open-brainy/src/db/portableGraph.ts

/**
 * @module db/portableGraph
 * @description Portable graph export/import engine for Brainy 8.0 — the
 * `PortableGraph v1` document format plus the `export`/`import` engines that power
 * `db.export()` / `brain.export()` (read, at a pinned generation) and
 * `brain.import(graph)` (write, applied as one atomic transaction).
 *
 * A `PortableGraph` is a self-describing, versioned, **portable** representation
 * of a graph (entities + relations, optionally vectors and file blobs) — the unit
 * Brainy exports for interchange between instances, versions, and products, and the
 * payload other artifacts embed (e.g. a workbench file's `graph` field). It is NOT
 * a backup: that role belongs to `db.persist()`/`Brainy.load()`, the native
 * whole-brain snapshot which preserves generation history. `PortableGraph` is the
 * human-readable, partial-or-whole, cross-version round-trip; distinct also from
 * `brain.import(file)` (foreign-file ingestion).
 *
 * Every document carries `format: 'brainy-portable-graph'` and `formatVersion: 1`;
 * import gates on `formatVersion` for forward migration. The format is shared
 * verbatim across the 7.x and 8.0 lines.
 *
 * **Reserved-field split:** each `PortableGraphEntity` carries Brainy's standard fields
 * (`subtype`, `visibility`, `data`, `confidence`, `weight`, `service`, `createdBy`,
 * `createdAt`) at the TOP LEVEL and `metadata` holds ONLY custom user fields —
 * mirroring the in-memory `Entity`, so import maps each to its dedicated `add`/
 * `relate` parameter rather than the metadata bag.
 */

import { Entity, Relation, Result } from '../types/brainy.types.js'
import { NounType, VerbType } from '../types/graphTypes.js'
import { StorageAdapter, HNSWVerbWithMetadata } from '../coreTypes.js'
import { getBrainyVersion } from '../utils/version.js'
import { TxOperation } from './types.js'

/** Fixed entity id of the VFS root collection (a `system` entity; excluded unless `includeSystem`). */
const VFS_ROOT_ID = '00000000-0000-0000-0000-000000000000'
/** Magic string identifying a Brainy `PortableGraph` document (the `format` tag). */
export const PORTABLE_GRAPH_FORMAT = 'brainy-portable-graph'
/** Current portable-format version (import gates on this for cross-version migration). */
export const PORTABLE_GRAPH_FORMAT_VERSION = 1
/** Default embedding model label (informational; `dimensions` is the real compat gate). */
const DEFAULT_EMBED_MODEL = 'all-MiniLM-L6-v2'
/** Page size for find-based enumeration (whole-brain / predicate selectors). */
const ENUM_PAGE = 1000
/** Per-node relation fetch ceiling (source/target-filtered, so no unfiltered-scan warning). */
const RELATION_FETCH_LIMIT = 100_000

// ============================================================================
// Public types — identical wire shape to the 7.x line
// ============================================================================

/**
 * @description Selects WHICH part of the graph to export. Reuses `find()`'s grammar
 * plus export-specific selectors. Omit (or pass `{}`) to export the whole brain.
 * Structural selectors (`ids`/`collection`/`connected`/`vfsPath`) and predicate
 * selectors (`type`/`subtype`/`where`/`service`/`visibility`) **compose**.
 */
export interface ExportSelector {
  /** Exactly these entity ids. */
  ids?: string[]
  /** A collection id → the collection + its transitive `Contains` members. */
  collection?: string
  /** Alias for `collection`. */
  memberOf?: string
  /** An entity + its N-hop neighbourhood. */
  connected?: {
    from: string
    depth?: number
    verbs?: VerbType[]
    direction?: 'out' | 'in' | 'both'
  }
  /** A VFS path → the directory/file + (for a directory) its `Contains` subtree. */
  vfsPath?: string
  /** For `vfsPath` directories: include the whole subtree (default: true). */
  recursive?: boolean
  /** For `collection`/`vfsPath`: cap traversal depth (default: unbounded). */
  depth?: number
  /** Predicate: entity type(s). */
  type?: NounType | NounType[]
  /** Predicate: entity subtype(s). */
  subtype?: string | string[]
  /** Predicate: exact-match metadata fields. */
  where?: Record<string, any>
  /** Predicate: multi-tenancy service id. */
  service?: string
  /** Predicate: visibility (`'public'` matches entities with no explicit visibility). */
  visibility?: string
}

