brainy/src/db/generationStore.ts

/**
 * @module db/generationStore
 * @description The generational record layer behind Brainy 8.0's MVCC Db API.
 *
 * One `GenerationStore` is attached to every writable `Brainy` instance. It
 * owns:
 *
 * - the **monotonic generation counter** (`_system/generation.json`),
 *   reserved per `transact()` commit and bumped per single-operation write
 *   batch via the storage-layer hook;
 * - the **commit protocol** for `brain.transact()`: capture before-images →
 *   write the generation delta → fsync → execute the operation batch →
 *   atomically rename `_system/manifest.json` (the rename IS the commit
 *   point) → append the tx-log line;
 * - **crash recovery**: on open, any `_generations/<N>` directory with
 *   `N > manifest.generation` is an uncommitted transaction — its
 *   before-images are restored to the canonical entity paths and the
 *   directory is removed, so a crash anywhere before the manifest rename
 *   leaves the store byte-identical to its pre-transaction state;
 * - **pin/release refcounting** for live `Db` values, which is what makes
 *   `compactHistory()` safe: a generation record-set is reclaimed only when
 *   no live pin could ever need it;
 * - **point-in-time resolution**: the state of id X at pinned generation G
 *   is the before-image stored by the *first* committed generation after G
 *   that touched X — or the live storage state when nothing after G touched
 *   it (so current-generation reads stay on the existing fast paths,
 *   untouched).
 *
 * Durability and atomicity guarantees, failure modes, and the intellectual
 * lineage (Datomic database-as-value, LMDB reader pins, LSM immutable
 * segments) are specified in `docs/ADR-001-generational-mvcc.md`.
 */

import { prodLog } from '../utils/logger.js'
import { GenerationCompactedError, GenerationConflictError, PendingFlushDurabilityError, StoreInconsistentError } from './errors.js'
import type { UnreconciledRecord } from './errors.js'
import { TransactionRollbackError } from '../transaction/errors.js'
import type {
  ChangedIds,
  CompactHistoryOptions,
  CompactHistoryResult,
  GenerationDelta,
  GenerationManifest,
  GenerationRecord,
  GenerationStorage,
  TxLogEntry
} from './types.js'
import { FactLog, storageSupportsFactLog, type CommitFact, type FactOp } from './factLog.js'

/**
 * The byte-identical before-images of every id a commit touches, read UNDER
 * the commit mutex immediately before the write applies. Passed to a commit's
 * optional `precommit` hook so callers can enforce compare-and-swap
 * preconditions (e.g. the per-entity `ifRev` check) atomically with the apply
 * — the same conditional-commit idea as `ifAtGeneration`, generalized to
 * arbitrary per-record predicates. An absent id maps to the create sentinel
 * (`metadata: null, vector: null`).
 */
export interface CommitBeforeImages {
  nouns: Map<string, GenerationRecord>
  verbs: Map<string, GenerationRecord>
}

/** Storage-root-relative path of the persisted generation counter. */
export const GENERATION_COUNTER_PATH = '_system/generation.json'
/** Storage-root-relative path of the commit manifest. */
export const MANIFEST_PATH = '_system/manifest.json'
/** Storage-root-relative prefix of the per-generation record directories. */
export const GENERATIONS_PREFIX = '_generations'

/**
 * @description Phases of the {@link GenerationStore.commitTransaction} commit
 * protocol at which a test-only fault injector can simulate a process crash.
 * The phases map 1:1 onto the protocol steps documented on
 * `commitTransaction`:
 *
 * - `'after-staging'` — before-images + the `tx.json` delta are written and
 *   fsynced; the operation batch has NOT executed yet.
 * - `'after-execute'` — the operation batch has been applied to canonical
 *   storage and indexes; the manifest still points at the prior generation.
 * - `'before-manifest-rename'` — the batch has executed and the counter is
 *   persisted; the atomic manifest rename — the commit point — has NOT
 *   happened. A crash here is the canonical "fully staged, never committed"
 *   case that recovery must roll back byte-identically.
 * - `'after-manifest-rename'` — the manifest rename landed (the transaction
 *   IS committed); the tx-log append has NOT happened yet. A crash here must
 *   keep the transaction (the tx-log is advisory metadata, not the source of
 *   commit truth).
 */
export type CommitFaultPhase =
  | 'after-staging'
  | 'after-execute'
  | 'before-manifest-rename'
  | 'after-manifest-rename'

/**
 * @description Identifies which ids a transaction touches, split by kind.
 */
export interface TouchedIds {
  /** Entity ids the transaction writes or deletes. */
  nouns: string[]
  /** Relationship ids the transaction writes or deletes. */
  verbs: string[]
}

/**
 * @description Result of {@link GenerationStore.open} — how many uncommitted
 * generations crash recovery rolled back (a non-zero value tells `Brainy` to
 * force an index reconciliation, since derived index state may reference the
 * rolled-back writes).
 */
export interface GenerationStoreOpenResult {
  /** Count of uncommitted generation directories rolled back and removed. */
  rolledBackGenerations: number
}

/**
 * @description The generational MVCC record layer. See the module
 * documentation for the full protocol; every public method documents its
 * own contract.
 */
export class GenerationStore {
  private readonly storage: GenerationStorage

  /**
   * The generation FACT LOG (dual-write transition) — an append-only,
   * CRC-framed record of every committed generation as an AFTER-IMAGE fact.
   * `null` when the storage layer lacks the binary raw-byte primitives.
   * Appends ride the same commit protocol: a fact-append failure FAILS the
   * write (loud — a silent fact gap would make the log a lie that a later
   * replay discovers), and open() reconciles the log to committed truth.
   */
  private factLog: FactLog | null = null

  /** Latest reserved/observed generation (≥ {@link committed}). */
  private counter = 0
  /** Committed-transaction watermark (manifest generation). */
  private committed = 0
  /** Compaction horizon — record-sets ≤ this are reclaimed. */
  private horizonGen = 0

  /**
   * Committed generations whose record dirs exist, stored as a SORTED, DISJOINT,
   * ascending list of INCLUSIVE `[start, end]` intervals (a run-length set).
   *
   * Committed generations come from the monotonic `counter++`, so they are dense
   * contiguous integers EXCEPT where {@link compact} punches a hole (reclaiming
   * an oldest contiguous prefix). Storing them as intervals makes the resident
   * size O(number-of-compaction-gaps) — typically a SINGLE interval
   * `[firstGen, lastGen]` — instead of one array element per generation. A
   * billion-insert corpus (every single-op reserves a distinct generation) would
   * otherwise hold a ~1B-element array resident for the process lifetime; the
   * interval form collapses that to O(1). All access goes through the helpers
   * ({@link appendCommittedGen}, {@link committedGensAsc}, {@link committedCount},
   * {@link removeCommittedUpTo}, {@link lastCommittedGen},
   * {@link largestReservedAtOrBefore}) so the multiset and ordering are identical
   * to the old flat array at every point.
   */
  private committedRanges: Array<[number, number]> = []
  /** Delta cache, keyed by generation (lazily loaded from `tx.json`). `bytes`
   *  is the serialized record-set size, summed by {@link historyBytes} for the
   *  `maxBytes` / adaptive retention caps (0 for generations written before
   *  byte accounting, or for un-flushed pending generations). */
  private readonly deltaCache = new Map<
    number,
    { nouns: Set<string>; verbs: Set<string>; timestamp: number; bytes: number }
  >()

  /**
   * Per-id inverted history index (Model-B scalability) — `id → ascending
   * generations that touched it`, one map per kind. {@link resolveAt} binary-
   * searches the id's OWN chain (O(log chain)) for the first generation after
   * the pin, instead of linearly scanning the global committed-generation set
   * ({@link committedRanges})
   * (which is O(database-age) — confirmed by the scalability spike: a read of an
   * unchanged entity at an old pin scaled 11.9x for 10x history depth). This
   * mirrors cor's `delta_history` chain.
   *
   * BOUNDED-RAM design (Approach C — hot-tail window + cold LRU). The old
   * unbounded `Map<id, number[]>` held ONE resident chain per id ever touched
   * across retained history → O(distinct-ids-ever-touched) RAM, which defeats
   * billion-scale time travel (the chains, not the data, become the heap floor).
   * These three structures replace it with RAM bounded by the knobs W/L,
   * INDEPENDENT of the corpus size:
   *
   * 1. {@link recentNounChains}/{@link recentVerbChains} — FULL per-id chains,
   *    but only over the HOT TAIL window `[windowLo, counter]` (the last
   *    {@link recentWindowGenerations} generations). Pins at-or-above
   *    `windowLo` (the overwhelming common case — recent history) resolve here
   *    with the same O(log chain) binary search as before.
   * 2. {@link coldNounChains}/{@link coldVerbChains} — a bounded LRU of FULL
   *    per-id chains for DEEP pins (`gen < windowLo`). Reconstructed on demand
   *    from the persisted deltas ({@link reconstructColdChain}), cached so a
   *    re-read is O(log chain), evicted oldest-first past
   *    {@link coldChainLruMax}. Empty chains are cached too (bounded negative
   *    cache for untouched-in-cold-region ids).
   * 3. {@link windowNounDeltas}/{@link windowVerbDeltas} — the window's INVERSE
   *    index (`gen → ids touched at that gen`), fed from the touched sets
   *    already handed to {@link extendChains}. It makes sliding `windowLo`
   *    forward an O(d̄) front-trim of exactly the ids leaving the window, with
   *    ZERO disk I/O — the write path never reads `tx.json` on a slide.
   *
   * INVARIANT — purely derived, NEVER persisted. All three structures (and the
   * cold LRU) are reconstructable from the on-disk deltas at any time; eviction
   * (cold LRU) and window slide-out drop entries that are then rebuilt on the
   * next read that needs them. Correctness across eviction/reconstruction is
   * guaranteed by pin-gating: a live historical `Db` pins at `g_old`, and the
   * before-image a read needs lives at the FIRST generation after `g_old`
   * (`> g_old ≥ minPinned`), which compaction can never reclaim while the pin is
   * held — so a reconstructed chain yields the identical answer as the original.
   */
  private readonly recentNounChains = new Map<string, number[]>()
  private readonly recentVerbChains = new Map<string, number[]>()
  /** Window inverse index (`gen → ids touched`), one map per kind — the O(d̄),
   *  I/O-free slide-trim driver. Holds exactly the gens in `[windowLo, counter]`. */
  private readonly windowNounDeltas = new Map<number, string[]>()
  private readonly windowVerbDeltas = new Map<number, string[]>()
  /** Bounded LRU of reconstructed deep-pin chains (`gen < windowLo`). JS Map
   *  insertion order is the LRU order: a hit re-inserts (delete+set) to mark
   *  MRU; eviction removes `keys().next().value` (the oldest) past the cap. */
  private readonly coldNounChains = new Map<string, number[]>()
  private readonly coldVerbChains = new Map<string, number[]>()
  /** Inclusive lower bound of the resident hot-tail window (`0` until built). */
  private windowLo = 0
  /** Whether the hot-tail window + its inverse index are built and current. */
  private windowReady = false
  /** De-dupes concurrent first-callers of {@link ensureWindow}. */
  private windowBuilding: Promise<void> | null = null

  /**
   * Hot-tail window width W — how many of the newest generations keep a resident
   * per-id chain ({@link recentNounChains}). Bounds recent-chain + window-delta
   * RAM to O(W·d̄), independent of corpus size. Deep pins below the window fall to
   * the cold LRU. Not `readonly` so tests can shrink it to exercise the
   * cold/slide paths without building thousands of generations.
   */
  private recentWindowGenerations = 4096

  /**
   * Cold-chain LRU capacity L — how many reconstructed deep-pin chains stay
   * resident. Bounds cold-chain RAM to O(L·avg-chain-length), independent of how
   * many distinct ids are deep-read. Not `readonly` so tests can shrink it to
   * exercise eviction + reconstruction.
   */
  private coldChainLruMax = 4096

  /**
   * Cap on resident {@link deltaCache} entries (Model-B RAM bound). The per-id
   * {@link recentNounChains}/{@link recentVerbChains} window is the hot read
   * structure; the raw
   * deltas are only needed for range queries ({@link changedBetween}, `diff`,
   * `since`) and chain (re)builds, and {@link getDelta} transparently re-reads an
   * evicted delta from storage. Bounding this keeps a long-lived high-write
   * process's heap O(cap) instead of O(generations) on the disk-backed path.
   * Not `readonly` so tests can lower it to exercise the eviction/re-read path
   * without building thousands of generations.
   */
  private deltaCacheMax = 4096

  /**
   * Running total of on-disk history bytes across committed generations —
   * `null` until {@link historyBytes} pays its one seeding walk. Maintained
   * incrementally at commit/reclaim so the adaptive retention check on every
   * flush() is O(1), never a tail re-walk. Never updated by cache re-reads
   * ({@link setDelta} inserts are cache population, not new history).
   */
  private historyBytesTotal: number | null = null

