

@ -46,7 +46,13 @@ import type {
TxLogEntry
} from './types.js'
import { readLogAuthority } from './logAuthority.js'
import { FactLog, storageSupportsFactLog, type CommitFact, type FactOp } from './factLog.js'
import {
FactLog,
storageSupportsFactLog,
type CommitFact,
type FactOp,
type FactIntMinter
} from './factLog.js'
import { GenerationSegmentStore, type FoldGeneration } from './generationSegments.js'
import { crc32c } from '../utils/crc32c.js'
@ -182,6 +188,22 @@ export class GenerationStore {
this.logDurability = mode
}
/**
* The fact log's v2 int minter injected by the OWNER (brainy wires the
* metadata index's id mapper here right after the index is ready), because
* this store cannot know the mapper. With the minter installed, new fact
* segments write the v2 format and after-image records carry minted dense
* ints reproducible by an id-mapper rebuild. Survives reopen: `open()`
* re-installs it on the fresh {@link FactLog} instance.
*/
private intMinter: FactIntMinter | null = null
/** Install the fact log's v2 int minter (see {@link intMinter}). */
setIntMinter(mint: FactIntMinter): void {
this.intMinter = mint
this.factLog?.setIntMinter(mint)
}
/** Latest reserved/observed generation (≥ {@link committed}). */
private counter = 0
/** Committed-transaction watermark (manifest generation). */
@ -493,6 +515,7 @@ export class GenerationStore {
// hosts no fact log (readers fall back to canonical enumeration).
if (storageSupportsFactLog(this.storage)) {
this.factLog = new FactLog(this.storage)
if (this.intMinter) this.factLog.setIntMinter(this.intMinter)
// LOG-AUTHORITY REPLAY (durable-at-ack's recovery half): when this
// brain's stored authority is the log, an intact fact ABOVE the
// manifest is an ACKED write whose canonical bytes may not have