chore(release): 10.4.9

The delta gate caught the backgrounded fold breaking six pinned
crash-recovery cases: a reopened brain must have its markers re-armed
when open() returns, and a background latch races every consumer of that
contract. The backgrounding is reverted; the low-water mark stays — it
is the part that kills the whole-history scan, and with it the
foreground fold costs the log's tail on any brain that has ever drained.
The unmarked first open after upgrade pays one full scan, once, and the
open narrates it as its own step.

CHANGELOG.md

@@ -2,6 +2,29 @@
 
 All notable changes to this project will be documented in this file. See [standard-version](https://github.com/conventional-changelog/standard-version) for commit guidelines.
 
+### [10.4.9](https://source.soulcraft.com/soulcraftlabs/open-brainy/compare/v10.4.6...v10.4.9) (2026-09-02)
+
+- Merge branch 'fix/pending-embed-low-water' into rel/10.4.9-candidate (2648f56d)
+- fix(open): pending-embed recovery keeps the crash-recovery contract — foreground, bounded by the mark (8a2ebacf)
+- Merge branches 'fix/connected-find-order', 'fix/pending-embed-low-water' and 'fix/related-verb-array' into rel/10.4.9-candidate (d5147ed6)
+- fix(graph): the verb fast paths honour every requested type, source, and target (6a89adc4)
+- perf(open): pending-embed recovery is bounded by a low-water mark and runs behind the doors (88e79729)
+- fix(find): connected finds are graph-first — neighbours, then the filter over those ids, then the page (077cbc0b)
+- fix(storage): counts persistence is single-flight, coalesced, and never races its own temp file (5e3b343a)
+
+
+### [10.4.6](https://source.soulcraft.com/soulcraftlabs/open-brainy/compare/v10.4.5...v10.4.6) (2026-08-31)
+
+- fix(transact): metadata-index ops take their JSON-safe view at the crossing, not at construction (73500e7d)
+
+
+### [10.4.5](https://source.soulcraft.com/soulcraftlabs/open-brainy/compare/v10.4.4...v10.4.5) (2026-08-31)
+
+- build(release): the docs-push step retires — this engine documents itself in its own repository (d6bcb14f)
+- fix(generations): a sealed segment may only declare the generations it holds (a963a744)
+- fix(recovery): a torn generation-log tail is a terminal verdict, never a wait (c9930871)
+
+
 ### [10.4.4](https://source.soulcraft.com/soulcraftlabs/open-brainy/compare/v10.4.3...v10.4.4) (2026-08-28)
 
 - fix(vfs): the old-root sweep narrates only when it has something to say (d49148e1)

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@soulcraftlabs/brainy",
-  "version": "10.4.4",
+  "version": "10.4.9",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@soulcraftlabs/brainy",
-      "version": "10.4.4",
+      "version": "10.4.9",
       "license": "MIT",
       "dependencies": {
         "@msgpack/msgpack": "^3.1.2",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@soulcraftlabs/brainy",
-  "version": "10.4.4",
+  "version": "10.4.9",
   "brainyContract": 1,
   "description": "Universal Knowledge Protocol™ - World's first Triple Intelligence database unifying vector, graph, and document search in one API. Stage 3 CANONICAL: 42 nouns × 127 verbs covering 96-97% of all human knowledge.",
   "main": "dist/index.js",