  /**
   * Model-B per-write group-commit — the in-memory PENDING tier.
   *
   * Each single-operation write reserves its OWN generation (the locked
   * cross-project contract: "every write versioned; a distinct generation per
   * single-op"), captures byte-identical before-images, and buffers them here
   * instead of paying a synchronous per-write manifest fsync (the 3-5x write
   * regression). {@link flushPendingSingleOps} persists the whole buffer to
   * disk in ONE fsync (durability batching — NOT a generation collapse).
   *
   * Crucially, these pending generations participate in point-in-time
   * resolution EXACTLY like committed ones ({@link resolveAt}, the per-id
   * chains, {@link changedBetween}, {@link hasCommittedAfter} all read the
   * union via {@link reservedGensAsc}; before-images resolve from this buffer
   * while pending, from disk once flushed). That is what lets the SYNCHRONOUS
   * `now()` pin freeze against un-flushed single-ops with no forced flush.
   *
   * Durability boundary: a hard crash before a flush loses only the buffered
   * *history* of un-flushed writes (live data is already in canonical storage);
   * a crash mid-flush is recovered by drop-without-restore (see
   * {@link GenerationDelta.groupCommit}).
   */
  private pendingGens: number[] = []
  private readonly pendingBuffer = new Map<
    number,
    { nouns: Map<string, GenerationRecord>; verbs: Map<string, GenerationRecord>; timestamp: number }
  >()
  /** Pending timer-coalesced flush handle (cleared on flush/close). */
  private pendingFlushTimer: ReturnType<typeof setTimeout> | null = null
  /**
   * Flush the pending tier once it holds this many generations (size trigger),
   * complementing the {@link PENDING_FLUSH_DELAY_MS} timer trigger. Not
   * `readonly` so tests can drive the boundary deterministically.
   */
  private pendingFlushThreshold = 256
  /** Coalescing window (ms) before an idle pending tier is flushed. */
  private static readonly PENDING_FLUSH_DELAY_MS = 50

  /**
   * Latched pending-flush durability failure. Set once background flushes have
   * failed {@link PENDING_FLUSH_FAILURE_THRESHOLD} consecutive times (a real,
   * persistent storage fault — not a blip); while set, writes are REFUSED with
   * a {@link PendingFlushDurabilityError} rather than silently accumulate
   * un-durable history in memory. Cleared the moment a flush finally succeeds
   * (the store self-heals). `null` = history-durability path is healthy.
   */
  private pendingFlushError: Error | null = null
  /** Consecutive failed background pending-flush attempts (resets on success). */
  private pendingFlushFailures = 0
  /**
   * Consecutive failed flush attempts before the durability latch trips and
   * writes start refusing. A small tolerance so a single transient fault does
   * not trip the alarm, while a genuinely stuck disk stops the bleed fast.
   */
  private static readonly PENDING_FLUSH_FAILURE_THRESHOLD = 3
  /** Upper bound (ms) on the exponential retry backoff between failed flushes. */
  private static readonly PENDING_FLUSH_RETRY_CAP_MS = 30_000

  /** Live pin refcounts, keyed by pinned generation. */
  private readonly pins = new Map<number, number>()

  /** Serializes transact commits, compaction, and counter persistence. */
  private mutexTail: Promise<unknown> = Promise.resolve()

  /** True while a transact batch is executing (suppresses the bump hook). */
  private inTransact = false

  /** Coalescing state for lazy persistence of single-op counter bumps. */
  private counterPersistScheduled = false

  /**
   * Test-only fault injector for the commit protocol (see
   * {@link GenerationStore.setCommitFaultInjector}). `undefined` in production.
   */
  private commitFaultInjector?: (phase: CommitFaultPhase) => void

  private opened = false

  /**
   * @param storage - The storage adapter, narrowed to the
   *   {@link GenerationStorage} contract (`BaseStorage` implements it).
   */
  constructor(storage: GenerationStorage) {
    this.storage = storage
  }

  // ==========================================================================
  // Lifecycle
  // ==========================================================================

  /**
   * @description Load the persisted counter + manifest and run crash
   * recovery. Must be called once, before indexes are built, because
   * recovery may rewrite canonical entity files (restoring before-images of
   * an uncommitted transaction).
   *
   * @param options.readOnly - When true (reader-mode brains), recovery is
   *   skipped: readers never write. An un-repaired crashed directory is
   *   repaired by the next *writer* open.
   * @returns How many uncommitted generations were rolled back.
   */
  async open(options?: { readOnly?: boolean }): Promise<GenerationStoreOpenResult> {
    const counterFile = (await this.storage.readRawObject(GENERATION_COUNTER_PATH)) as
      | { generation?: number }
      | null
    const manifest = (await this.storage.readRawObject(MANIFEST_PATH)) as GenerationManifest | null

    this.committed = manifest?.generation ?? 0
    this.horizonGen = manifest?.horizon ?? 0
    this.counter = Math.max(counterFile?.generation ?? 0, this.committed)

    // Discover existing generation record directories.
    const recordPaths = await this.storage.listRawObjects(GENERATIONS_PREFIX)
    const seenGens = new Set<number>()
    for (const p of recordPaths) {
      const gen = parseGenerationFromPath(p)
      if (gen !== null) seenGens.add(gen)
    }

    let rolledBack = 0
    // Coalesce the ascending on-disk committed gens into interval form: each
    // contiguous run becomes a single `[start, end]` range (appendCommittedGen
    // extends the last range on a +1 step, opens a new one across a gap).
    this.committedRanges = []
    for (const gen of [...seenGens].sort((a, b) => a - b)) {
      if (gen <= this.committed) {
        this.appendCommittedGen(gen)
      } else if (!options?.readOnly) {
        // Uncommitted (crashed) generation. A transact generation is rolled
        // back by RESTORING its before-images (its before-image + execute are
        // one atomic unit). A single-op group-commit generation is DROPPED
        // WITHOUT RESTORE — its live write already landed and was acknowledged
        // before the flush, so restoring would silently revert it. The branch
        // is decided by the persisted delta's `groupCommit` flag.
        await this.recoverUncommittedGeneration(gen)
        rolledBack++
      }
      // Reader mode: leave orphan dirs alone; resolution ignores them because
      // committedRanges only includes generations ≤ the manifest watermark.
      this.counter = Math.max(this.counter, gen)
    }
    // The history window is rebuilt lazily from the (possibly changed) generation
    // set on the next historical read — covers reopen and reopen-after-restore.
    this.invalidateChains()

    if (rolledBack > 0) {
      prodLog.warn(
        `[GenerationStore] Crash recovery rolled back ${rolledBack} uncommitted ` +
          `transaction(s); store restored to generation ${this.committed}.`
      )
    }

    this.opened = true

    // Generation FACT LOG (dual-write transition): when the storage layer
    // exposes the binary raw-byte primitives, open the after-image fact log
    // and reconcile it to committed truth — facts are appended BEFORE the
    // commit point, so a crash can only leave the log AHEAD; open truncates
    // any fact beyond `committed`. Storage without the primitives simply
    // hosts no fact log (readers fall back to canonical enumeration).
    if (storageSupportsFactLog(this.storage)) {
      this.factLog = new FactLog(this.storage)
      await this.factLog.open(this.committed)
    } else {
      this.factLog = null
    }

    // Hook single-op write batches so generation() is always meaningful.
    // Suppressed while a transact batch executes (the batch is ONE generation).
    if (!options?.readOnly) {
      this.storage.setGenerationBumpHook(() => this.noteSingleOpWrite())
    }

    return { rolledBackGenerations: rolledBack }
  }

  /**
   * @description Detach the storage hook and persist the counter. Called
   * from `brain.close()`.
   */
  async close(): Promise<void> {
    // Persist any buffered single-op history before detaching, so a clean close
    // never loses generations the caller already observed.
    this.clearPendingFlushTimer()
    await this.flushPendingSingleOps()
    this.storage.setGenerationBumpHook(undefined)
    await this.persistCounterNow()
  }

  /**
   * @description TEST-ONLY: install (or clear, with `undefined`) a fault
   * injector that is invoked at each {@link CommitFaultPhase} of the commit
   * protocol. A **throwing** injector simulates a process crash at that
   * phase: the abort cleanup (staging-directory removal, generation-counter
   * reservation return) is deliberately skipped — exactly as if the process
   * had died — so crash-consistency tests exercise the REAL recovery path
   * ({@link GenerationStore.open} rolling back the uncommitted generation on
   * the next open). Never installed in production code; the injector exists
   * solely so the durability protocol can be proven, not trusted.
   *
   * @param injector - Callback invoked synchronously at each phase, or
   *   `undefined` to detach.
   */
  setCommitFaultInjector(injector: ((phase: CommitFaultPhase) => void) | undefined): void {
    this.commitFaultInjector = injector
  }

  // ==========================================================================
  // Generation counter
  // ==========================================================================

  /** @returns The store's current generation (latest write watermark). */
  generation(): number {
    return this.counter
  }

  /** @returns The committed-transaction watermark (manifest generation). */
  committedGeneration(): number {
    return this.committed
  }

  /** @returns The compaction horizon (generations ≤ horizon are reclaimed). */
  horizon(): number {
    return this.horizonGen
  }

  /**
   * @description Read-only history footprint for fleet audits: how much
   * generational history this store holds on disk. `bytes` pays (and seeds)
   * the one-time {@link historyBytes} walk on first call — subsequent calls
   * are O(1). The oldest/newest timestamps come from those generations'
   * deltas (cache-bounded reads).
   * @returns Counts, bytes, generation range, and the compaction horizon.
   */
  async historyStats(): Promise<{
    generations: number
    bytes: number
    oldestGeneration: number | null
    newestGeneration: number | null
    oldestTimestamp: number | null
    newestTimestamp: number | null
    horizon: number
  }> {
    let oldest: number | null = null
    let newest: number | null = null
    for (const gen of this.committedGensAsc()) {
      if (oldest === null) oldest = gen
      newest = gen
    }
    return {
      generations: this.committedCount(),
      bytes: await this.historyBytes(),
      oldestGeneration: oldest,
      newestGeneration: newest,
      oldestTimestamp: oldest !== null ? (await this.getDelta(oldest)).timestamp : null,
      newestTimestamp: newest !== null ? (await this.getDelta(newest)).timestamp : null,
      horizon: this.horizonGen
    }
  }

  /**
   * @description Read one generation's persisted before-image records — the
   * compaction fallback for generations written before deltas carried
   * `blobHashes`. O(that generation's records).
   * @param gen - The generation whose `prev/` records to read.
   * @returns The record-set (empty when absent).
   */
  private async readGenerationRecords(gen: number): Promise<GenerationRecord[]> {
    let paths: string[] = []
    try {
      paths = await this.storage.listRawObjects(`${GENERATIONS_PREFIX}/${gen}/prev`)
    } catch {
      return []
    }
    const records: GenerationRecord[] = []
    for (const p of paths) {
      const record = (await this.storage.readRawObject(p)) as GenerationRecord | null
      if (record) records.push(record)
    }
    return records
  }

  /**
   * @description Single-operation write hook (registered with the storage
   * layer in {@link open}). Bumps the in-memory counter and schedules a
   * coalesced persist of `_system/generation.json` — durable artifacts
   * (records, manifests, snapshots) always persist the counter
   * synchronously at their own commit points, so a crash inside the
   * coalescing window can never move the counter backwards relative to
   * anything durable.
   */
  private noteSingleOpWrite(): void {
    if (this.inTransact) return
    this.counter++
    this.schedulePersistCounter()
  }

  /** Schedule one coalesced counter persist for the current write burst. */
  private schedulePersistCounter(): void {
    if (this.counterPersistScheduled) return
    this.counterPersistScheduled = true
    setImmediate(() => {
      this.counterPersistScheduled = false
      void this.withMutex(() => this.persistCounterUnlocked()).catch((err) => {
        prodLog.warn(`[GenerationStore] generation counter persist failed: ${(err as Error).message}`)
      })
    })
  }

  /**
   * @description Persist the generation counter now (atomic tmp+rename).
   * Called from `flush()`, `close()`, snapshotting, and the commit protocol.
   */
  async persistCounterNow(): Promise<void> {
    if (!this.opened) return
    await this.withMutex(() => this.persistCounterUnlocked())
  }

  private async persistCounterUnlocked(): Promise<void> {
    await this.storage.writeRawObject(GENERATION_COUNTER_PATH, {
      generation: this.counter,
      updatedAt: new Date().toISOString()
    })
  }

  // ==========================================================================
  // Pins
  // ==========================================================================

  /**
   * @description Pin a generation (refcounted). Compaction never reclaims a
   * record-set a live pin could need (i.e. any generation directory `N`
   * with `N > G` for some pinned `G`).
   * @param gen - The generation to pin.
   */
  pin(gen: number): void {
    this.pins.set(gen, (this.pins.get(gen) ?? 0) + 1)
  }

  /**
   * @description Release one pin on a generation. Safe to call for a
   * generation with no live pins (no-op with a warning).
   * @param gen - The generation to release.
   */
  release(gen: number): void {
    const count = this.pins.get(gen)
    if (count === undefined) {
      prodLog.warn(`[GenerationStore] release(${gen}) called with no live pin at that generation`)
      return
    }
    if (count <= 1) this.pins.delete(gen)
    else this.pins.set(gen, count - 1)
  }

  /** @returns Total number of live pins across all generations. */
  activePinCount(): number {
    let total = 0
    for (const count of this.pins.values()) total += count
    return total
  }

  /** @returns The smallest pinned generation, or `Infinity` with no pins. */
  private minPinnedGeneration(): number {
    let min = Infinity
    for (const gen of this.pins.keys()) min = Math.min(min, gen)
    return min
  }

  // ==========================================================================
  // Commit protocol
  // ==========================================================================