/** Controls HOW the selected graph is serialized. */
export interface ExportOptions {
  /** Include embedding vectors verbatim (default: false → `import()` re-embeds from `data`). */
  includeVectors?: boolean
  /** Include VFS file bytes in `blobs` so files round-trip byte-identically (default: false). */
  includeContent?: boolean
  /** Include `visibility:'system'` entities (e.g. the VFS root) (default: false). */
  includeSystem?: boolean
  /**
   * Admit BOTH hidden visibility tiers — `'internal'` AND `'system'` — into the
   * whole-brain/predicate candidate set, in EITHER `enumeration` mode (default:
   * false — today's behavior is byte-identical). `includeSystem` alone only ever
   * reached `'system'` for a structural selector's per-entity gate; whole-brain/
   * predicate enumeration never forwarded it into the candidate walk at all, so a
   * hidden-tier row could never survive a whole-brain export regardless of any
   * flag — the gap this option closes. `includeHidden: true` IMPLIES
   * `includeSystem: true` (both tiers are admitted together; there is no
   * "system but not internal" combination via this flag) — `includeSystem` on
   * its own keeps its narrower, pre-existing meaning for back-compat.
   *
   * Migration-grade exports set `includeHidden: true` — a complete-canon export
   * must carry every visibility tier; consumer-facing exports leave it off.
   */
  includeHidden?: boolean
  /** Which edges to include (default: `'induced'`). */
  edges?: 'induced' | 'incident' | 'none'
  /**
   * How the whole-brain / predicate selector (no `ids`/`collection`/`connected`/
   * `vfsPath`) resolves its candidate id set:
   *
   * - `'index'` (DEFAULT — unchanged behavior) — the generation-correct
   *   paginated `find()` walk. Fast (O(matches), not O(N)) but rides the
   *   metadata index as an acceleration structure: a lost/stale index posting
   *   can silently OMIT a canonical record, and a stale posting pointing at a
   *   record that no longer matches can silently produce a phantom (dropped
   *   later by the same predicate re-check `'canonical'` mode also runs, so
   *   phantoms never reach `entities` — but they ARE lost silently unless
   *   {@link reportIndexDrift} is set).
   * - `'canonical'` — walks every live noun/verb directly off the storage
   *   adapter's canonical shard layout (`storage.getNouns()`/`getVerbs()` —
   *   the same primitive `repairIndex()`'s recount and every index-heal walk
   *   use), then applies the selector as a plain predicate over the walked
   *   records. This GUARANTEES canon-completeness — the metadata/graph
   *   indexes are never consulted, so their corruption cannot hide a live
   *   record — at the cost of an O(N) walk regardless of selector
   *   selectivity (unlike the index path's O(matches)). Structural selectors
   *   (`ids`/`collection`/`connected`/`vfsPath`) resolve their node set
   *   exactly as in `'index'` mode either way (they never rode the metadata
   *   index); `'canonical'` additionally walks relations canonically for
   *   EVERY selector, since a lost adjacency-index posting can hide a
   *   relation regardless of how the node set was produced. Requires a
   *   storage adapter and the CURRENT generation — throws on a historical
   *   `asOf()` view or a speculative `with()` overlay (see
   *   {@link CanonicalEnumerationUnavailableError}), because the canonical
   *   walk has no notion of "as of a past generation."
   */
  enumeration?: 'index' | 'canonical'
  /**
   * Only meaningful with `enumeration:'canonical'` (ignored otherwise): ALSO
   * run the `'index'` enumeration in parallel and diff it against the
   * canonical ground truth, attaching the result as {@link PortableGraph.drift}.
   * Never auto-heals anything — this is migration-audit evidence, reported
   * loudly (`console.warn` with the counts) whenever either list is non-empty,
   * never silently. Default: false.
   */
  reportIndexDrift?: boolean
}

/**
 * @description Index-vs-canonical drift for one `export({ enumeration: 'canonical',
 * reportIndexDrift: true })` call. Only populated for the whole-brain / predicate
 * selector (structural selectors never consulted the metadata index for their node
 * set, so there is nothing to diff — both lists are empty for those).
 */
export interface ExportIndexDrift {
  /**
   * Ids the canonical storage walk confirmed (live, selector-matching) that the
   * index-based `find()` enumeration did NOT return — canonical records the
   * metadata index has lost track of.
   */
  canonicalOnly: string[]
  /**
   * Ids the index-based `find()` enumeration returned for this selector that
   * canonical ground truth (the storage walk + the same predicate check) does
   * NOT support — phantom index rows (stale or cross-bucket postings).
   */
  indexOnly: string[]
}

/** Controls how a `PortableGraph` is applied on `import()`. */
export interface ImportOptions {
  /** Conflict policy for an existing id (default: `'merge'` — dedup-by-id). */
  onConflict?: 'merge' | 'replace' | 'skip'
  /** Vector policy (default: `'auto'` — re-embed from `data` when no vector is carried). */
  reembed?: 'auto' | 'never'
  /** Rewrite every id on the way in (e.g. to clone a template subgraph under fresh ids). */
  remapIds?: (id: string) => string
  /** Transaction metadata, recorded in the tx-log alongside the new generation. */
  meta?: Record<string, unknown>
}

/** One entity in a `PortableGraph`. Standard fields top-level; `metadata` is custom-only. */
export interface PortableGraphEntity {
  id: string
  type: string
  subtype?: string
  visibility?: string
  data?: any
  confidence?: number
  weight?: number
  service?: string
  createdBy?: any
  createdAt?: number
  vector?: number[]
  metadata?: any
}

/** One relation (edge) in a `PortableGraph`. */
export interface PortableGraphRelation {
  id: string
  from: string
  to: string
  type: string
  subtype?: string
  visibility?: string
  weight?: number
  confidence?: number
  metadata?: any
}

/** A self-describing, versioned, portable graph document (identical on 7.x and 8.0). */
export interface PortableGraph {
  format: typeof PORTABLE_GRAPH_FORMAT
  formatVersion: number
  brainyVersion: string
  createdAt: string
  embedding: { model: string; dimensions: number }
  selector?: ExportSelector
  entities: PortableGraphEntity[]
  relations: PortableGraphRelation[]
  blobs?: Record<string, string>
  danglingIds?: string[]
  /** Present only when `export()` was called with `reportIndexDrift: true`. */
  drift?: ExportIndexDrift
  stats: { entityCount: number; relationCount: number; blobCount: number; vectorDimensions?: number }
}

/** Outcome of an `import()`. */
export interface ImportResult {
  imported: number
  merged: number
  skipped: number
  reembedded: number
  blobsWritten: number
  errors: Array<{ id: string; error: string }>
}

/** The minimal read surface the export engine needs — satisfied by `Db` and `Brainy`. */
export interface PortableGraphReader<T = any> {
  get(id: string, options?: { includeVectors?: boolean }): Promise<Entity<T> | null>
  find(query: any): Promise<Result<T>[]>
  related(paramsOrId?: any): Promise<Relation<T>[]>
}