releases/brainy.json

@@ -1,76 +0,0 @@
-{
-  "product": "brainy",
-  "entries": [
-    {
-      "version": "11.0.5",
-      "date": "2026-09-02",
-      "headline": "Graph-first finds in production, and opens that stop rescanning history",
-      "items": [
-        "find({ connected, where }) now walks the neighbours first and filters only those rows through a native door — correct at every page and O(neighbours), never the whole store.",
-        "related() with a list of verb types returns every requested kind (a fast path had silently kept only the first).",
-        "Deferred-embedding recovery resumes from a low-water mark instead of rescanning the whole generation log at every open — measured at two minutes on a large brain, now milliseconds."
-      ],
-      "url": null,
-      "thumb": null
-    },
-    {
-      "version": "11.0.4",
-      "date": "2026-09-01",
-      "headline": "Closes in milliseconds, index rebuilds without the disk-sync storm",
-      "items": [
-        "close() no longer pays deferred compaction or waits out an in-flight rebuild — measured 8 ms against the 4-minute closes it replaces; deferred work resumes at the next open, in the background.",
-        "The metadata index's rebuild syncs to disk per shard instead of per row, and the durability point moved to the publish step — the same guarantee, a fraction of the disk traffic.",
-        "A new native filter door evaluates queries over exactly the candidate rows a graph walk found, never the whole store."
-      ],
-      "url": null,
-      "thumb": null
-    },
-    {
-      "version": "11.0.3",
-      "date": "2026-09-01",
-      "headline": "The embedding upgrade ceremony runs on every brain",
-      "items": [
-        "A brain opened through the standard plugin now carries its embedding-model identity, so the full-precision upgrade ceremony can run on it.",
-        "A one-fix release; nothing else changed."
-      ],
-      "url": null,
-      "thumb": null
-    },
-    {
-      "version": "11.0.2",
-      "date": "2026-08-31",
-      "headline": "One embedding quality everywhere, 3–4× faster imports",
-      "items": [
-        "Every runtime embeds with the same full-precision model — search quality no longer depends on where you run.",
-        "Bulk embedding measured 3.1–4.2× faster, and an online re-embed ceremony upgrades existing stores without downtime.",
-        "The engine's change feed is documented, with the SSE/WebSocket fan-out pattern for realtime surfaces."
-      ],
-      "url": null,
-      "thumb": null
-    },
-    {
-      "version": "11.0.1",
-      "date": "2026-08-31",
-      "headline": "Deletes inside transactions are safe",
-      "items": [
-        "Deleting relations inside a transact() no longer corrupts index bookkeeping.",
-        "A store that deletes its last relation keeps serving instead of refusing."
-      ],
-      "url": null,
-      "thumb": null
-    },
-    {
-      "version": "11.0.0",
-      "date": "2026-08-28",
-      "headline": "One install, one engine — Brainy",
-      "items": [
-        "The former two-package pair is one package: the native engine under the familiar API. One import is the whole install.",
-        "A missing native build refuses loudly with its cures named; nothing falls back silently.",
-        "Stores open in place — no migration."
-      ],
-      "url": null,
-      "thumb": null
-    }
-  ],
-  "history": "The version line continues from the 4.3.x native-engine releases; their record lives in the product repository's CHANGELOG.md."
-}

releases/open-brainy.json

@@ -1,110 +0,0 @@
-{
-  "product": "open-brainy",
-  "entries": [
-    {
-      "version": "10.4.9",
-      "date": "2026-09-02",
-      "headline": "Graph-first finds, honest verb arrays, and opens that stop rescanning history",
-      "items": [
-        "find({ connected, where }) now walks the neighbours first and filters only those rows — correct at every page, and O(neighbours) instead of O(store).",
-        "related() with a list of verb types (or sources, or targets) returns every requested kind — four fast paths silently kept only the first.",
-        "Deferred-embedding recovery resumes from a low-water mark instead of rescanning the whole generation log at every open — measured at two minutes on a large brain, now milliseconds."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.9",
-      "thumb": null
-    },
-    {
-      "version": "10.4.7",
-      "date": "2026-09-01",
-      "headline": "Count ledgers can no longer race themselves",
-      "items": [
-        "Concurrent count flushes coalesce into one writer with a trailing pass — parallel flushes can no longer corrupt a store's count ledger.",
-        "Atomic writes carry a per-process sequence, so two processes' temp files can never collide."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.7",
-      "thumb": null
-    },
-    {
-      "version": "10.4.6",
-      "date": "2026-08-31",
-      "headline": "Transactions cross the index seam safely",
-      "items": [
-        "Deleting relations inside a transact() no longer fails against the metadata index — operations take a JSON-safe view at the moment they execute.",
-        "Fixes a class of transaction failures on stores with integer-mapped relation endpoints."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.6",
-      "thumb": null
-    },
-    {
-      "version": "10.4.5",
-      "date": "2026-08-31",
-      "headline": "Recovery tells the truth, docs live at home",
-      "items": [
-        "A torn generation-log tail is a terminal verdict with a named cure — never an endless wait at open.",
-        "A sealed segment declares only the generations it actually holds.",
-        "The engine's documentation now publishes from its own repository."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.5",
-      "thumb": null
-    },
-    {
-      "version": "10.4.4",
-      "date": "2026-08-28",
-      "headline": "Faster opens, quieter idle",
-      "items": [
-        "Opening a store discovers generations from directory names instead of walking the log, and answers \"any entities?\" with one directory read.",
-        "The flush-request watch is event-driven; idle stores stop paying a polling heartbeat.",
-        "A slow open now names the exact step it is in, so operators see what is being paid and why."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.4",
-      "thumb": null
-    },
-    {
-      "version": "10.4.3",
-      "date": "2026-08-27",
-      "headline": "Open Brainy, under its own name",
-      "items": [
-        "The same engine as 10.4.2, now published as @soulcraftlabs/brainy — the MIT reference engine, on The Source.",
-        "No code changes; your imports change once and everything else stays put."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.3",
-      "thumb": null
-    },
-    {
-      "version": "10.4.2",
-      "date": "2026-08-27",
-      "headline": "Vectors that lie are refused, counts that drift are caught",
-      "items": [
-        "A zero-norm vector is not a vector: the index refuses them, rebuilds skip them, and a sanctioned unvector door removes them cleanly.",
-        "The canonical count ledger derives from identity records and marks legacy-derived ledgers suspect at load.",
-        "Plugin activation failures keep their original error as cause, so the real frame reaches your logs."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.2",
-      "thumb": null
-    },
-    {
-      "version": "10.4.1",
-      "date": "2026-08-26",
-      "headline": "Writes that change nothing cost nothing",
-      "items": [
-        "The read gate is per index family, and a write carrying unchanged data never re-embeds.",
-        "The vectored-row count joins the ledger, so vector coverage is a number you can read, not a guess."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.1",
-      "thumb": null
-    },
-    {
-      "version": "10.4.0",
-      "date": "2026-08-26",
-      "headline": "Repair routing, the vector ledger, and honest empties",
-      "items": [
-        "Repairs route to the index that owns the damage, and the open gate closes the vector leg until coverage is proven.",
-        "An empty string is real data, not a missing field.",
-        "The metadata crossing never carries raw integer relation endpoints — a whole class of serialization faults closed."
-      ],
-      "url": "https://source.soulcraft.com/soulcraftlabs/open-brainy/releases/tag/v10.4.0",
-      "thumb": null
-    }
-  ],
-  "history": "Earlier releases are recorded in CHANGELOG.md in this repository."
-}