  /**
   * @description Commit one transaction. The caller (Brainy.transact) plans
   * the batch up front — resolving every touched id, including generated
   * entity/relationship ids and delete cascades — and hands this method the
   * touched-id sets plus an `execute` closure that runs the already-built
   * operation batch through the TransactionManager.
   *
   * Protocol (under the commit mutex):
   * 1. `ifAtGeneration` CAS check (rejects before anything is staged).
   * 2. Reserve generation `N` (counter increment).
   * 3. Capture before-images of every touched id and write them, plus the
   *    `tx.json` delta, under `_generations/N/`; fsync. From this point a
   *    crash anywhere is recoverable to the exact pre-transaction bytes.
   * 4. Run `execute()`. On failure the TransactionManager has already
   *    rolled back its applied operations; the staging directory is removed
   *    and the reservation is returned (when no concurrent bump consumed a
   *    later number), so a failed transaction leaves the generation
   *    unchanged.
   * 5. Persist the counter, then atomically rename the manifest to
   *    generation `N` — **the rename is the commit point** — and fsync it.
   * 6. Append the tx-log line and update in-memory bookkeeping.
   *
   * @param args.touched - Every id the batch writes or deletes.
   * @param args.meta - Transaction metadata for the tx-log.
   * @param args.ifAtGeneration - Whole-store CAS expectation.
   * @param args.execute - Runs the planned operation batch atomically.
   * @returns The committed generation and its commit timestamp.
   * @throws GenerationConflictError when the CAS expectation fails.
   */
  /**
   * The generation fact log, or `null` when the storage layer cannot host one.
   * Consumers scan committed facts through it (`scanFacts` / `segmentPaths`).
   */
  getFactLog(): FactLog | null {
    return this.factLog
  }

  /**
   * @description Build one commit's AFTER-IMAGE fact by reading canonical
   * state back for every touched id — under the commit mutex, immediately
   * after the operations applied, so canonical IS the after-image (and the
   * reads are page-cache-warm: the operations just wrote these files). An
   * absent id (both legs null) becomes a body-less TOMBSTONE — the delete
   * fact needs no body, so removal never requires reading the removed thing.
   * The fact's blobHashes are extracted from the AFTER records (the content
   * this generation's state references), unlike the history path's
   * before-image hashes.
   */
  private async buildCommitFact(args: {
    generation: number
    timestamp: number
    nouns: string[]
    verbs: string[]
    meta?: Record<string, unknown>
  }): Promise<CommitFact> {
    const ops: FactOp[] = []
    const afterRecords: GenerationRecord[] = []
    for (const id of args.nouns) {
      const after = await this.storage.readNounRaw(id)
      const absent = after.metadata === null && after.vector === null
      ops.push({ kind: 'noun', id, record: absent ? null : after })
      if (!absent) afterRecords.push({ kind: 'noun', metadata: after.metadata, vector: after.vector })
    }
    for (const id of args.verbs) {
      const after = await this.storage.readVerbRaw(id)
      const absent = after.metadata === null && after.vector === null
      ops.push({ kind: 'verb', id, record: absent ? null : after })
      if (!absent) afterRecords.push({ kind: 'verb', metadata: after.metadata, vector: after.vector })
    }
    const blobHashes = this.storage.extractBlobHashesFromRecords
      ? this.storage.extractBlobHashesFromRecords(afterRecords)
      : []
    return {
      generation: args.generation,
      timestamp: args.timestamp,
      ops,
      ...(args.meta ? { meta: args.meta } : {}),
      ...(blobHashes.length > 0 ? { blobHashes } : {})
    }
  }

  async commitTransaction(args: {
    touched: TouchedIds
    meta?: Record<string, unknown>
    ifAtGeneration?: number
    /** Optional compare-and-swap precondition over the {@link CommitBeforeImages},
     *  run under the commit mutex BEFORE anything is staged or applied — the
     *  per-record analogue of `ifAtGeneration`. A throw aborts the whole batch:
     *  the generation reservation is returned and no staging I/O has happened. */
    precommit?: (before: CommitBeforeImages) => void
    execute: () => Promise<void>
  }): Promise<{ generation: number; timestamp: number }> {
    return this.withMutex(async () => {
      // A latched history-durability failure compromises the whole generation
      // spine — refuse a transact too (advancing the manifest past stuck,
      // un-durable single-op generations would be inconsistent). Same loud
      // error; self-clears when the pending tier drains.
      this.assertHistoryDurable()
      if (args.ifAtGeneration !== undefined && args.ifAtGeneration !== this.counter) {
        throw new GenerationConflictError(args.ifAtGeneration, this.counter)
      }

      const nouns = [...new Set(args.touched.nouns)]
      const verbs = [...new Set(args.touched.verbs)]
      const gen = ++this.counter
      const dir = `${GENERATIONS_PREFIX}/${gen}`
      const timestamp = Date.now()

      // Test-only crash simulation (see setCommitFaultInjector): a throwing
      // injector must bypass the abort cleanup below, exactly as a real
      // process death would, so recovery-on-open is what restores the store.
      let crashSimulated = false
      const faultPoint = (phase: CommitFaultPhase): void => {
        if (!this.commitFaultInjector) return
        try {
          this.commitFaultInjector(phase)
        } catch (err) {
          crashSimulated = true
          throw err
        }
      }

      // Temporal-blob contract: hashes this record-set references, recorded
      // before staging; scoped outside the try so an abort can compensate.
      let txBlobHashes: string[] = []

      // Hoisted so the catch can reconcile canonical state against them after a
      // failed rollback (the trapdoor). Empty until populated below.
      const nounBefore = new Map<string, GenerationRecord>()
      const verbBefore = new Map<string, GenerationRecord>()

      try {
        // -- 3. Before-images + delta (the durable undo log) ------------------
        // Read every before-image FIRST, then run the caller's CAS
        // precondition against them, and only then stage to disk — so a
        // conflicting batch aborts with zero staging I/O. The maps hold the
        // byte-identical records the staged files are written from.
        for (const id of nouns) {
          const prev = await this.storage.readNounRaw(id)
          nounBefore.set(id, { kind: 'noun', metadata: prev.metadata, vector: prev.vector })
        }
        for (const id of verbs) {
          const prev = await this.storage.readVerbRaw(id)
          verbBefore.set(id, { kind: 'verb', metadata: prev.metadata, vector: prev.vector })
        }

        // Conditional commit: the per-record CAS analogue of the
        // `ifAtGeneration` check above, but against authoritative
        // before-images under the mutex. A throw lands in the catch below,
        // which returns the generation reservation; nothing was applied.
        args.precommit?.({ nouns: nounBefore, verbs: verbBefore })

        // Temporal-blob contract: count this record-set's content-hash
        // references BEFORE it is staged (over-count-only crash ordering —
        // see flushPendingSingleOps' matching note).
        txBlobHashes = this.storage.extractBlobHashesFromRecords
          ? this.storage.extractBlobHashesFromRecords([
              ...nounBefore.values(),
              ...verbBefore.values()
            ])
          : []
        if (txBlobHashes.length > 0 && this.storage.recordHistoryBlobReferences) {
          await this.storage.recordHistoryBlobReferences(txBlobHashes)
        }

        const stagedPaths: string[] = []
        let recordBytes = 0 // serialized record-set size, for retention accounting
        for (const [id, record] of nounBefore) {
          const recordPath = `${dir}/prev/${id}.json`
          await this.storage.writeRawObject(recordPath, record)
          recordBytes += serializedBytes(record)
          stagedPaths.push(recordPath)
        }
        for (const [id, record] of verbBefore) {
          const recordPath = `${dir}/prev/${id}.json`
          await this.storage.writeRawObject(recordPath, record)
          recordBytes += serializedBytes(record)
          stagedPaths.push(recordPath)
        }
        const delta: GenerationDelta = {
          generation: gen,
          timestamp,
          ...(args.meta && { meta: args.meta }),
          nouns,
          verbs,
          // Always present on new deltas (empty = "no blobs"), so compaction
          // only falls back to record reads for pre-contract generations.
          blobHashes: txBlobHashes
        }
        delta.bytes = recordBytes + serializedBytes(delta)
        const deltaPath = `${dir}/tx.json`
        await this.storage.writeRawObject(deltaPath, delta)
        stagedPaths.push(deltaPath)
        await this.storage.syncRawObjects(stagedPaths)
        faultPoint('after-staging')

        // -- 4. Execute the planned batch -------------------------------------
        // Durability barrier: record every canonical write/delete the operations
        // make, then fsync them BELOW (before the counter advances). Without
        // this, a hard kill after commitTransaction returns could leave the
        // fsync'd generation counter ahead of still-page-cached entity bytes —
        // phantom progress for any generation-based consumer.
        this.storage.beginWriteBarrier?.()
        this.inTransact = true
        try {
          await args.execute()
        } finally {
          this.inTransact = false
        }
        // The transaction's entire canonical footprint is now durable, so the
        // counter/manifest advance below can never outrun the entity bytes.
        await this.storage.flushWriteBarrier?.()
        faultPoint('after-execute')

        // Fact log (dual-write): append + fsync this generation's AFTER-IMAGE
        // fact BEFORE the commit point, inside the same durability window —
        // so a crash can only leave the log AHEAD (open truncates), never a
        // committed generation without its fact. A real abort below this point
        // compensates via dropAbove in the catch. An append failure fails the
        // write, loudly — a silent fact gap would be a lie a replay discovers.
        if (this.factLog) {
          const fact = await this.buildCommitFact({
            generation: gen,
            timestamp,
            nouns,
            verbs,
            ...(args.meta ? { meta: args.meta } : {})
          })
          await this.factLog.append(fact)
          await this.factLog.sync()
        }

        // -- 5. Counter + manifest rename (COMMIT POINT) ----------------------
        await this.persistCounterUnlocked()
        faultPoint('before-manifest-rename')
        const manifest: GenerationManifest = {
          version: 1,
          generation: gen,
          committedAt: new Date(timestamp).toISOString(),
          horizon: this.horizonGen
        }
        await this.storage.writeRawObject(MANIFEST_PATH, manifest)
        await this.storage.syncRawObjects([MANIFEST_PATH])
        faultPoint('after-manifest-rename')

        // -- 6. Post-commit bookkeeping ---------------------------------------
        this.committed = gen
        this.appendCommittedGen(gen)
        this.setDelta(gen, {
          nouns: new Set(nouns),
          verbs: new Set(verbs),
          timestamp,
          bytes: delta.bytes ?? 0
        })
        if (this.historyBytesTotal !== null) {
          this.historyBytesTotal += delta.bytes ?? 0
        }
        this.extendChains(gen, nouns, verbs)
        const logEntry: TxLogEntry = { generation: gen, timestamp, ...(args.meta && { meta: args.meta }) }
        await this.storage.appendTxLogLine(JSON.stringify(logEntry))

        return { generation: gen, timestamp }
      } catch (err) {
        // Simulated crash (test-only fault injector): skip the abort cleanup
        // entirely — a dead process cannot clean up. Recovery on the next
        // open() is what must (and does) restore the store.
        if (crashSimulated) {
          throw err
        }
        // The trapdoor for a batch: if rollback FAILED to fully apply, canonical
        // storage may be inconsistent. A batch is never adopted forward (its
        // other ops were rolled back — partial commit would break atomicity), so
        // any unreconciled record is a fail-loud StoreInconsistentError. Compute
        // it here (before the staging cleanup, which is always safe to run) and
        // throw it in place of the raw error at the end.
        let inconsistent: UnreconciledRecord[] = []
        if (err instanceof TransactionRollbackError) {
          inconsistent = await this.reconcileFailedRollback(nounBefore, verbBefore)
        }
        // Failed before the manifest rename: nothing is committed. Remove the
        // staging directory; the TransactionManager already restored any
        // applied operation byte-identically.
        try {
          await this.storage.removeRawPrefix(dir)
        } catch (cleanupErr) {
          prodLog.warn(
            `[GenerationStore] failed to remove staging dir ${dir} after aborted transaction: ` +
              `${(cleanupErr as Error).message} (recovery will remove it on next open)`
          )
        }
        // Compensate the pre-staging history-reference increments. Best
        // effort — a failure here only over-counts (a leak the scrub
        // repairs). Reclaim inside is safe on an abort: the before-image
        // hashes are the entities' still-live content, so live references
        // block any physical delete.
        if (txBlobHashes.length > 0 && this.storage.releaseHistoryBlobReferences) {
          try {
            await this.storage.releaseHistoryBlobReferences(txBlobHashes)
          } catch {
            // over-count-safe; the scrub restores exactness
          }
        }
        // Fact-log compensation: a real (non-crash) abort after the fact was
        // appended must take the fact back out — the generation never
        // committed. A crash instead reaches open(), whose truncation does the
        // same reconcile from disk.
        if (this.factLog && this.factLog.headGeneration() >= gen) {
          await this.factLog.dropAbove(gen - 1)
        }
        // Return the reservation when no concurrent bump consumed a later
        // number, so a failed transaction leaves generation() unchanged.
        if (this.counter === gen) this.counter = gen - 1
        // A failed rollback that left canonical records unreconciled surfaces as
        // a loud StoreInconsistentError (brainy quarantines writes until repair);
        // otherwise the raw error (clean abort, or a derived-only undo failure
        // the egress guard + rebuild handle).
        if (inconsistent.length > 0) {
          throw new StoreInconsistentError(
            inconsistent,
            (err as TransactionRollbackError).originalError
          )
        }
        throw err
      }
    })
  }