/** The minimal write surface the import engine needs — satisfied by `Brainy`. */
export interface PortableGraphWriter<T = any> {
  get(id: string, options?: { includeVectors?: boolean }): Promise<Entity<T> | null>
  transact(ops: TxOperation<T>[], options?: { meta?: Record<string, unknown> }): Promise<unknown>
}

/** Type guard: is this value a Brainy `PortableGraph` document (vs a file/foreign object)? */
export function isPortableGraph(value: unknown): value is PortableGraph {
  return (
    typeof value === 'object' &&
    value !== null &&
    (value as any).format === PORTABLE_GRAPH_FORMAT &&
    Array.isArray((value as any).entities)
  )
}

/** Result of {@link validatePortableGraph}. `valid` is true iff `errors` is empty. */
export interface PortableGraphValidation {
  valid: boolean
  errors: string[]
  warnings: string[]
}

/**
 * @description Dry-run check a value before `import()`, without mutating the brain:
 * structural validity, a supported `formatVersion`, entity-id uniqueness + required
 * fields, and relation-endpoint coverage. `import()` throws on a non-`PortableGraph` or a
 * too-new `formatVersion`; `validatePortableGraph()` lets a consumer (e.g. a serializer checking
 * a `.wbench` graph payload) surface issues first.
 * @param data - The value to validate (need not be a `PortableGraph`).
 * @returns `{ valid, errors, warnings }`.
 * @example
 * const { valid, errors } = validatePortableGraph(payload)
 * if (!valid) throw new Error(`invalid PortableGraph: ${errors.join('; ')}`)
 * await brain.import(payload)
 */
export function validatePortableGraph(data: unknown): PortableGraphValidation {
  const errors: string[] = []
  const warnings: string[] = []

  if (!isPortableGraph(data)) {
    errors.push(`not a PortableGraph document (expected format: '${PORTABLE_GRAPH_FORMAT}')`)
    return { valid: false, errors, warnings }
  }
  if (typeof data.formatVersion !== 'number') {
    errors.push('formatVersion is missing or not a number')
  } else if (data.formatVersion > PORTABLE_GRAPH_FORMAT_VERSION) {
    errors.push(
      `formatVersion ${data.formatVersion} is newer than supported (max ${PORTABLE_GRAPH_FORMAT_VERSION})`
    )
  }

  const ids = new Set<string>()
  for (const e of data.entities) {
    if (!e || typeof e.id !== 'string' || e.id.length === 0) {
      errors.push('an entity is missing an id')
      continue
    }
    if (ids.has(e.id)) errors.push(`duplicate entity id: ${e.id}`)
    ids.add(e.id)
    if (e.type === undefined || e.type === null || (e.type as string) === '') {
      errors.push(`entity ${e.id} is missing a type`)
    }
  }

  const dangling = new Set(data.danglingIds ?? [])
  for (const r of data.relations) {
    if (!r || typeof r.id !== 'string') {
      errors.push('a relation is missing an id')
      continue
    }
    if (r.type === undefined || r.type === null || (r.type as string) === '') {
      errors.push(`relation ${r.id} is missing a type`)
    }
    if (!r.from || !r.to) {
      errors.push(`relation ${r.id} is missing from/to`)
      continue
    }
    if (!ids.has(r.from) && !dangling.has(r.from)) {
      warnings.push(`relation ${r.id}: from-endpoint ${r.from} is not in entities`)
    }
    if (!ids.has(r.to) && !dangling.has(r.to)) {
      warnings.push(`relation ${r.id}: to-endpoint ${r.to} is not in entities`)
    }
  }

  if (!data.embedding || typeof data.embedding.dimensions !== 'number') {
    warnings.push('embedding manifest is missing or has no dimensions')
  }

  return { valid: errors.length === 0, errors, warnings }
}

// ============================================================================
// EXPORT
// ============================================================================

/**
 * @description Serialize part or all of a graph (read through `reader` at its pinned
 * generation) into a portable `PortableGraph` document.
 *
 * `enumeration:'canonical'` (see {@link ExportOptions.enumeration}) requires a
 * storage adapter and the current generation: it throws
 * {@link CanonicalEnumerationUnavailableError} if `storage` is absent, and the
 * caller (`Db.export()`) throws the same error before this runs if the view is
 * historical or a speculative overlay — the canonical storage walk has no
 * generation parameter, so it can only ever answer "as of right now."
 *
 * @param reader - Generation-correct read surface (`Db` or `Brainy`).
 * @param storage - Storage adapter (VFS blob bytes when `includeContent`; the
 *   canonical noun/verb walk when `enumeration:'canonical'`).
 * @param selector - WHAT to export (omit for the whole brain).
 * @param options - HOW to export (vectors / file bytes / edge policy / enumeration mode).
 */
export async function exportGraph<T = any>(
  reader: PortableGraphReader<T>,
  storage: StorageAdapter | undefined,
  selector: ExportSelector,
  options: ExportOptions
): Promise<PortableGraph> {
  const {
    includeVectors = false,
    includeContent = false,
    includeSystem = false,
    includeHidden = false,
    edges = 'induced',
    enumeration = 'index',
    reportIndexDrift = false
  } = options

  if (enumeration === 'canonical' && !storage) {
    throw new Error(
      `export(): enumeration:'canonical' requires a storage adapter, but none was supplied ` +
        `to this reader. Use enumeration:'index' (the default), or export through a Db/Brainy ` +
        `that carries its storage adapter.`
    )
  }
  const wantDrift = enumeration === 'canonical' && reportIndexDrift
  // includeHidden IMPLIES includeSystem (see ExportOptions.includeHidden's JSDoc) — every
  // system-tier gate below reads THIS combined value, never the raw option, so
  // `includeHidden` alone is always sufficient to see system-tier rows too.
  const effectiveIncludeSystem = includeSystem || includeHidden