scripts/release.sh

@@ -237,7 +237,7 @@ fi
 # and RELEASES.md are the record; this just gives The Source's UI a release page).
 echo -e "${BLUE}🔟 Creating release page on The Source...${NC}"
 if [ -n "${FORGEJO_RELEASE_TOKEN:-}" ]; then
-  if curl -sf -X POST "https://source.soulcraft.com/api/v1/repos/soulcraftlabs/open-brainy/releases" \
+  if curl -sf -X POST "https://source.soulcraft.com/api/v1/repos/soulcraft/brainy/releases" \
     -H "Authorization: token ${FORGEJO_RELEASE_TOKEN}" -H "Content-Type: application/json" \
     -d "{\"tag_name\":\"v${NEW_VERSION}\",\"name\":\"v${NEW_VERSION}\",\"prerelease\":${PRERELEASE}}" >/dev/null; then
     echo -e "${GREEN}✅ Release page created on The Source${NC}\n"

src/brainy.ts

@@ -15,6 +15,7 @@ import { JsHnswVectorIndex } from './hnsw/hnswIndex.js'
 import { createStorage, resolveFilesystemRoot } from './storage/storageFactory.js'
 import type { StorageOptions } from './storage/storageFactory.js'
 import { rebuildCounts } from './utils/rebuildCounts.js'
+import { jsonSafeIndexMetadata } from './utils/jsonSafeIndexMetadata.js'
 import type { MetadataWriteBuffer } from './utils/metadataWriteBuffer.js'
 import { BaseStorage } from './storage/baseStorage.js'
 import {
@@ -1819,6 +1820,11 @@ export class Brainy<T = any> implements BrainyInterface<T> {
       // a deferred write's ack and its background embed DELAYED a vector;
       // this is where it lands.
       if (!this.isReadOnly) {
+        // Foreground, as the crash-recovery contract pins it: a reopened brain
+        // has its markers re-armed when open() returns. The low-water mark
+        // bounds this to the log's tail on any brain that has ever drained —
+        // milliseconds — so the foreground cost is the unmarked first open
+        // only, once per upgraded brain.
         try {
           await step(
             'bridge-pending-embed-sidecars',
@@ -1827,7 +1833,7 @@ export class Brainy<T = any> implements BrainyInterface<T> {
           )
           await step(
             'recover-pending-embeds',
-            'folding the generation log\'s deferred-embed markers back into the pending set',
+            'folding the generation log\'s deferred-embed markers (from the low-water mark) into the pending set',
             () => this.recoverPendingEmbedsFromLog()
           )
           if (this._pendingEmbedIds.size > 0) {
@@ -2407,6 +2413,17 @@ export class Brainy<T = any> implements BrainyInterface<T> {
    */
   private static readonly PENDING_EMBED_PREFIX = '_system/pending_embeds/'
 