  /**
   * @description After a transaction's rollback FAILED to fully apply
   * (`TransactionRollbackError`), determine which touched records are now
   * unreconciled by comparing current canonical state to the before-images the
   * commit captured. This observes reality rather than trusting the opaque undo
   * closures — the authoritative signal for adopt-forward vs fail-loud.
   *
   * @param nounBefore - Byte-identical entity before-images captured pre-execute.
   * @param verbBefore - Byte-identical relationship before-images.
   * @returns Every record whose canonical state no longer matches its
   *   before-image, tagged `'orphan'` (durably present when it should be gone —
   *   an add whose delete-undo failed) or `'loss'` (durably gone/wrong when it
   *   should have been restored — a remove/update whose restore-undo failed). An
   *   empty array means canonical storage is cleanly rolled back (only a derived
   *   index undo failed — the egress guard + a rebuild handle that).
   */
  private async reconcileFailedRollback(
    nounBefore: Map<string, GenerationRecord>,
    verbBefore: Map<string, GenerationRecord>
  ): Promise<UnreconciledRecord[]> {
    const out: UnreconciledRecord[] = []
    const classify = (
      before: GenerationRecord,
      current: { metadata: unknown | null; vector: unknown | null }
    ): 'orphan' | 'loss' | null => {
      const beforeEmpty = before.metadata == null && before.vector == null
      const currentEmpty = current.metadata == null && current.vector == null
      if (beforeEmpty && currentEmpty) return null // add cleanly undone
      if (beforeEmpty) return 'orphan' // add's delete-undo failed → still present
      if (currentEmpty) return 'loss' // remove's restore-undo failed → gone
      // Both present: same value = reconciled; different = restore left a wrong value.
      const same =
        JSON.stringify({ m: before.metadata, v: before.vector }) ===
        JSON.stringify({ m: current.metadata, v: current.vector })
      return same ? null : 'loss'
    }
    for (const [id, before] of nounBefore) {
      const d = classify(before, await this.storage.readNounRaw(id))
      if (d) out.push({ id, kind: 'noun', disposition: d })
    }
    for (const [id, before] of verbBefore) {
      const d = classify(before, await this.storage.readVerbRaw(id))
      if (d) out.push({ id, kind: 'verb', disposition: d })
    }
    return out
  }

  // ==========================================================================
  // Single-operation commit (Model-B per-write group-commit)
  // ==========================================================================

  /**
   * @description Commit ONE single-operation write (`add`/`update`/`remove`/
   * `relate`/`updateRelation`/`unrelate` and the per-item calls of `addMany`/
   * `updateMany`/`relateMany`) as its own immutable generation — the Model-B
   * "every write versioned" contract. (`removeMany` commits each delete *chunk*
   * as a single generation for bulk efficiency — the one non-per-item path.) Structurally a one-operation {@link commitTransaction} with
   * **deferred durability**:
   *
   * - Under the commit mutex it reserves generation `N`, captures byte-identical
   *   before-images of every touched id (so the pin/`asOf` read source is
   *   exact), runs `execute()` (the existing single-op `executeTransaction`
   *   batch) with the storage bump hook suppressed, then BUFFERS the
   *   before-images in the in-memory pending tier and returns — **no per-write
   *   manifest fsync** (that synchronous fsync is the 3-5x write regression this
   *   design avoids).
   * - The generation is immediately visible to point-in-time reads (chains +
   *   {@link reservedGensAsc}), so a synchronous `now()` pin freezes against it
   *   with no forced flush.
   * - {@link flushPendingSingleOps} later persists the buffer to disk in one
   *   fsync (async group-commit). A crash before that flush loses only the
   *   buffered *history*; live data is already in canonical storage. A crash
   *   mid-flush is recovered by drop-without-restore
   *   (see {@link GenerationDelta.groupCommit}).
   *
   * **Durability contract (single-op vs transact).** A single-op write's live
   * canonical bytes are written via tmp+rename but NOT individually fsync'd —
   * they are durable at the next {@link flushPendingSingleOps} or `close()`, not
   * the instant the write resolves. On a hard kill before that flush, a
   * single-op write can be lost even though the call returned; the group-commit
   * counter is buffered alongside, so counter and data are lost together (no
   * counter-ahead-of-state torn store). {@link commitTransaction} is the
   * stronger contract: it runs a write barrier that fsyncs the whole batch's
   * canonical footprint BEFORE advancing the generation counter, so a committed
   * transact is durable on return. Callers that need per-write durability should
   * use `transact()` (or `flush()` after a single-op); Model-B group-commit
   * deliberately trades single-op fsync latency (a 3-5x write regression) for
   * throughput.
   *
   * The per-write generation is read by graph-index ops through the same
   * `generation()` watermark `transact()` uses — because this method, like
   * {@link commitTransaction}, holds the mutex across the counter bump AND
   * `execute()`, so the watermark equals `N` for the duration of the write (a
   * distinct generation threaded to cor per single-op, the locked contract).
   *
   * @param args.touched - The ids this write creates/updates/deletes, by kind.
   * @param args.execute - Runs the single-op's existing operation batch.
   * @param args.precommit - Optional compare-and-swap precondition, invoked
   *   under the commit mutex with the just-read {@link CommitBeforeImages} and
   *   BEFORE `execute()`. A throw aborts the commit atomically: the generation
   *   reservation is returned and nothing is applied or buffered. This is the
   *   one place a per-record CAS (e.g. `ifRev`) is race-free — any check done
   *   before this mutex can interleave with a concurrent writer.
   * @returns The reserved generation and its timestamp.
   */
  async commitSingleOp(args: {
    touched: { nouns?: string[]; verbs?: string[] }
    execute: () => Promise<void>
    precommit?: (before: CommitBeforeImages) => void
  }): Promise<{ generation: number; timestamp: number; degraded?: string[] }> {
    return this.withMutex(async () => {
      // Refuse to accept a write whose history we cannot make durable: if the
      // pending-tier flush has latched a persistent failure, accepting more
      // writes would silently pile un-durable before-images into memory. Fail
      // loud instead (the latch self-clears when a flush finally succeeds).
      this.assertHistoryDurable()
      const nouns = args.touched.nouns ? [...new Set(args.touched.nouns)] : []
      const verbs = args.touched.verbs ? [...new Set(args.touched.verbs)] : []
      const gen = ++this.counter
      const timestamp = Date.now()

      // Before-images BEFORE execute overwrites live storage (mirrors
      // commitTransaction's staging, but buffered in memory — durability is
      // deferred to the flush). readNounRaw/readVerbRaw on an absent id return
      // {metadata:null, vector:null} = the create sentinel.
      const nounBefore = new Map<string, GenerationRecord>()
      for (const id of nouns) {
        const prev = await this.storage.readNounRaw(id)
        nounBefore.set(id, { kind: 'noun', metadata: prev.metadata, vector: prev.vector })
      }
      const verbBefore = new Map<string, GenerationRecord>()
      for (const id of verbs) {
        const prev = await this.storage.readVerbRaw(id)
        verbBefore.set(id, { kind: 'verb', metadata: prev.metadata, vector: prev.vector })
      }

      // Conditional commit: the caller's CAS precondition runs here — under
      // the mutex, against the authoritative before-images, before any write.
      // A throw aborts cleanly: return the generation reservation (nothing was
      // applied or buffered) and surface the conflict to the caller.
      if (args.precommit) {
        try {
          args.precommit({ nouns: nounBefore, verbs: verbBefore })
        } catch (err) {
          if (this.counter === gen) this.counter = gen - 1
          throw err
        }
      }

      // Execute the live write. inTransact suppresses the storage bump hook so
      // the counter is not double-advanced (this generation already reserved
      // it) — the same suppression transact() uses.
      this.inTransact = true
      try {
        await args.execute()
      } catch (err) {
        this.inTransact = false
        // A failed rollback (TransactionRollbackError) may have left canonical
        // storage inconsistent — the trapdoor. Reconcile against the
        // before-images to decide the honest response (David's ruling:
        // adopt-forward when safe, else fail loud).
        if (err instanceof TransactionRollbackError) {
          const records = await this.reconcileFailedRollback(nounBefore, verbBefore)
          if (records.length > 0 && records.every((r) => r.disposition === 'orphan')) {
            // ADOPT FORWARD: the durably-present orphan(s) are exactly what this
            // single-op add wrote. Keep + buffer the generation so the record is
            // legitimately committed and readable; the derived index may be
            // incomplete for these ids until the next rebuild/repairIndex (the
            // egress guard prevents wrong results meanwhile). Loud, honest,
            // no double-write.
            this.pendingBuffer.set(gen, { nouns: nounBefore, verbs: verbBefore, timestamp })
            this.pendingGens.push(gen)
            this.extendChains(gen, nouns, verbs)
            // The adopted generation is committed — it gets its fact like any
            // other (durability rides the group-commit flush, same as the
            // buffered history).
            if (this.factLog) {
              await this.factLog.append(
                await this.buildCommitFact({ generation: gen, timestamp, nouns, verbs })
              )
            }
            prodLog.warn(
              `[GenerationStore] Recovered a failed rollback FORWARD: single-op write ` +
                `committed as generation ${gen} because its canonical undo could not be ` +
                `applied (record(s) ${records.map((r) => r.id).join(', ')} are durable). ` +
                `The derived index may be incomplete for these ids — run repairIndex() to heal.`
            )
            return { generation: gen, timestamp, degraded: records.map((r) => r.id) }
          }
          if (records.length > 0) {
            // FAIL LOUD: a loss (or mixed) cannot be safely adopted. Return the
            // reservation and throw — brainy quarantines writes until repair.
            if (this.counter === gen) this.counter = gen - 1
            throw new StoreInconsistentError(records, err.originalError)
          }
          // records.length === 0: canonical is cleanly rolled back (only a
          // derived undo failed) — fall through to the clean-abort path below.
        }
        // Live write failed with canonical cleanly rolled back: nothing buffered,
        // nothing durable. Return the reservation (when no concurrent bump
        // consumed a later number) so generation() is unchanged.
        if (this.counter === gen) this.counter = gen - 1
        throw err
      }
      this.inTransact = false

      // Buffer the pending generation + make it instantly visible to reads.
      this.pendingBuffer.set(gen, { nouns: nounBefore, verbs: verbBefore, timestamp })
      this.pendingGens.push(gen)
      this.extendChains(gen, nouns, verbs)
      // Fact log (dual-write): the acked write's AFTER-IMAGE fact, appended
      // now (read back warm, under the mutex — group-commit means flush-time
      // canonical only holds the LATEST state, so each generation's after-image
      // exists only here). Durability rides the group-commit flush, exactly
      // like the buffered before-image history: a crash before the flush loses
      // the fact AND the generation together — never a torn state.
      if (this.factLog) {
        await this.factLog.append(
          await this.buildCommitFact({ generation: gen, timestamp, nouns, verbs })
        )
      }
      this.schedulePendingFlush()
      return { generation: gen, timestamp }
    })
  }

  /**
   * @description Run a write WITHOUT creating a generation — for init-time /
   * infrastructure writes (e.g. the VFS root directory) that constitute the
   * generation-0 baseline of a freshly-materialized brain rather than a user
   * mutation. The storage bump hook is suppressed (`inTransact`) so the
   * generation counter does not advance, so a brand-new brain reports
   * `generation() === 0` and an empty `transactionLog()`, and the first USER
   * write is generation 1. Mirrors Datomic semantics where the empty database
   * value is the starting point and bootstrap is not a transaction.
   * @param execute - The infrastructure write to apply to the baseline.
   */
  async runWithoutGeneration(execute: () => Promise<void>): Promise<void> {
    this.inTransact = true
    try {
      await execute()
    } finally {
      this.inTransact = false
    }
  }

  /**
   * @description Persist the in-memory pending single-op generations to disk in
   * ONE fsync (async group-commit) and advance the committed watermark. Called
   * on the size/timer triggers and forced by `flush()`/`close()`/`transact()`/
   * `compactHistory()` (so generations stay strictly ordered and durable before
   * any of those). A no-op when the pending tier is empty.
   *
   * Each pending generation is written as a normal `_generations/<N>/` record
   * set (`prev/<id>.json` before-images + `tx.json` delta) but with the delta's
   * `groupCommit: true` flag set — the drop-without-restore recovery marker.
   * The single manifest rename to the highest generation commits the whole
   * batch atomically (committed iff `gen ≤ manifest.generation`).
   */
  async flushPendingSingleOps(): Promise<void> {
    if (this.pendingGens.length === 0) return
    // Centralize durability accounting here so EVERY failure path — the
    // background scheduler AND an explicit flush()/close() — latches
    // consistently, and success (from any caller) clears the latch. The
    // scheduler owns only the retry cadence.
    try {
      await this.flushPendingSingleOpsUnlocked()
      this.onPendingFlushSuccess()
    } catch (err) {
      this.recordPendingFlushFailure(err as Error)
      throw err
    }
  }

  /** The actual pending-tier flush under the mutex (see {@link flushPendingSingleOps},
   *  which wraps this with durability accounting). */
  private async flushPendingSingleOpsUnlocked(): Promise<void> {
    return this.withMutex(async () => {
      if (this.pendingGens.length === 0) return
      this.clearPendingFlushTimer()

      const gens = [...this.pendingGens].sort((a, b) => a - b)
      const stagedPaths: string[] = []
      const logEntries: TxLogEntry[] = []
      const genBytes = new Map<number, number>() // for the post-commit setDelta
      for (const gen of gens) {
        const buf = this.pendingBuffer.get(gen)
        if (!buf) continue
        const dir = `${GENERATIONS_PREFIX}/${gen}`