  // 1. Resolve the node-id set (+ the index's raw candidate set, only when diffing it).
  // Both `enumerateAllCanonical` and `enumerateAllIndexed` receive the SAME
  // `effectiveIncludeSystem`/`includeHidden` pair below, so a drift diff can never
  // contain tier-policy noise — only genuine index-vs-canonical disagreement.
  const { idSet, indexCandidateIds } = await resolveSelector(
    reader,
    storage,
    selector,
    effectiveIncludeSystem,
    enumeration,
    wantDrift,
    includeHidden
  )

  // 2. Read canonical entities (reserved fields top-level), applying any predicate filter.
  // Identical for both enumeration modes: 'canonical' only changes WHICH ids reach this
  // loop, never how a candidate is verified — so the two modes can disagree on candidacy,
  // never on what counts as a match.
  const usePredicate = hasPredicate(selector)
  const entityMap = new Map<string, Entity<T>>()
  const entities: PortableGraphEntity[] = []
  for (const id of idSet) {
    const e = await reader.get(id, { includeVectors })
    if (!e) continue
    if (!effectiveIncludeSystem && (e as any).visibility === 'system') continue
    if (usePredicate && !matchesPredicate(e, selector)) continue
    entityMap.set(id, e)
    entities.push(toPortableGraphEntity(e, includeVectors))
  }
  const keptIds = new Set(entityMap.keys())

  // 2b. Finalize the drift report now that ground truth (keptIds) is known.
  let drift: ExportIndexDrift | undefined
  if (wantDrift) {
    const indexIds = indexCandidateIds ?? new Set<string>()
    const canonicalOnly = [...keptIds].filter((id) => !indexIds.has(id))
    const indexOnly = [...indexIds].filter((id) => !keptIds.has(id))
    drift = { canonicalOnly, indexOnly }
    if (canonicalOnly.length > 0 || indexOnly.length > 0) {
      console.warn(
        `[Brainy] export() index drift: ${canonicalOnly.length} canonical-only id(s) ` +
          `(canon-present, the index-based enumeration missed them) and ${indexOnly.length} ` +
          `index-only id(s) (index-visible, canon-absent — phantom rows). ` +
          `See the returned PortableGraph's 'drift' field for the exact ids. Nothing was ` +
          `auto-healed — run brain.repairIndex() to reconcile the metadata index.`
      )
    }
  }

  // 3. Edges per policy. Canonical mode ALSO walks verbs canonically for every
  // selector (not just whole-brain) — a lost adjacency-index posting can hide a
  // relation regardless of how the node set was produced.
  const { relations, danglingIds } =
    enumeration === 'canonical'
      ? await collectEdgesCanonical(storage!, keptIds, edges, effectiveIncludeSystem, includeHidden)
      : await collectEdges(reader, keptIds, edges, includeHidden)

  // 4. VFS blob bytes (only when requested).
  let blobs: Record<string, string> | undefined
  if (includeContent) blobs = await collectBlobs(storage, entityMap)
  const blobCount = blobs ? Object.keys(blobs).length : 0

  // Embedding dimensionality for the manifest: detect from a carried vector,
  // else the default model's 384 (informational — `dimensions` is the compat gate).
  const dimensions = entities.find((e) => e.vector && e.vector.length)?.vector?.length ?? 384

  return {
    format: PORTABLE_GRAPH_FORMAT,
    formatVersion: PORTABLE_GRAPH_FORMAT_VERSION,
    brainyVersion: getBrainyVersion(),
    createdAt: new Date().toISOString(),
    embedding: { model: DEFAULT_EMBED_MODEL, dimensions },
    selector,
    entities,
    relations,
    ...(blobs && blobCount > 0 ? { blobs } : {}),
    ...(danglingIds && danglingIds.length > 0 ? { danglingIds } : {}),
    ...(drift ? { drift } : {}),
    stats: {
      entityCount: entities.length,
      relationCount: relations.length,
      blobCount,
      vectorDimensions: dimensions
    }
  }
}

// ============================================================================
// IMPORT
// ============================================================================

/**
 * @description Apply a `PortableGraph` to the brain as ONE atomic transaction (a single
 * new generation). Dedup-by-id merge by default; re-embeds from `data` when no vector
 * is carried. VFS blob bytes are written first (content-addressed, idempotent).
 * @param writer - The brain (`get` + `transact`).
 * @param storage - Storage adapter (for VFS blob bytes).
 * @param data - A `PortableGraph` document.
 * @param options - Conflict / vector / id-remap policy.
 * @throws If `data` is not a `PortableGraph`, or its `formatVersion` is newer than supported.
 */
export async function importGraph<T = any>(
  writer: PortableGraphWriter<T>,
  storage: StorageAdapter | undefined,
  data: PortableGraph,
  options: ImportOptions = {}
): Promise<ImportResult> {
  if (!isPortableGraph(data)) {
    throw new Error(
      `import() expects a PortableGraph document (format:'${PORTABLE_GRAPH_FORMAT}'). ` +
        `For file ingestion (CSV/PDF/Excel/JSON), pass a file/buffer instead.`
    )
  }
  if (typeof data.formatVersion === 'number' && data.formatVersion > PORTABLE_GRAPH_FORMAT_VERSION) {
    throw new Error(
      `PortableGraph formatVersion ${data.formatVersion} is newer than this Brainy supports ` +
        `(max ${PORTABLE_GRAPH_FORMAT_VERSION}). Upgrade Brainy to import this document.`
    )
  }

  const { onConflict = 'merge', reembed = 'auto', remapIds, meta } = options
  const result: ImportResult = {
    imported: 0,
    merged: 0,
    skipped: 0,
    reembedded: 0,
    blobsWritten: 0,
    errors: []
  }
  const mapId = (id: string) => (remapIds ? remapIds(id) : id)