+  /**
+   * Storage-root-relative path of the ADVISORY pending-embed low-water mark:
+   * `{ generation, writtenAt }`, written whenever the pending set drains to
+   * empty (and at clean close when empty). Every marker in facts at or below
+   * `generation` is consumed, so recovery scans from `generation + 1`. The
+   * mark is advisory and monotone-safe: stale-low costs a longer scan, never
+   * a lost marker; it is never required for correctness.
+   */
+  private static readonly PENDING_EMBED_LOWWATER_PATH = '_system/pending_embeds_lowwater.json'
+
+
   /**
    * @description Mark a deferred embed pending (MT5): the id joins the
    * in-memory fast-path set and the returned `embed.pending` record is
@@ -2434,6 +2451,40 @@ export class Brainy<T = any> implements BrainyInterface<T> {
    */
   private clearPendingEmbed(id: string): void {
     this._pendingEmbedIds.delete(id)
+    if (this._pendingEmbedIds.size === 0) this.maybeWriteEmbedLowWater()
+  }
+
+  /**
+   * @description Advance the advisory low-water mark: called at drain-to-empty
+   * (and at clean close when empty), it records the fact log's CURRENT head —
+   * with the set empty, every marker at or below the head has been consumed,
+   * so the next open's recovery fold scans only what comes after. Fire-and-
+   * forget at the drain (close() awaits the core); loud on failure: a missed
+   * write costs the next open a longer scan, never a marker. No-op without a
+   * fact log (no durable markers exist there) and on read-only opens.
+   */
+  private maybeWriteEmbedLowWater(): void {
+    void this.writeEmbedLowWater()
+  }
+
+  /** The awaitable core of {@link maybeWriteEmbedLowWater} — close() awaits it. */
+  private async writeEmbedLowWater(): Promise<void> {
+    if (this.isReadOnly) return
+    const log = this.generationStore ? this.generationStore.getFactLog() : null
+    if (!log) return
+    const generation = log.headGeneration()
+    if (!(generation > 0)) return
+    try {
+      await this.storage.writeRawObject(Brainy.PENDING_EMBED_LOWWATER_PATH, {
+        generation,
+        writtenAt: Date.now()
+      })
+    } catch (err) {
+      prodLog.warn(
+        `[Brainy] pending-embed low-water write failed at generation ${generation}: ` +
+          `${(err as Error).message} — the next open scans from the previous mark`
+      )
+    }
   }
 
   /**
@@ -2444,9 +2495,14 @@ export class Brainy<T = any> implements BrainyInterface<T> {
    * survives the fold is exactly the set of acknowledged deferred writes
    * whose vectors have not landed.
    *
-   * BOUND (honest): no durable low-water mark exists for the earliest
+   * BOUND: the scan starts at the advisory low-water mark
-   * unconsumed pending, so the fold scans the log's committed facts from
+   * ({@link Brainy.PENDING_EMBED_LOWWATER_PATH}) — the log head at which the
-   * generation 1 — a sequential read of the log at open, O(log bytes).
+   * pending set last drained to empty — so a settled brain reads only the
+   * facts since then, not its whole history. Without a mark (first open
+   * after upgrade) it scans from generation 1, once; a stale-low mark costs
+   * a longer scan, never a marker. The fold stays on the open's foreground —
+   * the crash-recovery contract pins that a reopened brain has its markers
+   * re-armed when open() returns — and the mark is what makes that cheap.
    * It is SKIPPED WHOLESALE when the log has never had a v2 tail
    * ({@link FactLog.hasV2History} — v1 facts cannot carry marker records),
    * so pre-cutover brains pay nothing; on a mixed log the scan still reads
@@ -2459,7 +2515,18 @@ export class Brainy<T = any> implements BrainyInterface<T> {
   private async recoverPendingEmbedsFromLog(): Promise<void> {
     const log = this.generationStore.getFactLog()
     if (!log || !log.hasV2History()) return
-    const scan = log.scanFacts({ fromGeneration: 1 })
+    let fromGeneration = 1
+    try {
+      const mark = (await this.storage.readRawObject(Brainy.PENDING_EMBED_LOWWATER_PATH)) as {
+        generation?: number
+      } | null
+      if (mark && typeof mark.generation === 'number' && mark.generation > 0) {
+        fromGeneration = mark.generation + 1
+      }
+    } catch {
+      // No mark (or unreadable): scan from 1 — correctness over cost.
+    }
+    const scan = log.scanFacts({ fromGeneration })
     for await (const batch of scan.batches()) {
       for (const fact of batch.facts) {
         for (const record of fact.records ?? []) {
@@ -4203,32 +4270,19 @@ export class Brainy<T = any> implements BrainyInterface<T> {
    */
   /**
    * @description A JSON-safe view of a record bound for the metadata-index
-   * crossing. The seam's metadata is JSON-safe BY CONTRACT (a native provider
+   * crossing — delegates to the shared {@link jsonSafeIndexMetadata} leaf,
-   * serializes it; u64 ints as Number corrupt above 2^53) — but
+   * which the metadata-index transaction operations ALSO apply at execute
-   * {@link resolveVerbEndpointInts} MIRRORS the resolved endpoint ints onto
+   * and rollback time. This plan-time wrap alone proved insufficient: it
-   * the verb object itself as BigInt (`verb.sourceInt`/`targetInt`), so a
+   * returns the same reference when the record is clean, and `transact()`'s
-   * verb object reused as index metadata carried BigInts into
+   * delete legs share that reference with a graph-retraction op whose
-   * JSON.stringify, which throws, aborting the whole transaction (found by
+   * execute-time endpoint resolution mirrors BigInt ints onto it (the full
-   * the first joint pair gate). Endpoint ints ride their OWN op params on the
+   * aliasing story lives on the leaf module's doc).
-   * graph legs — the metadata crossing drops every BigInt-valued top-level
-   * key instead of guessing at a lossy numeric encoding.
    * @param metadata - The candidate index-metadata record.
    * @returns The same object when already JSON-safe, else a shallow copy
    *   without the BigInt-valued keys.
    */
   private static jsonSafeIndexMetadata(metadata: unknown): unknown {
-    if (metadata === null || typeof metadata !== 'object') return metadata
+    return jsonSafeIndexMetadata(metadata)
-    const rec = metadata as Record<string, unknown>
-    let hasBigint = false
-    for (const k in rec) {
-      if (typeof rec[k] === 'bigint') { hasBigint = true; break }
-    }
-    if (!hasBigint) return metadata
-    const out: Record<string, unknown> = {}
-    for (const k in rec) {
-      if (typeof rec[k] !== 'bigint') out[k] = rec[k]
-    }
-    return out
   }
 