        // Temporal-blob contract: count this record-set's content-hash
        // references BEFORE persisting it (a crash between the two only
        // over-counts — the scrub repairs a leak; under-counting could let
        // compaction reclaim bytes a retained generation still needs).
        const blobHashes = this.storage.extractBlobHashesFromRecords
          ? this.storage.extractBlobHashesFromRecords([
              ...buf.nouns.values(),
              ...buf.verbs.values()
            ])
          : []
        if (blobHashes.length > 0 && this.storage.recordHistoryBlobReferences) {
          await this.storage.recordHistoryBlobReferences(blobHashes)
        }

        const nounIds: string[] = []
        const verbIds: string[] = []
        let recordBytes = 0
        for (const [id, record] of buf.nouns) {
          const path = `${dir}/prev/${id}.json`
          await this.storage.writeRawObject(path, record)
          recordBytes += serializedBytes(record)
          stagedPaths.push(path)
          nounIds.push(id)
        }
        for (const [id, record] of buf.verbs) {
          const path = `${dir}/prev/${id}.json`
          await this.storage.writeRawObject(path, record)
          recordBytes += serializedBytes(record)
          stagedPaths.push(path)
          verbIds.push(id)
        }
        const delta: GenerationDelta = {
          generation: gen,
          timestamp: buf.timestamp,
          groupCommit: true,
          nouns: nounIds,
          verbs: verbIds,
          // Always present on new deltas (empty = "no blobs"), so compaction
          // only falls back to record reads for pre-contract generations.
          blobHashes
        }
        delta.bytes = recordBytes + serializedBytes(delta)
        genBytes.set(gen, delta.bytes)
        const deltaPath = `${dir}/tx.json`
        await this.storage.writeRawObject(deltaPath, delta)
        stagedPaths.push(deltaPath)
        logEntries.push({ generation: gen, timestamp: buf.timestamp })
      }

      // ONE fsync for the whole window — the durability-batching win.
      await this.storage.syncRawObjects(stagedPaths)

      // Fact log (dual-write): make the window's buffered facts durable in the
      // same batch, BEFORE the commit point below — so a crash can only leave
      // the log AHEAD of the counter (open truncates), never a committed
      // generation without its durable fact.
      await this.factLog?.sync()

      // Test-only crash simulation: a throwing injector here leaves the staged
      // group-commit generation dirs on disk with NO manifest advance — the
      // exact "crashed mid-flush" state recovery must DROP-WITHOUT-RESTORE
      // (the pending in-memory state would be lost in a real crash; we abandon
      // this store and recover from disk on the next open()).
      if (this.commitFaultInjector) this.commitFaultInjector('before-manifest-rename')

      // Commit point: persist the counter, then atomically rename the manifest
      // to the highest pending generation — that commits ALL of them at once.
      const highest = gens[gens.length - 1]
      const highestTs = this.pendingBuffer.get(highest)?.timestamp ?? Date.now()
      await this.persistCounterUnlocked()
      const manifest: GenerationManifest = {
        version: 1,
        generation: highest,
        committedAt: new Date(highestTs).toISOString(),
        horizon: this.horizonGen
      }
      await this.storage.writeRawObject(MANIFEST_PATH, manifest)
      await this.storage.syncRawObjects([MANIFEST_PATH])

      // Move pending → committed. Chains already carry these gens; seed the
      // bounded delta cache from the buffer so the next read avoids a re-read,
      // then drop the in-memory before-images.
      this.committed = highest
      for (const gen of gens) {
        const buf = this.pendingBuffer.get(gen)
        if (!buf) continue
        this.appendCommittedGen(gen)
        this.setDelta(gen, {
          nouns: new Set(buf.nouns.keys()),
          verbs: new Set(buf.verbs.keys()),
          timestamp: buf.timestamp,
          bytes: genBytes.get(gen) ?? 0
        })
        if (this.historyBytesTotal !== null) {
          this.historyBytesTotal += genBytes.get(gen) ?? 0
        }
        this.pendingBuffer.delete(gen)
      }
      this.pendingGens = []

      for (const entry of logEntries) {
        await this.storage.appendTxLogLine(JSON.stringify(entry))
      }
    })
  }

  /**
   * @description Reset the durability-failure state after a successful flush.
   * If a failure was latched (writes were being refused), log the recovery so
   * the transition out of the loud state is as visible as the transition in.
   */
  private onPendingFlushSuccess(): void {
    if (this.pendingFlushError) {
      prodLog.info(
        `[GenerationStore] pending single-op history flush recovered after ` +
          `${this.pendingFlushFailures} failed attempt(s); writes resume.`
      )
    }
    this.pendingFlushFailures = 0
    this.pendingFlushError = null
  }

  /**
   * @description Account for a failed pending-tier flush (called from
   * {@link flushPendingSingleOps}' catch, so it fires on EVERY failure path —
   * background scheduler and explicit flush()/close() alike). Escalates a
   * transient blip (a warn, keep retrying) into a latched durability failure
   * once {@link PENDING_FLUSH_FAILURE_THRESHOLD} consecutive attempts fail —
   * from that point writes are refused with a {@link PendingFlushDurabilityError}
   * until a flush finally succeeds. Always ensures a backoff retry is scheduled
   * so a recovering disk self-heals without operator action, regardless of who
   * triggered the failing flush. The pending before-images are retained (the
   * failed flush never cleared them), so no history is lost — only not-yet-durable.
   */
  private recordPendingFlushFailure(err: Error): void {
    this.pendingFlushFailures++
    const attempts = this.pendingFlushFailures
    if (attempts >= GenerationStore.PENDING_FLUSH_FAILURE_THRESHOLD) {
      // Latch: refuse further writes rather than pile un-durable history.
      this.pendingFlushError = err
      prodLog.error(
        `[GenerationStore] pending single-op history flush has FAILED ${attempts} ` +
          `consecutive times (${err.message}). Generation history is not durable; ` +
          `writes are now REFUSED until it drains. Live canonical data is intact. ` +
          `Retrying with backoff; the store resumes writes when a flush succeeds.`
      )
    } else {
      prodLog.warn(
        `[GenerationStore] pending single-op flush failed (attempt ${attempts}/` +
          `${GenerationStore.PENDING_FLUSH_FAILURE_THRESHOLD}): ${err.message}; retrying with backoff.`
      )
    }
    this.scheduleFlushRetry(attempts)
  }

  /**
   * @description (Re)arm the pending-flush timer at a capped exponential
   * backoff after a failure, so a persistent fault is retried at a decaying
   * cadence instead of hot-looping. Shares the single {@link pendingFlushTimer}
   * slot with {@link schedulePendingFlush} (only one flush is ever scheduled).
   * The retry's own failure is re-accounted inside {@link flushPendingSingleOps}.
   */
  private scheduleFlushRetry(attempt: number): void {
    if (this.pendingFlushTimer !== null) return
    const backoff = Math.min(
      GenerationStore.PENDING_FLUSH_DELAY_MS * 2 ** Math.min(attempt, 6),
      GenerationStore.PENDING_FLUSH_RETRY_CAP_MS
    )
    this.pendingFlushTimer = setTimeout(() => {
      this.pendingFlushTimer = null
      // flushPendingSingleOps records the failure + reschedules internally.
      void this.flushPendingSingleOps().catch(() => {})
    }, backoff)
    const t = this.pendingFlushTimer as unknown as { unref?: () => void }
    if (t && typeof t.unref === 'function') t.unref()
  }

  /**
   * @description Throw if the store has a latched history-durability failure.
   * Called at the top of every write commit so a broken durability path fails
   * loud and immediately instead of silently accumulating un-durable history.
   */
  private assertHistoryDurable(): void {
    if (this.pendingFlushError) {
      throw new PendingFlushDurabilityError(this.pendingFlushError, this.pendingFlushFailures)
    }
  }

  /** Schedule a coalesced pending-tier flush (size trigger fires immediately on
   *  the next microtask; otherwise a {@link PENDING_FLUSH_DELAY_MS} timer). Both
   *  defer outside the current mutex section so the flush can re-acquire it. A
   *  failed flush is accounted + rescheduled inside {@link flushPendingSingleOps}
   *  (latch-on-persistent-failure), never swallowed as a bare log line. */
  private schedulePendingFlush(): void {
    if (this.pendingGens.length >= this.pendingFlushThreshold) {
      // Failure is recorded + retried inside flushPendingSingleOps.
      void Promise.resolve()
        .then(() => this.flushPendingSingleOps())
        .catch(() => {})
      return
    }
    if (this.pendingFlushTimer !== null) return
    this.pendingFlushTimer = setTimeout(() => {
      this.pendingFlushTimer = null
      void this.flushPendingSingleOps().catch(() => {})
    }, GenerationStore.PENDING_FLUSH_DELAY_MS)
    // A background flush must not keep the process alive (Node only).
    const t = this.pendingFlushTimer as unknown as { unref?: () => void }
    if (t && typeof t.unref === 'function') t.unref()
  }

  /** Cancel any scheduled pending-tier flush timer. */
  private clearPendingFlushTimer(): void {
    if (this.pendingFlushTimer !== null) {
      clearTimeout(this.pendingFlushTimer)
      this.pendingFlushTimer = null
    }
  }

  // ==========================================================================
  // Committed-generation interval set (run-length representation)
  // ==========================================================================

  /**
   * @description Record a newly committed generation. `gen` is ALWAYS greater
   * than every committed generation (the monotonic `counter++`), so it either
   * extends the last interval (the normal contiguous append, `gen ===
   * lastEnd + 1`) or opens a fresh single-element interval (the first
   * generation, or the first commit after a {@link compact} gap).
   * @param gen - The just-committed generation.
   */
  private appendCommittedGen(gen: number): void {
    const ranges = this.committedRanges
    const last = ranges[ranges.length - 1]
    if (last !== undefined && gen === last[1] + 1) {
      last[1] = gen
    } else {
      ranges.push([gen, gen])
    }
  }

  /** @returns The newest committed generation, or `undefined` when none exist. */
  private lastCommittedGen(): number | undefined {
    const ranges = this.committedRanges
    return ranges.length ? ranges[ranges.length - 1][1] : undefined
  }

  /** @returns How many committed generations the interval set represents. */
  private committedCount(): number {
    let total = 0
    for (const [start, end] of this.committedRanges) total += end - start + 1
    return total
  }

  /**
   * @description Yield every committed generation ascending — the exact multiset
   * (and order) the old flat `committedGens` array held.
   */
  private *committedGensAsc(): IterableIterator<number> {
    for (const [start, end] of this.committedRanges) {
      for (let g = start; g <= end; g++) yield g
    }
  }

  /**
   * @description Yield committed ∪ pending generations, ascending. Pending
   * generations are always greater than every committed one (reserved after the
   * last commit; `transact()`/`compact()` flush the pending tier first), so the
   * committed-then-pending concatenation is already sorted — identical to the old
   * `[...committedGens, ...pendingGens]`. This is the union historical reads
   * resolve over so un-flushed single-ops are visible to pins/`asOf`.
   */
  private *reservedGensAsc(): IterableIterator<number> {
    yield* this.committedGensAsc()
    for (const g of this.pendingGens) yield g
  }

  /**
   * @description Remove every committed generation `<= maxGen`. {@link compact}
   * always reclaims an oldest CONTIGUOUS prefix of the committed gens, so this
   * walk over the front of {@link committedRanges} is exact: drop whole ranges
   * that end at or below `maxGen`, then clip the first surviving range up to
   * `maxGen + 1` if `maxGen` falls inside it, and stop.
   * @param maxGen - Inclusive upper bound of the reclaimed prefix.
   */
  private removeCommittedUpTo(maxGen: number): void {
    const ranges = this.committedRanges
    let drop = 0 // count of leading ranges fully reclaimed
    while (drop < ranges.length) {
      const range = ranges[drop]
      if (range[1] <= maxGen) {
        drop++ // whole range is at or below the cut → reclaim it
      } else {
        if (range[0] <= maxGen) range[0] = maxGen + 1 // clip the straddling range
        break // this range survives; nothing newer is below the cut either
      }
    }
    if (drop > 0) ranges.splice(0, drop)
  }

  /**
   * @description The largest reserved (committed ∪ pending) generation
   * `<= gen` — exactly what `largestAtOrBefore([...committedGens, ...pendingGens],
   * gen)` returned, in O(log ranges + scanned pending) instead of materializing
   * the union. Pending generations are the newest reserved gens (all greater than
   * every committed one), so the largest pending `<= gen`, when one exists,
   * dominates every committed gen and is the answer; otherwise the answer is the
   * largest committed gen `<= gen`, found by binary-searching {@link committedRanges}
   * for the rightmost range whose `start <= gen` and clamping `gen` into it.
   * @param gen - The upper bound (inclusive).
   * @returns The largest reserved generation `<= gen`, or `undefined` when every
   *   reserved generation is greater than `gen`.
   */
  private largestReservedAtOrBefore(gen: number): number | undefined {
    // Pending (ascending, all > every committed): the largest one ≤ gen wins.
    let bestPending: number | undefined
    for (const p of this.pendingGens) {
      if (p <= gen) bestPending = p
      else break
    }
    if (bestPending !== undefined) return bestPending
    // No pending ≤ gen → the largest committed gen ≤ gen. Binary-search for the
    // rightmost range whose start ≤ gen; the answer is min(gen, that range.end).
    const ranges = this.committedRanges
    let lo = 0
    let hi = ranges.length // first range index whose start is > gen
    while (lo < hi) {
      const mid = (lo + hi) >>> 1
      if (ranges[mid][0] <= gen) lo = mid + 1
      else hi = mid
    }
    if (lo === 0) return undefined // every committed range starts above gen
    const [, end] = ranges[lo - 1]
    return Math.min(gen, end)
  }