  // 1. Blob bytes first (content-addressed → idempotent), so file entities resolve content.
  if (data.blobs && Object.keys(data.blobs).length > 0) {
    const blobStorage = (storage as any)?.blobStorage
    if (blobStorage?.write) {
      for (const [hash, b64] of Object.entries(data.blobs)) {
        try {
          await blobStorage.write(Buffer.from(b64, 'base64'))
          result.blobsWritten++
        } catch (e) {
          result.errors.push({ id: hash, error: `blob: ${(e as Error).message}` })
        }
      }
    } else {
      for (const hash of Object.keys(data.blobs)) {
        result.errors.push({ id: hash, error: 'blob: storage does not support binary blobs' })
      }
    }
  }

  // 2. Build the operation batch (one atomic transaction → one generation).
  const ops: TxOperation<T>[] = []
  for (const be of data.entities || []) {
    const id = mapId(be.id)
    try {
      const exists = (await writer.get(id)) !== null
      if (exists) {
        if (onConflict === 'skip') {
          result.skipped++
          continue
        }
        if (onConflict === 'merge') {
          ops.push({ op: 'update', id, ...entityUpdateFields(be), merge: true } as TxOperation<T>)
          result.merged++
          continue
        }
        ops.push({ op: 'remove', id } as TxOperation<T>) // 'replace'
      }
      const useVector = Array.isArray(be.vector) && be.vector.length > 0
      if (!useVector && reembed === 'never') {
        result.errors.push({ id, error: 'no vector carried and reembed:never' })
        continue
      }
      ops.push({
        op: 'add',
        id,
        ...entityAddFields(be),
        ...(useVector ? { vector: be.vector } : {})
      } as TxOperation<T>)
      result.imported++
      if (!useVector) result.reembedded++
    } catch (e) {
      result.errors.push({ id, error: (e as Error).message })
    }
  }

  for (const br of data.relations || []) {
    ops.push({
      op: 'relate',
      from: mapId(br.from),
      to: mapId(br.to),
      type: br.type as VerbType,
      ...(br.subtype !== undefined ? { subtype: br.subtype } : {}),
      ...(br.weight !== undefined ? { weight: br.weight } : {}),
      ...(br.confidence !== undefined ? { confidence: br.confidence } : {}),
      ...(br.metadata !== undefined ? { metadata: br.metadata } : {})
    } as TxOperation<T>)
  }

  // 3. Apply atomically — one generation for the whole graph, or none.
  if (ops.length > 0) {
    await writer.transact(ops, meta ? { meta } : undefined)
  }

  return result
}

// ============================================================================
// Selector resolution (private)
// ============================================================================

function hasStructural(s: ExportSelector): boolean {
  return !!(s.ids || s.collection || s.memberOf || s.connected || s.vfsPath)
}

function hasPredicate(s: ExportSelector): boolean {
  return (
    s.type !== undefined ||
    s.subtype !== undefined ||
    s.where !== undefined ||
    s.service !== undefined ||
    s.visibility !== undefined
  )
}

/**
 * @param reader - Generation-correct read surface.
 * @param storage - Storage adapter (only touched when `enumeration:'canonical'`
 *   resolves the whole-brain/predicate branch).
 * @param s - The export selector.
 * @param includeSystem - The ALREADY-COMBINED `includeSystem || includeHidden` value
 *   (see `exportGraph`'s `effectiveIncludeSystem`) — whether `visibility:'system'`
 *   entities are wanted.
 * @param enumeration - `'index'` (default) or `'canonical'` — see {@link ExportOptions.enumeration}.
 *   Only affects the whole-brain/predicate branch (the `else` below): structural
 *   selectors (`ids`/`collection`/`connected`/`vfsPath`) never rode the metadata
 *   index for their node set, so they resolve identically either way.
 * @param wantIndexCandidates - When true (only meaningful with `enumeration:'canonical'`
 *   on the whole-brain/predicate branch), ALSO run the index-based walk and return
 *   its raw candidate set as `indexCandidateIds`, for {@link ExportIndexDrift}.
 * @param includeHidden - Whether `visibility:'internal'` entities are ALSO wanted
 *   (see {@link ExportOptions.includeHidden}). Threaded to BOTH enumeration
 *   functions identically so a drift diff never contains tier-policy noise.
 */
async function resolveSelector<T>(
  reader: PortableGraphReader<T>,
  storage: StorageAdapter | undefined,
  s: ExportSelector,
  includeSystem: boolean,
  enumeration: 'index' | 'canonical',
  wantIndexCandidates: boolean,
  includeHidden: boolean
): Promise<{ idSet: Set<string>; indexCandidateIds?: Set<string> }> {
  let idSet: Set<string>
  let indexCandidateIds: Set<string> | undefined
  if (s.ids && s.ids.length) {
    idSet = new Set(s.ids)
  } else if (s.collection ?? s.memberOf) {
    idSet = await resolveCollectionSubtree(reader, (s.collection ?? s.memberOf)!, s.depth)
  } else if (s.connected) {
    idSet = await resolveConnected(reader, s.connected)
  } else if (s.vfsPath) {
    idSet = await resolveVfsPath(reader, s.vfsPath, s.recursive ?? true, s.depth)
  } else if (enumeration === 'canonical') {
    idSet = await enumerateAllCanonical(storage!, includeSystem, includeHidden)
    if (wantIndexCandidates) indexCandidateIds = await enumerateAllIndexed(reader, s, includeHidden)
  } else {
    idSet = await enumerateAllIndexed(reader, s, includeHidden)
  }
  if (!includeSystem) idSet.delete(VFS_ROOT_ID)
  return { idSet, indexCandidateIds }
}