   private metadataIndexRetractionOp(
@@ -7482,7 +7536,37 @@ export class Brainy<T = any> implements BrainyInterface<T> {
       // JS path — there the materialized `candidateIds` restricts the walk instead.
       let preResolvedAllowedIds: OpaqueIdSet | undefined
 
-      if (params.where || params.type || params.subtype || params.service || params.excludeVFS) {
+      // Graph-first law (10.4.8, BRAINY-PROD-LATENCY-TRIAD rounds 44/45): with
+      // `connected` present the NEIGHBOUR SET is the candidate universe. It is
+      // resolved first from the adjacency (O(neighbours)), the metadata filter
+      // is evaluated over those ids only, and paging happens LAST. The earlier
+      // order materialized the whole-store filtered id list, paged it, hydrated
+      // the page, and only then intersected with the neighbours — O(store) per
+      // call, and a neighbour outside the first page was silently dropped.
+      let graphFirstIds: string[] | null = null
+      if (hasGraphCriteria) {
+        graphFirstIds = await this.resolveConnectedIds(params)
+        if (hiddenIds.size > 0) {
+          graphFirstIds = graphFirstIds.filter((id) => !hiddenIds.has(id))
+        }
+        if (
+          graphFirstIds.length > 0 &&
+          (params.where || params.type || params.subtype || params.service || params.excludeVFS)
+        ) {
+          preResolvedFilter = this.buildMetadataFilter(params)
+          graphFirstIds = await this.filterIdsWithinBelted(preResolvedFilter, graphFirstIds)
+        }
+        if (graphFirstIds.length === 0) {
+          return []
+        }
+        if (!hasVectorSearchCriteria) {
+          return await this.pageConnectedIds(params, graphFirstIds)
+        }
+        // The vector leg walks ONLY the neighbours (its candidate walk). The
+        // filter is already applied above, so no opaque universe is produced —
+        // it would describe the whole store, not the neighbour set.
+        preResolvedMetadataIds = graphFirstIds
+      } else if (params.where || params.type || params.subtype || params.service || params.excludeVFS) {
         preResolvedFilter = this.buildMetadataFilter(params)
         preResolvedMetadataIds = await this.filterIdsBelted(preResolvedFilter)
 
@@ -7671,9 +7755,11 @@ export class Brainy<T = any> implements BrainyInterface<T> {
         }
       }
 
-      // Graph search component with O(1) traversal
+      // The text leg of a hybrid find has no candidate door, so its hits are
-      if (params.connected) {
+      // held to the neighbour set here; the vector leg walked only the neighbours.
-        results = await this.executeGraphSearch(params, results)
+      if (graphFirstIds !== null && results.length > 0) {
+        const neighbourSet = new Set(graphFirstIds)
+        results = results.filter((r) => neighbourSet.has(r.id))
       }
 
       // Apply fusion scoring if requested
@@ -12788,6 +12874,29 @@ export class Brainy<T = any> implements BrainyInterface<T> {
     }
   }
 