  // ==========================================================================
  // Point-in-time resolution
  // ==========================================================================

  /**
   * @description Whether any transaction committed after generation `gen`.
   * O(1) — this is the fast-path check that lets a `Db` pinned at the
   * current generation delegate every read to the live brain untouched.
   * @param gen - The pinned generation.
   */
  hasCommittedAfter(gen: number): boolean {
    // Pending single-op generations count: a now() pin must report itself
    // historical the moment any later write (committed OR un-flushed) lands,
    // else the Model-A "pin doesn't freeze against single-ops" hole reopens.
    const last =
      this.pendingGens.length > 0
        ? this.pendingGens[this.pendingGens.length - 1]
        : this.lastCommittedGen()
    return last !== undefined && last > gen
  }

  /**
   * @description Resolve the state of an entity or relationship at pinned
   * generation `gen`: the before-image stored by the first committed
   * generation after `gen` that touched the id — or `{ source: 'current' }`
   * when nothing after `gen` touched it (the live storage state *is* the
   * state at `gen`).
   *
   * @param kind - `'noun'` for entities, `'verb'` for relationships.
   * @param id - The id to resolve.
   * @param gen - The pinned generation.
   * @returns `{ source: 'current' }`, `{ source: 'absent' }` (did not exist
   *   at `gen`), or `{ source: 'record', metadata, vector }` with the raw
   *   stored objects as of `gen`.
   */
  async resolveAt(
    kind: 'noun' | 'verb',
    id: string,
    gen: number
  ): Promise<
    | { source: 'current' }
    | { source: 'absent' }
    | { source: 'record'; metadata: any; vector: any | null }
  > {
    await this.ensureWindow()
    // Snapshot windowLo, then read the chain synchronously (no await between):
    // a concurrent commit can only slide the window at the `await` above, so the
    // routing decision and the chain it reads are a consistent pair.
    const windowLo = this.windowLo
    let candidate: number | undefined
    if (gen >= windowLo) {
      // HOT path — every generation that could hold the before-image
      // (`> gen ≥ windowLo`) is inside the resident window, so the recent chain
      // is complete for this pin. O(log chain).
      const chain = (kind === 'noun' ? this.recentNounChains : this.recentVerbChains).get(id)
      if (chain === undefined) return { source: 'current' }
      candidate = firstGenerationAfter(chain, gen)
    } else {
      // COLD path — a deep pin below the window. Serve from the bounded LRU,
      // reconstructing (lock-light) on a miss. Caching empty chains too keeps
      // repeated deep reads of untouched-in-cold-region ids O(1).
      const cold = kind === 'noun' ? this.coldNounChains : this.coldVerbChains
      let chain = cold.get(id)
      if (chain !== undefined) {
        cold.delete(id) // re-insert to mark most-recently-used
        cold.set(id, chain)
      } else {
        chain = await this.reconstructColdChain(kind, id)
        cold.set(id, chain)
        while (cold.size > this.coldChainLruMax) {
          const oldest = cold.keys().next().value
          if (oldest === undefined) break
          cold.delete(oldest)
        }
      }
      candidate = firstGenerationAfter(chain, gen)
    }
    if (candidate === undefined) {
      // Nothing after `gen` touched `id`; the live state is its state at `gen`.
      return { source: 'current' }
    }
    const record = await this.readBeforeImage(kind, candidate, id)
    if (record === null) {
      // The record-set exists (candidate is committed) but the id's
      // before-image is missing — only possible through external tampering.
      throw new Error(
        `Generation record missing: ${GENERATIONS_PREFIX}/${candidate}/prev/${id}.json ` +
          `(store corrupted or records removed outside compactHistory())`
      )
    }
    if (record.metadata === null && record.vector === null) {
      return { source: 'absent' }
    }
    return { source: 'record', metadata: record.metadata, vector: record.vector }
  }

  /**
   * @description Resolve the FIRST reserved generation after `gen` that touched
   * each of `ids`, by kind — the bulk, MUTEX-FREE counterpart of
   * {@link resolveAt}'s chain lookup. A SINGLE ascending pass over the reserved
   * (committed ∪ pending) generations records `firstAfter[id] = g` for the first
   * `g > gen` whose delta touched `id`, stopping early once every id is resolved.
   *
   * Ids absent from the returned map were never touched after `gen` → their live
   * storage state IS their state at `gen` (the {@link resolveAt} `'current'`
   * answer). Ids present map to the generation whose before-image
   * ({@link readGenerationRecord}) holds their state as of `gen`.
   *
   * WHY MUTEX-FREE (load-bearing): {@link Brainy.materializeAtGeneration} runs its
   * final reconciliation inside {@link snapshotWith} (which HOLDS the commit
   * mutex). Resolving ids one-by-one through {@link resolveAt} there would (a)
   * re-enter the mutex via {@link ensureWindow} → DEADLOCK, and (b) regress an
   * O(R) bulk copy to O(N·R). This method takes no lock and makes one pass, so a
   * deep-pin materialize over N ids stays O(R·d̄ + |ids|) with O(R) `getDelta`
   * calls. It iterates SNAPSHOTS of the small committed-range list (O(gaps)) and
   * the pending list (O(pending)) — NOT a materialized O(R) gen array — so it is
   * safe against concurrent commits and bounded in memory.
   *
   * @param kind - `'noun'` for entities, `'verb'` for relationships.
   * @param ids - The ids to resolve.
   * @param gen - The pinned generation.
   * @returns `id → first reserved generation after `gen` that touched it` for the
   *   subset of `ids` touched after `gen`.
   */
  async resolveManyAt(
    kind: 'noun' | 'verb',
    ids: Set<string>,
    gen: number
  ): Promise<Map<string, number>> {
    const firstAfter = new Map<string, number>()
    if (ids.size === 0) return firstAfter
    // Snapshot the small range list + pending list (NOT the expanded gens) so the
    // pass is immune to a concurrent compaction splice / commit append and stays
    // O(gaps + pending) in memory.
    const ranges = this.committedRanges.map(([s, e]) => [s, e] as [number, number])
    const pending = [...this.pendingGens]
    const record = async (g: number): Promise<boolean> => {
      const delta = await this.getDelta(g)
      const touched = kind === 'noun' ? delta.nouns : delta.verbs
      for (const id of touched) {
        if (ids.has(id) && !firstAfter.has(id)) firstAfter.set(id, g)
      }
      return firstAfter.size === ids.size
    }
    for (const [s, e] of ranges) {
      if (e <= gen) continue
      for (let g = Math.max(s, gen + 1); g <= e; g++) {
        if (await record(g)) return firstAfter
      }
    }
    for (const g of pending) {
      if (g <= gen) continue
      if (await record(g)) return firstAfter
    }
    return firstAfter
  }

  /**
   * @description Read the before-image of `id` recorded by generation `gen` — the
   * public, MUTEX-FREE accessor {@link Brainy.materializeAtGeneration}'s `copyAt`
   * uses after {@link resolveManyAt} hands it the generation. Pairs with
   * `resolveManyAt` so the bulk historical copy never touches the commit mutex
   * (no deadlock inside {@link snapshotWith}, no per-id chain rebuild).
   * @param kind - `'noun'` for entities, `'verb'` for relationships.
   * @param gen - The generation whose before-image to read.
   * @param id - The id to read.
   * @returns The stored {@link GenerationRecord}, or `null` when the record-set
   *   has no before-image for `id`.
   */
  async readGenerationRecord(
    kind: 'noun' | 'verb',
    gen: number,
    id: string
  ): Promise<GenerationRecord | null> {
    return this.readBeforeImage(kind, gen, id)
  }

  /**
   * @description Read the before-image of `id` stored by generation `gen` from
   * the right tier: the in-memory pending buffer while the generation is
   * un-flushed, or `_generations/<gen>/prev/<id>.json` on disk once committed.
   * Returns `null` when the record-set has no before-image for `id`.
   */
  private async readBeforeImage(
    kind: 'noun' | 'verb',
    gen: number,
    id: string
  ): Promise<GenerationRecord | null> {
    const pending = this.pendingBuffer.get(gen)
    if (pending) {
      return (kind === 'noun' ? pending.nouns : pending.verbs).get(id) ?? null
    }
    return (await this.storage.readRawObject(
      `${GENERATIONS_PREFIX}/${gen}/prev/${id}.json`
    )) as GenerationRecord | null
  }

  /**
   * @description Build the resident hot-tail window — the per-id chains
   * ({@link recentNounChains}/{@link recentVerbChains}) and the inverse index
   * ({@link windowNounDeltas}/{@link windowVerbDeltas}) over the last
   * {@link recentWindowGenerations} reserved generations, `[windowLo, counter]`.
   * Built once, lazily, under the commit mutex (so no concurrent commit mutates
   * {@link committedRanges}/pending mid-build); thereafter maintained
   * incrementally by {@link extendChains} and invalidated by {@link compact} /
   * reopen. Concurrent first-callers share one build. O(W) `getDelta` reads the
   * first time (only the window, NOT all of history); O(1) afterwards. Deep pins
   * below `windowLo` are served by {@link reconstructColdChain} on demand.
   */
  private async ensureWindow(): Promise<void> {
    if (this.windowReady) return
    if (!this.windowBuilding) {
      this.windowBuilding = this.withMutex(async () => {
        if (this.windowReady) return
        this.recentNounChains.clear()
        this.recentVerbChains.clear()
        this.windowNounDeltas.clear()
        this.windowVerbDeltas.clear()
        // The window is the newest W reserved generations, clamped above the
        // compaction horizon (gens ≤ horizon are reclaimed — never resident).
        this.windowLo = Math.max(this.horizonGen + 1, this.counter - this.recentWindowGenerations + 1)
        // reservedGensAsc (committed ∪ pending) is ascending, so chains accrue in
        // order and un-flushed single-ops in-window are indexed too.
        for (const g of this.reservedGensAsc()) {
          if (g < this.windowLo) continue
          const delta = await this.getDelta(g)
          const nouns = [...delta.nouns]
          const verbs = [...delta.verbs]
          for (const id of nouns) appendToChain(this.recentNounChains, id, g)
          for (const id of verbs) appendToChain(this.recentVerbChains, id, g)
          this.windowNounDeltas.set(g, nouns)
          this.windowVerbDeltas.set(g, verbs)
        }
        this.windowReady = true
        this.windowBuilding = null
      })
    }
    await this.windowBuilding
  }

  /**
   * @description Reconstruct the FULL per-id chain for a deep pin (`gen <
   * windowLo`) — every generation that touched `id`, ascending. Built from the
   * persisted deltas BELOW the window plus the resident in-window tail, then
   * cached in the cold LRU by the caller.
   *
   * LOCK-LIGHT — deliberately does NOT hold the commit mutex (a deep read must
   * never stall writers). Correctness without the lock:
   * - The recent tail and `windowLo` are snapshotted SYNCHRONOUSLY up front, so
   *   even if the window slides during the (awaited) cold scan, the chain is the
   *   one consistent with the snapshot. New writes during the scan land ABOVE the
   *   snapshot and cannot change any deep pin's first-after answer; the cold
   *   entry is also dropped by {@link extendChains} the moment `id` is next
   *   touched, so it is never served stale.
   * - The cold scan iterates a SNAPSHOT of the small committed-range list
   *   (O(gaps) memory, immune to a concurrent compaction splice).
   * - A generation concurrently reclaimed by compaction surfaces as a missing
   *   delta. Such a gen is always `≤ minPinned` (compaction never reclaims above
   *   the lowest live pin) and therefore NEVER a live pin's answer (`first-after
   *   > pin ≥ minPinned`), so it is skipped, not raised. A missing delta ABOVE
   *   `minPinned` is genuine corruption and still throws.
   *
   * @param kind - `'noun'` for entities, `'verb'` for relationships.
   * @param id - The id whose chain to reconstruct.
   * @returns The id's full ascending generation chain (`[]` when never touched).
   */
  private async reconstructColdChain(kind: 'noun' | 'verb', id: string): Promise<number[]> {
    // Synchronous, consistent snapshot of the window state.
    const windowLo = this.windowLo
    const recentTail = (kind === 'noun' ? this.recentNounChains : this.recentVerbChains).get(id)
    const recentSnapshot = recentTail ? [...recentTail] : []
    const ranges = this.committedRanges.map(([s, e]) => [s, e] as [number, number])
    const pendingCold = this.pendingGens.filter((g) => g < windowLo)

    const cold: number[] = []
    const scan = async (g: number): Promise<void> => {
      let delta: { nouns: Set<string>; verbs: Set<string> }
      try {
        delta = await this.getDelta(g)
      } catch (err) {
        // Concurrent-compaction tolerance: a reclaimed gen (≤ minPinned, never a
        // live pin's answer) is skipped; a missing gen above the lowest pin is
        // real corruption and rethrows.
        if (g <= this.minPinnedGeneration()) return
        throw err
      }
      if ((kind === 'noun' ? delta.nouns : delta.verbs).has(id)) cold.push(g)
    }

    for (const [s, e] of ranges) {
      if (s >= windowLo) break // ascending → nothing newer is below the window
      const hi = Math.min(e, windowLo - 1)
      for (let g = s; g <= hi; g++) await scan(g)
    }
    // Cold pending gens (only when more than W writes are buffered, i.e. tests
    // with a small window) sit between the committed gens and the window.
    for (const g of pendingCold) await scan(g)