/**
 * @description Whole-brain / predicate enumeration via generation-correct
 * paginated `find()`. The metadata index is an acceleration structure over
 * this candidate set — see {@link enumerateAllCanonical} for the storage-level
 * counterpart that never consults it.
 *
 * @param includeHidden - When true, forwards `includeInternal: true` AND
 *   `includeSystem: true` into the SAME `find()` call — `find()` supports both
 *   flags simultaneously (confirmed via `FindParams.includeInternal`/`includeSystem`
 *   and `Brainy`'s `resolveHiddenIds`/`excludedVisibilityTiers`), so ONE pass
 *   reaches both hidden tiers; no per-tier union pass is needed. When false
 *   (default), neither flag is forwarded — the pre-existing behavior, preserved
 *   byte-identically for back-compat (`ExportOptions.includeSystem` alone never
 *   reached this far; see {@link ExportOptions.includeHidden}'s JSDoc).
 */
async function enumerateAllIndexed<T>(
  reader: PortableGraphReader<T>,
  s: ExportSelector,
  includeHidden = false
): Promise<Set<string>> {
  const params: any = {}
  if (s.type !== undefined) params.type = s.type
  if (s.subtype !== undefined) params.subtype = s.subtype
  if (s.where !== undefined) params.where = s.where
  if (s.service !== undefined) params.service = s.service
  if (includeHidden) {
    params.includeInternal = true
    params.includeSystem = true
  }
  const ids = new Set<string>()
  let offset = 0
  // eslint-disable-next-line no-constant-condition
  while (true) {
    const batch = await reader.find({ ...params, limit: ENUM_PAGE, offset })
    for (const r of batch) ids.add(r.id)
    if (batch.length < ENUM_PAGE) break
    offset += ENUM_PAGE
  }
  return ids
}

/**
 * @description Canonical (storage-level) counterpart of {@link enumerateAllIndexed}:
 * walks every live noun directly off the storage adapter's canonical shard layout
 * (`storage.getNouns()` — the same primitive `repairIndex()`'s recount and every
 * index-heal walk use) instead of going through the metadata index. Guarantees
 * canon-completeness — a lost or stale metadata-index posting cannot cause a
 * canonical record to be silently missing from the returned set — at the cost of
 * an O(N) walk regardless of selector selectivity (unlike the index path's
 * O(matches)). Returns the RAW candidate id set; `exportGraph`'s caller applies
 * `matchesPredicate` per-entity via `reader.get()` afterward, exactly as the index
 * path does, so both paths share one predicate-evaluation code path and can only
 * disagree on candidacy, never on what a match means.
 *
 * Mirrors `find()`'s hidden-tier policy given the SAME `includeSystem`/`includeHidden`
 * pair (see {@link enumerateAllIndexed}) so the two enumeration modes produce
 * identical id sets when the index is healthy, at ANY tier-visibility setting.
 *
 * @param includeSystem - The ALREADY-COMBINED `includeSystem || includeHidden` value.
 * @param includeHidden - Whether `'internal'`-tier nouns are ALSO admitted.
 */
async function enumerateAllCanonical(
  storage: StorageAdapter,
  includeSystem = false,
  includeHidden = false
): Promise<Set<string>> {
  const ids = new Set<string>()
  let offset = 0
  let cursor: string | undefined
  // eslint-disable-next-line no-constant-condition
  while (true) {
    const page = await storage.getNouns({ pagination: { limit: ENUM_PAGE, offset, cursor } })
    for (const item of page.items) {
      if (item.visibility === 'internal' && !includeHidden) continue
      if (item.visibility === 'system' && !includeSystem) continue
      ids.add(item.id)
    }
    if (!page.hasMore || page.items.length === 0) break
    if (page.nextCursor !== undefined) {
      cursor = page.nextCursor
    } else {
      offset += ENUM_PAGE
    }
  }
  return ids
}

async function resolveCollectionSubtree<T>(
  reader: PortableGraphReader<T>,
  rootId: string,
  depth?: number
): Promise<Set<string>> {
  const set = new Set<string>([rootId])
  const maxDepth = depth ?? Infinity
  let frontier = [rootId]
  let d = 0
  while (frontier.length && d < maxDepth) {
    const next: string[] = []
    for (const id of frontier) {
      const rels = await reader.related({ from: id, type: VerbType.Contains, limit: RELATION_FETCH_LIMIT })
      for (const r of rels) {
        if (!set.has(r.to)) {
          set.add(r.to)
          next.push(r.to)
        }
      }
    }
    frontier = next
    d++
  }
  return set
}

async function resolveConnected<T>(
  reader: PortableGraphReader<T>,
  c: NonNullable<ExportSelector['connected']>
): Promise<Set<string>> {
  const { from, depth = 1, verbs, direction = 'out' } = c
  const set = new Set<string>([from])
  let frontier = [from]
  for (let d = 0; d < depth; d++) {
    const next: string[] = []
    for (const id of frontier) {
      const neighbours = await neighboursOf(reader, id, direction, verbs)
      for (const n of neighbours) {
        if (!set.has(n)) {
          set.add(n)
          next.push(n)
        }
      }
    }
    frontier = next
    if (!next.length) break
  }
  return set
}

async function neighboursOf<T>(
  reader: PortableGraphReader<T>,
  id: string,
  direction: 'out' | 'in' | 'both',
  verbs?: VerbType[]
): Promise<string[]> {
  const out: string[] = []
  if (direction === 'out' || direction === 'both') {
    const rels = await reader.related({ from: id, limit: RELATION_FETCH_LIMIT })
    for (const r of rels) if (!verbs || verbs.includes(r.type)) out.push(r.to)
  }
  if (direction === 'in' || direction === 'both') {
    const rels = await reader.related({ to: id, limit: RELATION_FETCH_LIMIT })
    for (const r of rels) if (!verbs || verbs.includes(r.type)) out.push(r.from)
  }
  return out
}