+  /**
+   * The id-scoped twin of {@link filterIdsBelted}: evaluate `filter` over `ids`
+   * only, through the provider's own evaluation so the answer can never drift
+   * from `getIdsForFilter`'s. A provider without the door is served by its
+   * whole-store answer intersected here (the reference index implements the
+   * door itself). Same belt: field refusals cross as `BrainyFieldRefusal`.
+   */
+  private async filterIdsWithinBelted(filter: unknown, ids: readonly string[]): Promise<string[]> {
+    this.ensureIndexesLoaded(['metadata'])
+    const mip = this.metadataIndex as unknown as MetadataIndexProvider
+    try {
+      if (typeof mip.filterIdsWithin === 'function') {
+        return await mip.filterIdsWithin(filter, ids)
+      }
+      const matched = new Set(await this.metadataIndex.getIdsForFilter(filter))
+      return ids.filter((id) => matched.has(id))
+    } catch (err) {
+      const normalized = asBrainyFieldRefusal(err)
+      if (normalized) throw normalized
+      throw err
+    }
+  }
+
   async getIndexStatus(): Promise<{
     initialized: boolean
     /** `true` once open()'s index-build-if-needed step has run. Named for API
@@ -15771,16 +15880,16 @@ export class Brainy<T = any> implements BrainyInterface<T> {
   }
   
   /**
-   * Execute graph search component.
+   * Resolve `params.connected` to the neighbour id set — the graph-first
+   * find's candidate universe (deterministic traversal order, anchors excluded).
    *
    * Honors the full `GraphConstraints` contract: multi-hop `depth` (breadth-first via
-   * `neighbors()`), `via`/`type` verb-type filtering, and `direction`. Previously this read
+   * `neighbors()`), `via`/`type` verb-type filtering, and `direction`. An empty set
-   * only `from`/`to`/`direction` and did a single 1-hop `getNeighbors()`, so `depth` and `via`
+   * is re-verified against the adjacency before it is believed — a not-serving
-   * were silently ignored — `find({ connected: { from, depth: 3 } })` returned only the
+   * adjacency throws rather than answering `[]` as truth.
-   * immediate neighbour at every depth.
    */
-  private async executeGraphSearch(params: FindParams<T>, existingResults: Result<T>[]): Promise<Result<T>[]> {
+  private async resolveConnectedIds(params: FindParams<T>): Promise<string[]> {
-    if (!params.connected) return existingResults
+    if (!params.connected) return []
 
     const { from, to, depth, direction = 'both' } = params.connected
     const via = params.connected.via ?? params.connected.type
@@ -15834,8 +15943,8 @@ export class Brainy<T = any> implements BrainyInterface<T> {
           if (anchorInt === undefined) return new Set() // unmapped → no relations
 
           const verbTypeIndex = TypeUtils.getVerbIndex(via as VerbType)
-          // No limit: match the JS BFS exactly — overall result limiting happens
+          // No limit: match the JS BFS exactly — the page is cut downstream,
-          // downstream against existingResults.
+          // after the metadata filter, by pageConnectedIds / the candidate walk.
           const reachedInts = await provider.findConnectedSubtype(
             anchorInt, verbTypeIndex, subtypeArr[0], effectiveDepth, null
           )
@@ -15920,22 +16029,44 @@ export class Brainy<T = any> implements BrainyInterface<T> {
       await this.verifyGraphAdjacencyLive()
     }
 
-    // Filter existing results to only connected entities
+    return [...connectedIds]
-    if (existingResults.length > 0) {
+  }
-      return existingResults.filter(r => connectedIds.has(r.id))
-    }
 
-    // Batch-load connected entities for fast cloud-storage performance
+  /**
+   * Page and hydrate an already-filtered neighbour set — the pure graph (and
+   * graph + metadata) find's tail. `orderBy` sorts the WHOLE set by field value
+   * before the page is cut (never the page after), null values last on `asc`
+   * and first on `desc`; without `orderBy` the traversal order stands.
+   */
+  private async pageConnectedIds(params: FindParams<T>, ids: string[]): Promise<Result<T>[]> {
+    const limit = params.limit || 10
+    const offset = params.offset || 0
+    let ordered = ids
+    if (params.orderBy) {
+      const field = params.orderBy
+      const asc = (params.order || 'asc') === 'asc'
+      const valued = await Promise.all(
+        ids.map(async (id) => ({ id, value: await this.metadataIndex.getFieldValueForEntity(id, field) }))
+      )
+      valued.sort((a, b) => {
+        if (a.value == null && b.value == null) return 0
+        if (a.value == null) return asc ? 1 : -1
+        if (b.value == null) return asc ? -1 : 1
+        if (a.value === b.value) return 0
+        const comparison = a.value < b.value ? -1 : 1
+        return asc ? comparison : -comparison
+      })
+      ordered = valued.map((v) => v.id)
+    }
+    const pageIds = ordered.slice(offset, offset + limit)
+    const entitiesMap = await this.batchGet(pageIds)
     const results: Result<T>[] = []
-    const ids = [...connectedIds]
+    for (const id of pageIds) {
-    const entitiesMap = await this.batchGet(ids)
-    for (const id of ids) {
       const entity = entitiesMap.get(id)
       if (entity) {
         results.push(this.createResult(id, 1.0, entity))
       }
     }
-
     return results
   }
   