    // Cold gens (< windowLo) then the resident in-window tail (≥ windowLo) — both
    // ascending and disjoint, so the concatenation is the full ascending chain.
    return recentSnapshot.length ? cold.concat(recentSnapshot) : cold
  }

  /**
   * Record that generation `gen` (the newest) touched these ids and advance the
   * hot-tail window. Keeps the resident window current after a commit without a
   * rebuild. No-op until the window is built (the eventual build reads `gen` from
   * {@link reservedGensAsc}).
   *
   * Three steps, all in-memory and I/O-FREE:
   * 1. Append `gen` to each touched id's recent chain, and DROP each touched id
   *    from the cold LRU — a freshly-touched id's cached deep chain is now
   *    incomplete, so it must be reconstructed on the next deep read (else a
   *    deep pin whose first-after just became `gen` would wrongly read `current`).
   * 2. Record the window inverse index entry `gen → touched ids` (the slide-trim
   *    driver).
   * 3. Slide `windowLo` forward to `max(horizon+1, gen-W+1)`, front-trimming the
   *    chains of exactly the ids leaving the window using the inverse index — NO
   *    `getDelta`, so the write path never reads `tx.json` on a slide regardless
   *    of W vs the delta-cache size.
   */
  private extendChains(gen: number, nouns: Iterable<string>, verbs: Iterable<string>): void {
    if (!this.windowReady) return
    const nounArr = [...nouns]
    const verbArr = [...verbs]
    for (const id of nounArr) {
      appendToChain(this.recentNounChains, id, gen)
      this.coldNounChains.delete(id)
    }
    for (const id of verbArr) {
      appendToChain(this.recentVerbChains, id, gen)
      this.coldVerbChains.delete(id)
    }
    this.windowNounDeltas.set(gen, nounArr)
    this.windowVerbDeltas.set(gen, verbArr)

    const newLo = Math.max(this.horizonGen + 1, gen - this.recentWindowGenerations + 1)
    if (newLo <= this.windowLo) return
    // Collect the ids leaving the window from the inverse index (O(slid·d̄)), then
    // front-trim their chains to the new low watermark (each id once).
    const trimNouns = new Set<string>()
    const trimVerbs = new Set<string>()
    for (let g = this.windowLo; g < newLo; g++) {
      const n = this.windowNounDeltas.get(g)
      if (n) for (const id of n) trimNouns.add(id)
      this.windowNounDeltas.delete(g)
      const v = this.windowVerbDeltas.get(g)
      if (v) for (const id of v) trimVerbs.add(id)
      this.windowVerbDeltas.delete(g)
    }
    for (const id of trimNouns) dropChainFront(this.recentNounChains, id, newLo)
    for (const id of trimVerbs) dropChainFront(this.recentVerbChains, id, newLo)
    this.windowLo = newLo
  }

  /** Drop the resident window + cold LRU so the next read rebuilds from the
   *  surviving generations (called after compaction removes record-sets / on
   *  reopen). */
  private invalidateChains(): void {
    this.windowReady = false
    this.windowBuilding = null
    this.recentNounChains.clear()
    this.recentVerbChains.clear()
    this.windowNounDeltas.clear()
    this.windowVerbDeltas.clear()
    this.coldNounChains.clear()
    this.coldVerbChains.clear()
    this.windowLo = 0
  }

  /**
   * @description The ids touched by committed transactions in the interval
   * `(fromGen, toGen]`. Used by `db.since()` and by historical `find()` to
   * build its overlay of changed entities.
   * @param fromGen - Exclusive lower bound.
   * @param toGen - Inclusive upper bound.
   */
  async changedBetween(fromGen: number, toGen: number): Promise<ChangedIds> {
    const nouns = new Set<string>()
    const verbs = new Set<string>()
    for (const gen of this.reservedGensAsc()) {
      if (gen <= fromGen || gen > toGen) continue
      const delta = await this.getDelta(gen)
      for (const id of delta.nouns) nouns.add(id)
      for (const id of delta.verbs) verbs.add(id)
    }
    return {
      nouns: [...nouns].sort(),
      verbs: [...verbs].sort(),
      fromGeneration: fromGen,
      toGeneration: toGen
    }
  }

  /**
   * @description The committed generations in `(fromGen, toGen]` that touched a
   * single id, ascending. The per-id companion to {@link changedBetween} — used
   * by `history()` to walk one entity's change-points without scanning every id.
   * No all-ids index: it consults the same cached per-generation deltas.
   * @param kind - `'noun'` for entities, `'verb'` for relationships.
   * @param id - The id to track.
   * @param fromGen - Exclusive lower bound.
   * @param toGen - Inclusive upper bound.
   */
  async generationsTouching(
    kind: 'noun' | 'verb',
    id: string,
    fromGen: number,
    toGen: number
  ): Promise<number[]> {
    const result: number[] = []
    for (const gen of this.reservedGensAsc()) {
      if (gen <= fromGen || gen > toGen) continue
      const delta = await this.getDelta(gen)
      const touched = kind === 'noun' ? delta.nouns : delta.verbs
      if (touched.has(id)) result.push(gen)
    }
    return result
  }

  /**
   * @description Assert that generation `gen` is reachable (not compacted
   * away) and within the committed range, then return it. Used by `asOf()`.
   * Reachability boundary: reclaiming record-set `N` removes the
   * before-images that reads at generations BELOW `N` depend on, so
   * generations `< horizon` are unreachable while the horizon itself remains
   * servable from the record-sets above it.
   * @param gen - The requested generation.
   * @throws GenerationCompactedError when `gen` is below the horizon.
   * @throws RangeError when `gen` is negative or beyond the current counter.
   */
  assertReachable(gen: number): number {
    if (!Number.isInteger(gen) || gen < 0) {
      throw new RangeError(`asOf(): generation must be a non-negative integer (got ${gen})`)
    }
    if (gen > this.counter) {
      throw new RangeError(
        `asOf(): generation ${gen} is in the future (store is at generation ${this.counter})`
      )
    }
    if (gen < this.horizonGen) {
      throw new GenerationCompactedError(gen, this.horizonGen)
    }
    return gen
  }

  /**
   * @description Resolve a wall-clock timestamp to the generation that was
   * committed at-or-before it, via the tx-log. Under Model-B every write —
   * `transact()` AND single-op — has its own tx-log entry, so resolution is
   * per-write (the tx-log includes un-flushed single-op generations too, so
   * `asOf(Date)` is consistent with the rest of the temporal surface).
   * @param timestampMs - Milliseconds since epoch.
   * @returns The resolved generation (`0` when `timestampMs` predates the
   *   first commit) and the matched tx-log entry when one exists.
   */
  async resolveTimestamp(
    timestampMs: number
  ): Promise<{ generation: number; entry: TxLogEntry | null }> {
    let best: TxLogEntry | null = null
    for (const entry of await this.txLog()) {
      if (entry.timestamp <= timestampMs) {
        if (best === null || entry.generation > best.generation) best = entry
      }
    }
    return { generation: best?.generation ?? 0, entry: best }
  }

  /**
   * @description Read all committed tx-log entries, oldest first. Torn
   * trailing lines from a crashed append are tolerated and skipped, and
   * entries beyond the committed watermark are excluded — the tx-log is
   * advisory metadata; the manifest rename is the source of commit truth.
   * Compaction never rewrites the log, so entries may reference generations
   * whose record-sets were already reclaimed.
   * @returns Every committed {@link TxLogEntry}, in commit order.
   */
  async txLog(): Promise<TxLogEntry[]> {
    const lines = await this.storage.readTxLogLines()
    const entries: TxLogEntry[] = []
    for (const line of lines) {
      let entry: TxLogEntry
      try {
        entry = JSON.parse(line) as TxLogEntry
      } catch {
        continue // tolerate a torn trailing line from a crashed append
      }
      if (entry.generation <= this.committed) entries.push(entry)
    }
    // Include un-flushed single-op generations so the tx-log is consistent with
    // the rest of the temporal surface (`since`/`diff`/`asOf`/`changedBetween`
    // already resolve over the pending tier). Reads are thus flush-agnostic —
    // whether the async group-commit has run yet never changes results, only
    // crash durability. Pending single-ops are always newer than every
    // committed generation, so they extend the ascending list.
    for (const gen of this.pendingGens) {
      const buf = this.pendingBuffer.get(gen)
      if (buf) entries.push({ generation: gen, timestamp: buf.timestamp })
    }
    return entries
  }

  /**
   * @description The commit timestamp of the newest committed generation at
   * or below `gen`, when its record-set is still resident. Used to populate
   * `Db.timestamp` for `asOf()` values.
   * @param gen - The pinned generation.
   */
  async commitTimestampAtOrBefore(gen: number): Promise<number | null> {
    // The largest reserved (committed ∪ pending) generation ≤ gen, found in
    // O(log ranges) over the interval set (not an O(database-age) backward
    // scan). Pending single-op generations carry timestamps too.
    const candidate = this.largestReservedAtOrBefore(gen)
    if (candidate === undefined) return null
    const delta = await this.getDelta(candidate)
    return delta.timestamp
  }

  private async getDelta(
    gen: number
  ): Promise<{ nouns: Set<string>; verbs: Set<string>; timestamp: number; bytes: number }> {
    const cached = this.deltaCache.get(gen)
    if (cached) return cached
    // A pending (un-flushed) generation has no tx.json on disk yet — derive its
    // delta from the in-memory before-image buffer. Not cached: the bounded
    // cache is for the disk path; the buffer is the source of truth here.
    // bytes = 0: pending generations are not on disk, so they do not count
    // toward the on-disk history byte budget ({@link historyBytes}).
    const pending = this.pendingBuffer.get(gen)
    if (pending) {
      return {
        nouns: new Set(pending.nouns.keys()),
        verbs: new Set(pending.verbs.keys()),
        timestamp: pending.timestamp,
        bytes: 0
      }
    }
    const delta = (await this.storage.readRawObject(
      `${GENERATIONS_PREFIX}/${gen}/tx.json`
    )) as GenerationDelta | null
    if (delta === null) {
      throw new Error(
        `Generation delta missing: ${GENERATIONS_PREFIX}/${gen}/tx.json ` +
          `(store corrupted or records removed outside compactHistory())`
      )
    }
    const entry = {
      nouns: new Set(delta.nouns),
      verbs: new Set(delta.verbs),
      timestamp: delta.timestamp,
      bytes: delta.bytes ?? 0
    }
    this.setDelta(gen, entry)
    return entry
  }

  /**
   * @description Insert a delta into {@link deltaCache}, evicting the oldest
   * entries (FIFO, by insertion order) past {@link deltaCacheMax}. Eviction is
   * safe: {@link getDelta} re-reads any evicted delta from storage on demand.
   */
  private setDelta(
    gen: number,
    entry: { nouns: Set<string>; verbs: Set<string>; timestamp: number; bytes: number }
  ): void {
    this.deltaCache.set(gen, entry)
    while (this.deltaCache.size > this.deltaCacheMax) {
      const oldest = this.deltaCache.keys().next().value
      if (oldest === undefined) break
      this.deltaCache.delete(oldest)
    }
  }

  // ==========================================================================
  // Compaction
  // ==========================================================================

  /**
   * @description Total serialized bytes of the ON-DISK generational history —
   * the sum of every committed generation's recorded `bytes`. Backs the
   * `maxBytes` and adaptive retention caps. O(1) after the first call: the
   * total is computed by ONE walk over committed deltas, then maintained
   * incrementally at every commit (+bytes) and reclaim (−bytes) and dropped on
   * a wholesale state replacement (restore). Without the running total, the
   * adaptive auto-compaction on every flush() re-walked the ENTIRE history —
   * O(committed generations) file reads per flush past the delta-cache bound —
   * which is how a 70k-generation production brain turned every write into a
   * full-tail scan (SELF-GENERATIONS-GROWTH). Pending (un-flushed) generations
   * are excluded (they are not on disk).
   * @returns The total on-disk history byte count.
   */
  async historyBytes(): Promise<number> {
    if (this.historyBytesTotal !== null) {
      return this.historyBytesTotal
    }
    let total = 0
    for (const gen of this.committedGensAsc()) {
      total += (await this.getDelta(gen)).bytes
    }
    this.historyBytesTotal = total
    return total
  }