async function resolveVfsPath<T>(
  reader: PortableGraphReader<T>,
  path: string,
  recursive: boolean,
  depth?: number
): Promise<Set<string>> {
  const dirId = await resolveVfsPathToId(reader, path)
  if (!dirId) return new Set()
  if (!recursive) {
    const set = new Set<string>([dirId])
    const rels = await reader.related({ from: dirId, type: VerbType.Contains, limit: RELATION_FETCH_LIMIT })
    for (const r of rels) set.add(r.to)
    return set
  }
  return resolveCollectionSubtree(reader, dirId, depth)
}

async function resolveVfsPathToId<T>(reader: PortableGraphReader<T>, path: string): Promise<string | null> {
  const segments = path.split('/').filter(Boolean)
  let currentId = VFS_ROOT_ID
  for (const seg of segments) {
    const rels = await reader.related({ from: currentId, type: VerbType.Contains, limit: RELATION_FETCH_LIMIT })
    let found: string | null = null
    for (const r of rels) {
      const child = await reader.get(r.to)
      if ((child?.metadata as any)?.name === seg) {
        found = r.to
        break
      }
    }
    if (!found) return null
    currentId = found
  }
  return currentId
}

function matchesPredicate<T>(e: Entity<T>, s: ExportSelector): boolean {
  if (s.type !== undefined) {
    const types = Array.isArray(s.type) ? s.type : [s.type]
    if (!types.includes(e.type)) return false
  }
  if (s.subtype !== undefined) {
    const subs = Array.isArray(s.subtype) ? s.subtype : [s.subtype]
    if (e.subtype === undefined || !subs.includes(e.subtype)) return false
  }
  if (s.service !== undefined && e.service !== s.service) return false
  if (s.visibility !== undefined) {
    const vis = (e as any).visibility ?? 'public'
    if (vis !== s.visibility) return false
  }
  if (s.where && !matchesWhere(e.metadata, s.where)) return false
  return true
}

function matchesWhere(metadata: any, where: Record<string, any>): boolean {
  if (!metadata) return false
  for (const [key, value] of Object.entries(where)) {
    if (metadata[key] !== value) return false
  }
  return true
}

// ============================================================================
// Serialization helpers (private)
// ============================================================================

function toPortableGraphEntity<T>(e: Entity<T>, includeVectors: boolean): PortableGraphEntity {
  const be: PortableGraphEntity = { id: e.id, type: e.type as string }
  if (e.subtype !== undefined) be.subtype = e.subtype
  const vis = (e as any).visibility
  if (vis !== undefined && vis !== 'public') be.visibility = vis
  if (e.data !== undefined) be.data = e.data
  if (e.confidence !== undefined) be.confidence = e.confidence
  if (e.weight !== undefined) be.weight = e.weight
  if (e.service !== undefined) be.service = e.service
  if ((e as any).createdBy !== undefined) be.createdBy = (e as any).createdBy
  if (e.createdAt !== undefined) be.createdAt = e.createdAt
  if (includeVectors && e.vector && e.vector.length) be.vector = e.vector
  if (e.metadata && Object.keys(e.metadata as any).length) be.metadata = e.metadata
  return be
}

function toPortableGraphRelation<T>(r: Relation<T>): PortableGraphRelation {
  const br: PortableGraphRelation = { id: r.id, from: r.from, to: r.to, type: r.type as string }
  if (r.subtype !== undefined) br.subtype = r.subtype
  const vis = (r as any).visibility
  if (vis !== undefined && vis !== 'public') br.visibility = vis
  if (r.weight !== undefined) br.weight = r.weight
  if (r.confidence !== undefined) br.confidence = r.confidence
  if (r.metadata && Object.keys(r.metadata as any).length) br.metadata = r.metadata
  return br
}

/**
 * @param includeHidden - When true, forwards `includeInternal`/`includeSystem` into
 *   every `related()` call so hidden-tier relations reach candidacy too — mirrors
 *   {@link enumerateAllIndexed}'s `includeHidden` handling, and preserves back-compat
 *   when false/omitted (the pre-existing, unconditional hidden-tier exclusion).
 */
async function collectEdges<T>(
  reader: PortableGraphReader<T>,
  idSet: Set<string>,
  edges: 'induced' | 'incident' | 'none',
  includeHidden = false
): Promise<{ relations: PortableGraphRelation[]; danglingIds?: string[] }> {
  if (edges === 'none') return { relations: [] }

  const tierOptIn = includeHidden ? { includeInternal: true, includeSystem: true } : {}
  const relations: PortableGraphRelation[] = []
  const dangling = new Set<string>()
  const seen = new Set<string>()

  for (const id of idSet) {
    const rels = await reader.related({ from: id, limit: RELATION_FETCH_LIMIT, ...tierOptIn })
    for (const r of rels) {
      if (seen.has(r.id)) continue
      const toIn = idSet.has(r.to)
      if (edges === 'induced' && !toIn) continue
      if (!toIn) dangling.add(r.to)
      seen.add(r.id)
      relations.push(toPortableGraphRelation(r))
    }
  }

  if (edges === 'incident') {
    for (const id of idSet) {
      const rels = await reader.related({ to: id, limit: RELATION_FETCH_LIMIT, ...tierOptIn })
      for (const r of rels) {
        if (seen.has(r.id)) continue
        if (!idSet.has(r.from)) {
          dangling.add(r.from)
          seen.add(r.id)
          relations.push(toPortableGraphRelation(r))
        }
      }
    }
  }

  return dangling.size > 0 ? { relations, danglingIds: Array.from(dangling) } : { relations }
}