@@ -19455,6 +19586,7 @@ export class Brainy<T = any> implements BrainyInterface<T> {
    *   terminal releases have run.
    */
   async close(): Promise<void> {
+    if (this._pendingEmbedIds.size === 0) await this.writeEmbedLowWater()
     let closeFailure: unknown = null
     try {
       await this.closeDurableSteps()

src/plugin.ts

@@ -411,6 +411,19 @@ export interface MetadataIndexProvider {
    * @returns The matching id universe as an opaque set.
    */
   getIdSetForFilter?(filter: any): Promise<OpaqueIdSet>
+  /**
+   * @description OPTIONAL: evaluate `filter` over `ids` ONLY and return the
+   * survivors in the caller's order — the door a graph-first
+   * `find({ connected, where })` walks. The neighbour set is the universe there,
+   * so the filter must cost O(|ids|) membership checks, never a whole-store
+   * materialization. A native index answers from its roaring filter result
+   * (membership by entity int); the reference index answers from its own
+   * `getIdsForFilter`, so the two doors can never disagree. Absent → Brainy
+   * intersects `getIdsForFilter`'s answer with `ids` itself (correct, O(store)).
+   * @param filter - The same filter shape accepted by `getIdsForFilter`.
+   * @param ids - The candidate ids (canonical). The answer is a subsequence.
+   */
+  filterIdsWithin?(filter: any, ids: readonly string[]): Promise<string[]>
   getIdsForTextQuery(query: string): Promise<Array<{ id: string; matchCount: number }>>
   getSortedIdsForFilter(filter: any, orderBy: string, order?: 'asc' | 'desc', topK?: number): Promise<string[]>
   getFilterValues(field: string): Promise<string[]>

src/storage/adapters/baseStorageAdapter.ts

@@ -1089,6 +1089,10 @@ export abstract class BaseStorageAdapter implements StorageAdapter {
 
   // Counts changed since the last persist? Drives the write-through flush.
   protected pendingCountPersist = false
+  /** The one persist running right now, if any (single-flight law — see flushCounts). */
+  private countPersistInFlight: Promise<void> | null = null
+  /** The one trailing persist a burst has queued behind the in-flight one. */
+  private countPersistTrailing: Promise<void> | null = null
 
   /**
    * Get total noun count - O(1) operation
@@ -1341,15 +1345,46 @@ export abstract class BaseStorageAdapter implements StorageAdapter {
       return
     }
 
-    try {
+    // SINGLE-FLIGHT, COALESCED. Counts are write-through on every change, so
-      // Persist to storage (implemented by subclass)
+    // a burst of writes used to launch one persist per change, all in flight
-      await this.persistCounts()
+    // together. Two of them inside the same millisecond shared the atomic
-      this.pendingCountPersist = false
+    // writer's temp path (`.tmp-<pid>-<ms>`): both wrote it, the first rename
-    } catch (error) {
+    // consumed it, the second rename found nothing — ENOENT, ~1,500 times a
-      console.error('CRITICAL: Failed to flush counts to storage:', error)
+    // day on a busy production brain, with a full ledger write per change
-      // Keep pending flag set so we retry on next operation
+    // behind it. Now exactly one persist runs at a time; requests that arrive
-      throw error
+    // while it runs collapse into ONE trailing persist that carries the final
+    // state. A burst of N changes costs at most two writes and never races
+    // itself.
+    if (this.countPersistInFlight) {
+      // The in-flight write may have already serialised a stale snapshot —
+      // ask for one more pass after it, and let every caller in this burst
+      // await that same pass.
+      if (!this.countPersistTrailing) {
+        this.countPersistTrailing = this.countPersistInFlight
+          .catch(() => undefined)
+          .then(() => {
+            this.countPersistTrailing = null
+            return this.flushCounts()
+          })
+      }
+      return this.countPersistTrailing
     }
+
+    this.countPersistInFlight = (async () => {
+      try {
+        // Persist to storage (implemented by subclass)
+        this.pendingCountPersist = false
+        await this.persistCounts()
+      } catch (error) {
+        // Keep the flag set so the next operation retries.
+        this.pendingCountPersist = true
+        console.error('CRITICAL: Failed to flush counts to storage:', error)
+        throw error
+      } finally {
+        this.countPersistInFlight = null
+      }
+    })()
+    return this.countPersistInFlight
   }
 