  /**
   * @description Reclaim generation record-sets per the retention CAPS. A
   * generation is removed only when it is at or below **every** live pin
   * (deleting `N` can only break readers pinned *below* `N`, because resolution
   * reads before-images from generations strictly greater than the pin) — pins
   * are ALWAYS exempt. Among the unpinned generations, the OLDEST are reclaimed
   * while **ANY** supplied cap is exceeded:
   *
   * - `maxGenerations` — reclaim while the committed-generation count exceeds it.
   * - `maxAge` — reclaim generations older than `now − maxAge`.
   * - `maxBytes` — reclaim while total on-disk history bytes exceed it.
   *
   * Because reclamation is oldest-first and every cap is monotone in age, once
   * the oldest unpinned generation violates no cap, no newer one does either, so
   * the scan stops. With no options, every unpinned generation is reclaimed.
   *
   * @param options - Retention caps (see {@link CompactHistoryOptions}).
   * @returns Count of removed record-sets and the new horizon.
   */
  async compact(options?: CompactHistoryOptions): Promise<CompactHistoryResult> {
    return this.withMutex(async () => {
      const minPinned = this.minPinnedGeneration()
      const maxGenerations = options?.maxGenerations
      const maxAge = options?.maxAge
      const maxBytes = options?.maxBytes
      const ageCutoff = maxAge !== undefined ? Date.now() - maxAge : undefined
      // Bounded maintenance pass (8.9.0): stop reclaiming once the budget is
      // spent. Safe mid-loop — reclamation is oldest-first, so an early stop
      // leaves a consistent contiguous prefix and the next pass resumes.
      const deadline =
        options?.timeBudgetMs !== undefined ? Date.now() + options.timeBudgetMs : undefined
      const noCaps =
        maxGenerations === undefined && maxAge === undefined && maxBytes === undefined

      // Running totals the caps are evaluated against; updated as we reclaim.
      let remainingCount = this.committedCount()
      let remainingBytes = maxBytes !== undefined ? await this.historyBytes() : 0

      // Snapshot the committed gens ascending so the reclaim loop iterates safely
      // while removeCommittedUpTo mutates the interval set afterwards.
      const removed: number[] = []
      for (const gen of [...this.committedGensAsc()]) {
        // Pins are always exempt: never reclaim a generation a live pin needs.
        if (gen > minPinned) break // committedGensAsc ascending → nothing newer is eligible either
        if (deadline !== undefined && Date.now() >= deadline) break // budget spent — resume next pass
        const delta = await this.getDelta(gen)
        if (!noCaps) {
          const violatesCount = maxGenerations !== undefined && remainingCount > maxGenerations
          const violatesBytes = maxBytes !== undefined && remainingBytes > maxBytes
          const violatesAge = ageCutoff !== undefined && delta.timestamp < ageCutoff
          // Oldest-first: once the oldest unpinned gen trips no cap, none newer do.
          if (!violatesCount && !violatesBytes && !violatesAge) break
        }
        const genBytes = maxBytes !== undefined ? delta.bytes : 0

        // Temporal-blob contract: resolve the content-blob hashes this
        // generation's record-set references BEFORE deleting it. New
        // generations carry the multiset in their persisted delta (empty
        // array when none — distinguishing "new format, no blobs" from a
        // pre-contract delta); legacy generations fall back to reading the
        // records themselves. Skipped entirely on non-blob-aware storage.
        let blobHashes: string[] | undefined
        if (this.storage.releaseHistoryBlobReferences) {
          const rawDelta = (await this.storage.readRawObject(
            `${GENERATIONS_PREFIX}/${gen}/tx.json`
          )) as GenerationDelta | null
          blobHashes = rawDelta?.blobHashes
          if (blobHashes === undefined && this.storage.extractBlobHashesFromRecords) {
            blobHashes = this.storage.extractBlobHashesFromRecords(
              await this.readGenerationRecords(gen)
            )
          }
        }

        await this.storage.removeRawPrefix(`${GENERATIONS_PREFIX}/${gen}`)
        this.deltaCache.delete(gen)
        if (this.historyBytesTotal !== null) {
          this.historyBytesTotal -= delta.bytes
        }

        // AFTER the record-set is gone (over-count-only crash ordering):
        // release its history references and reclaim any blob left with zero
        // live AND zero history references — the system's one byte-reclaim point.
        if (blobHashes && blobHashes.length > 0 && this.storage.releaseHistoryBlobReferences) {
          await this.storage.releaseHistoryBlobReferences(blobHashes)
        }

        remainingCount--
        remainingBytes -= genBytes
        removed.push(gen)
      }

      if (removed.length > 0) {
        // The reclaim loop walks committedGensAsc() oldest-first and stops at the
        // first survivor, so `removed` is the oldest CONTIGUOUS prefix of the
        // committed gens — every committed gen ≤ max(removed) was reclaimed, which
        // is exactly what removeCommittedUpTo strips. Guard that prefix invariant.
        const highestRemoved = Math.max(...removed)
        const countBefore = this.committedCount()
        this.removeCommittedUpTo(highestRemoved)
        if (this.committedCount() !== countBefore - removed.length) {
          throw new Error(
            `[GenerationStore] compaction invariant violated: reclaimed ${removed.length} ` +
              `generation(s) but committedCount dropped by ${countBefore - this.committedCount()} ` +
              `(the reclaimed set is not the oldest contiguous prefix)`
          )
        }
        // Reclaimed generations leave the per-id chains stale → rebuild on next read.
        this.invalidateChains()
        this.horizonGen = Math.max(this.horizonGen, highestRemoved)
        const manifest: GenerationManifest = {
          version: 1,
          generation: this.committed,
          committedAt: new Date().toISOString(),
          horizon: this.horizonGen
        }
        await this.storage.writeRawObject(MANIFEST_PATH, manifest)
        await this.storage.syncRawObjects([MANIFEST_PATH])
      }

      return { removedGenerations: removed.length, horizon: this.horizonGen }
    })
  }

  // ==========================================================================
  // Restore support
  // ==========================================================================

  /**
   * @description Re-read all persisted state after `brain.restore()`
   * replaced the store's contents from a snapshot, enforcing counter
   * monotonicity: the counter never moves below `floorGeneration` (the
   * pre-restore counter), so generation numbers observed before a restore
   * are never reissued.
   * @param floorGeneration - The counter value before the restore.
   */
  async reopenAfterRestore(floorGeneration: number): Promise<void> {
    await this.withMutex(async () => {
      this.deltaCache.clear()
      // The running history-byte total describes the REPLACED store — drop it;
      // the next historyBytes() re-seeds with one walk over the new state.
      this.historyBytesTotal = null
      // A wholesale state replacement invalidates any buffered single-op
      // history — discard the pending tier (its live writes are gone with the
      // replaced store).
      this.clearPendingFlushTimer()
      this.pendingGens = []
      this.pendingBuffer.clear()
      this.opened = false
      // open() re-reads counter/manifest and re-registers the bump hook.
      await this.open()
      if (this.counter < floorGeneration) {
        this.counter = floorGeneration
        await this.persistCounterUnlocked()
      }
    })
  }

  // ==========================================================================
  // Internals
  // ==========================================================================

  /**
   * Recover one uncommitted (crashed) generation, branching on whether it is a
   * single-op group-commit generation (drop-without-restore) or a `transact()`
   * generation (restore before-images). The decision reads the generation's
   * `tx.json` `groupCommit` flag:
   *
   * - **`groupCommit: true` → DROP** the directory, never restore. The single-op
   *   write already mutated canonical storage and was acknowledged to the caller
   *   before the flush; restoring its before-images would silently revert an
   *   acknowledged user write (the data-corruption trap). Only the *history* of
   *   the crashed flush window is lost — live data is correct as-is.
   * - **absent/`false` (a transact generation) → RESTORE** then drop (the
   *   existing atomic-rollback path).
   * - **`tx.json` unreadable/missing → DROP** (safe for both): a transact
   *   fsyncs `tx.json` BEFORE it executes, so an unreadable delta means the
   *   execute never ran → live storage is unchanged → dropping is a no-op on
   *   live state, while a partial group-commit must be dropped anyway.
   */
  private async recoverUncommittedGeneration(gen: number): Promise<void> {
    const dir = `${GENERATIONS_PREFIX}/${gen}`
    const delta = (await this.storage.readRawObject(`${dir}/tx.json`)) as GenerationDelta | null
    if (delta === null || delta.groupCommit === true) {
      // Drop-without-restore (group-commit, or an indeterminate partial dir).
      await this.storage.removeRawPrefix(dir)
      prodLog.warn(
        `[GenerationStore] dropped uncommitted ${
          delta === null ? 'indeterminate' : 'group-commit'
        } generation ${gen} (live data left intact; history of that flush window discarded)`
      )
      return
    }
    await this.rollBackUncommittedGeneration(gen)
  }

  /**
   * Restore the before-images of an uncommitted `transact()` generation to the
   * canonical entity paths, then remove its record directory. Restores are
   * idempotent (writing identical bytes / deleting already-absent files), so
   * recovery is safe at every crash point of the transact commit protocol.
   */
  private async rollBackUncommittedGeneration(gen: number): Promise<void> {
    const dir = `${GENERATIONS_PREFIX}/${gen}`
    const prevPaths = await this.storage.listRawObjects(`${dir}/prev`)
    for (const recordPath of prevPaths) {
      const id = recordIdFromPath(recordPath)
      if (id === null) continue
      const record = (await this.storage.readRawObject(recordPath)) as GenerationRecord | null
      if (record === null) continue
      const image = { metadata: record.metadata, vector: record.vector }
      if (record.kind === 'verb') {
        await this.storage.writeVerbRaw(id, image)
      } else {
        await this.storage.writeNounRaw(id, image)
      }
    }
    await this.storage.removeRawPrefix(dir)
    prodLog.warn(
      `[GenerationStore] rolled back uncommitted generation ${gen} ` +
        `(${prevPaths.length} record(s) restored)`
    )
  }

  /**
   * @description Run a snapshot section serialized against transact commits,
   * compaction, and counter persistence (the store's single commit mutex),
   * with the generation counter durably persisted first. Used by
   * `db.persist()`: the lock guarantees a snapshot can never interleave with
   * a commit's record-write/manifest-rename sequence, and the up-front
   * counter persist guarantees the snapshot's `_system/generation.json`
   * reflects every single-operation bump (whose persistence is otherwise
   * coalesced and may still be scheduled).
   * @param fn - The exclusive snapshot section.
   * @returns `fn`'s result.
   */
  snapshotWith<R>(fn: () => Promise<R>): Promise<R> {
    return this.withMutex(async () => {
      await this.persistCounterUnlocked()
      return fn()
    })
  }

  /** Run `fn` exclusively, chained behind every prior exclusive section. */
  private withMutex<R>(fn: () => Promise<R>): Promise<R> {
    const run = this.mutexTail.then(fn, fn)
    // Keep the chain alive regardless of fn's outcome.
    this.mutexTail = run.catch(() => {})
    return run
  }
}

/**
 * @description Approximate serialized byte size of a record or delta, for the
 * Model-B retention byte caps (`maxBytes` / adaptive). The JSON string length
 * (chars ≈ bytes) is a deliberately cheap soft estimate — retention is a budget,
 * not exact accounting — and avoids a storage size API. Returns 0 on any
 * serialization failure (e.g. a cyclic structure, which records never are).
 * @param obj - The record or delta about to be persisted.
 */
function serializedBytes(obj: unknown): number {
  try {
    return JSON.stringify(obj)?.length ?? 0
  } catch {
    return 0
  }
}

/**
 * @description Extract the generation number from a record path of the form
 * `_generations/<N>/...`. Returns `null` for paths that don't match.
 * @param path - A storage-root-relative object path.
 */
function parseGenerationFromPath(path: string): number | null {
  const match = /^_generations[/\\](\d+)[/\\]/.exec(path)
  if (!match) return null
  const gen = Number(match[1])
  return Number.isSafeInteger(gen) ? gen : null
}

/**
 * @description Extract the record id (file basename without `.json`) from a
 * `prev/` before-image record path. Returns `null` for non-record paths.
 * @param path - A storage-root-relative object path.
 */
function recordIdFromPath(path: string): string | null {
  const match = /[/\\]prev[/\\]([^/\\]+)\.json$/.exec(path)
  return match ? match[1] : null
}

/**
 * @description Append `gen` to an id's ascending history chain in `chains`,
 * creating the chain on first touch. Callers append in ascending generation
 * order, so the chain stays sorted without a re-sort.
 */
function appendToChain(chains: Map<string, number[]>, id: string, gen: number): void {
  const chain = chains.get(id)
  if (chain === undefined) {
    chains.set(id, [gen])
  } else if (chain[chain.length - 1] !== gen) {
    chain.push(gen)
  }
}

/**
 * @description Drop the leading entries of an id's ascending chain that fall
 * below `lo` (generations that slid out of the hot-tail window), deleting the
 * chain entirely when it empties. Used by the O(d̄) window slide in
 * {@link GenerationStore.extendChains}. Tolerates an already-trimmed front.
 */
function dropChainFront(chains: Map<string, number[]>, id: string, lo: number): void {
  const chain = chains.get(id)
  if (chain === undefined) return
  let i = 0
  while (i < chain.length && chain[i] < lo) i++
  if (i === chain.length) chains.delete(id)
  else if (i > 0) chain.splice(0, i)
}

/**
 * @description Binary-search an ascending generation chain for the FIRST entry
 * strictly greater than `gen` (the generation whose before-image holds the
 * value as of `gen`). Returns `undefined` when every entry is ≤ `gen`. O(log n).
 */
function firstGenerationAfter(chain: number[], gen: number): number | undefined {
  let lo = 0
  let hi = chain.length // upper-bound search over [lo, hi)
  while (lo < hi) {
    const mid = (lo + hi) >>> 1
    if (chain[mid] > gen) hi = mid
    else lo = mid + 1
  }
  return lo < chain.length ? chain[lo] : undefined
}