/** Converts a canonical verb record (as returned by `storage.getVerbs()`) into the wire shape. */
function hnswVerbToPortableGraphRelation(v: HNSWVerbWithMetadata): PortableGraphRelation {
  const br: PortableGraphRelation = { id: v.id, from: v.sourceId, to: v.targetId, type: v.verb as string }
  if (v.subtype !== undefined) br.subtype = v.subtype
  if (v.visibility !== undefined && v.visibility !== 'public') br.visibility = v.visibility
  if (v.weight !== undefined) br.weight = v.weight
  if (v.confidence !== undefined) br.confidence = v.confidence
  if (v.metadata && Object.keys(v.metadata as any).length) br.metadata = v.metadata
  return br
}

/**
 * @description Canonical (storage-level) counterpart of {@link collectEdges}:
 * walks every live verb directly off `storage.getVerbs()` — the same primitive
 * `repairIndex()`'s recount and every index-heal walk use — instead of the graph
 * adjacency index (`reader.related()`), so a lost/stale adjacency posting cannot
 * cause a canonical relationship to be silently dropped from the export. Used for
 * EVERY selector in `enumeration:'canonical'` mode, not just the whole-brain
 * branch: relations can be blinded by adjacency-index corruption regardless of
 * how `idSet` (the kept node ids) was produced.
 *
 * Mirrors `related()`'s default hidden-tier policy (hides `'internal'` and
 * `'system'` unless `includeHidden`/`includeSystem` say otherwise) so the two
 * enumeration modes produce identical relation sets when the index is healthy.
 *
 * @param includeSystem - Whether `'system'`-tier verbs are admitted (the
 *   caller passes the ALREADY-combined `includeSystem || includeHidden` value —
 *   see `exportGraph`'s `effectiveIncludeSystem`).
 * @param includeHidden - Whether `'internal'`-tier verbs are ALSO admitted.
 */
async function collectEdgesCanonical(
  storage: StorageAdapter,
  idSet: Set<string>,
  edges: 'induced' | 'incident' | 'none',
  includeSystem = false,
  includeHidden = false
): Promise<{ relations: PortableGraphRelation[]; danglingIds?: string[] }> {
  if (edges === 'none') return { relations: [] }

  const relations: PortableGraphRelation[] = []
  const dangling = new Set<string>()
  const seen = new Set<string>()
  let offset = 0
  let cursor: string | undefined

  // eslint-disable-next-line no-constant-condition
  while (true) {
    const page = await storage.getVerbs({ pagination: { limit: ENUM_PAGE, offset, cursor } })
    for (const v of page.items) {
      if (seen.has(v.id)) continue
      if (v.visibility === 'internal' && !includeHidden) continue
      if (v.visibility === 'system' && !includeSystem) continue
      const fromIn = idSet.has(v.sourceId)
      const toIn = idSet.has(v.targetId)
      if (edges === 'induced') {
        if (!fromIn || !toIn) continue
      } else if (!fromIn && !toIn) {
        continue // 'incident': neither endpoint kept — irrelevant to this export
      }
      if (fromIn && !toIn) dangling.add(v.targetId)
      if (toIn && !fromIn) dangling.add(v.sourceId)
      seen.add(v.id)
      relations.push(hnswVerbToPortableGraphRelation(v))
    }
    if (!page.hasMore || page.items.length === 0) break
    if (page.nextCursor !== undefined) {
      cursor = page.nextCursor
    } else {
      offset += ENUM_PAGE
    }
  }

  return dangling.size > 0 ? { relations, danglingIds: Array.from(dangling) } : { relations }
}

async function collectBlobs<T>(
  storage: StorageAdapter | undefined,
  entityMap: Map<string, Entity<T>>
): Promise<Record<string, string>> {
  const blobs: Record<string, string> = {}
  const blobStorage = (storage as any)?.blobStorage
  if (!blobStorage?.read) return blobs

  for (const e of entityMap.values()) {
    const storageMeta = (e.metadata as any)?.storage
    const hash = storageMeta?.hash
    if (storageMeta?.type === 'blob' && hash && !blobs[hash]) {
      try {
        const buf = await blobStorage.read(hash)
        blobs[hash] = Buffer.from(buf).toString('base64')
      } catch {
        // Referenced blob bytes unreadable (storage drift): skip — file structure still
        // travels, and stats.blobCount reflects what was actually captured.
      }
    }
  }
  return blobs
}

// ============================================================================
// add()/update() param mapping (private)
// ============================================================================

function entityAddFields(be: PortableGraphEntity): Record<string, any> {
  const fields: Record<string, any> = { data: be.data, type: be.type as NounType }
  if (be.subtype !== undefined) fields.subtype = be.subtype
  if (be.visibility !== undefined) fields.visibility = be.visibility
  if (be.service !== undefined) fields.service = be.service
  if (be.confidence !== undefined) fields.confidence = be.confidence
  if (be.weight !== undefined) fields.weight = be.weight
  if (be.metadata !== undefined) fields.metadata = be.metadata
  return fields
}

function entityUpdateFields(be: PortableGraphEntity): Record<string, any> {
  const fields: Record<string, any> = {}
  if (be.data !== undefined) fields.data = be.data
  if (be.type !== undefined) fields.type = be.type as NounType
  if (be.subtype !== undefined) fields.subtype = be.subtype
  if (be.visibility !== undefined) fields.visibility = be.visibility
  if (be.confidence !== undefined) fields.confidence = be.confidence
  if (be.weight !== undefined) fields.weight = be.weight
  if (be.metadata !== undefined) fields.metadata = be.metadata
  if (Array.isArray(be.vector) && be.vector.length) fields.vector = be.vector
  return fields
}