   /**

src/storage/adapters/fileSystemStorage.ts

@@ -2400,8 +2400,15 @@ export class FileSystemStorage extends BaseStorage {
    * Atomic write via temp-file-then-rename so concurrent readers never see a
    * half-written lock JSON. Reused by writer-lock writes + heartbeat.
    */
+  /** Monotonic per-process sequence so two atomic writes never share a temp path. */
+  private static atomicWriteSeq = 0
+
   private async writeFileAtomic(filePath: string, contents: string): Promise<void> {
-    const tmp = `${filePath}.tmp-${process.pid}-${Date.now()}`
+    // pid + timestamp alone collided: two writers of the same target inside
+    // one millisecond shared this path, and the loser's rename found the
+    // winner had already moved it (ENOENT). The sequence makes every call's
+    // temp path its own.
+    const tmp = `${filePath}.tmp-${process.pid}-${Date.now()}-${++FileSystemStorage.atomicWriteSeq}`
     await fs.promises.writeFile(tmp, contents)
     await fs.promises.rename(tmp, filePath)
   }

src/storage/baseStorage.ts

@@ -2942,19 +2942,33 @@ export abstract class BaseStorage extends BaseStorageAdapter {
         !options.filter.service &&
         !options.filter.metadata
       ) {
-        const sourceId = Array.isArray(options.filter.sourceId)
+        const sourceIds = Array.isArray(options.filter.sourceId)
-          ? options.filter.sourceId[0]
+          ? options.filter.sourceId
-          : options.filter.sourceId
+          : [options.filter.sourceId]
 
-        const verbType = Array.isArray(options.filter.verbType)
+        // EVERY requested verb type is honoured — an array used to collapse to
-          ? options.filter.verbType[0]
+        // its first element here, silently dropping the rest of the ask.
-          : options.filter.verbType
+        const verbTypes = new Set(
+          Array.isArray(options.filter.verbType)
+            ? options.filter.verbType
+            : [options.filter.verbType]
+        )
 
-        // Get verbs by source, then filter by type (O(1) graph lookup + O(n) type filter),
+        // Get verbs by source (union over every requested source), filter by the
-        // then apply the subtype / visibility metadata filters on the candidate set.
+        // requested type SET (O(1) graph lookup + O(n) type filter), then apply
-        const verbsBySource = await this.getVerbsBySource_internal(sourceId)
+        // the subtype / visibility metadata filters on the candidate set.
+        const bySource: HNSWVerbWithMetadata[] = []
+        const seenVerbIds = new Set<string>()
+        for (const oneSource of sourceIds) {
+          for (const v of await this.getVerbsBySource_internal(oneSource)) {
+            if (!seenVerbIds.has(v.id)) {
+              seenVerbIds.add(v.id)
+              bySource.push(v)
+            }
+          }
+        }
         const filteredVerbs = this.applyVerbMetadataFilters(
-          verbsBySource.filter(v => v.verb === verbType),
+          bySource.filter(v => verbTypes.has(v.verb)),
           options.filter
         )
 
@@ -2985,16 +2999,22 @@ export abstract class BaseStorage extends BaseStorageAdapter {
         !options.filter.service &&
         !options.filter.metadata
       ) {
-        const sourceId = Array.isArray(options.filter.sourceId)
+        // EVERY requested source is honoured — an array used to collapse to
-          ? options.filter.sourceId[0]
+        // its first element here, silently dropping the rest of the ask.
-          : options.filter.sourceId
+        const onlySourceIds = Array.isArray(options.filter.sourceId)
-
+          ? options.filter.sourceId
-        // Get verbs by source directly (hydrated with metadata), then apply the
+          : [options.filter.sourceId]
-        // subtype / visibility metadata filters on the O(degree) candidate set.
+        const sourceUnion: HNSWVerbWithMetadata[] = []
-        const verbsBySource = this.applyVerbMetadataFilters(
+        const seenSourceVerbIds = new Set<string>()
-          await this.getVerbsBySource_internal(sourceId),
+        for (const oneSource of onlySourceIds) {
-          options.filter
+          for (const v of await this.getVerbsBySource_internal(oneSource)) {
-        )
+            if (!seenSourceVerbIds.has(v.id)) {
+              seenSourceVerbIds.add(v.id)
+              sourceUnion.push(v)
+            }
+          }
+        }
+        const verbsBySource = this.applyVerbMetadataFilters(sourceUnion, options.filter)
 
         // Apply pagination
         const paginatedVerbs = verbsBySource.slice(offset, offset + limit)
@@ -3023,16 +3043,22 @@ export abstract class BaseStorage extends BaseStorageAdapter {
         !options.filter.service &&
         !options.filter.metadata
       ) {
-        const targetId = Array.isArray(options.filter.targetId)
+        // EVERY requested target is honoured — an array used to collapse to
-          ? options.filter.targetId[0]
+        // its first element here, silently dropping the rest of the ask.
-          : options.filter.targetId
+        const onlyTargetIds = Array.isArray(options.filter.targetId)
-
+          ? options.filter.targetId
-        // Get verbs by target directly (hydrated with metadata), then apply the
+          : [options.filter.targetId]
-        // subtype / visibility metadata filters on the O(degree) candidate set.
+        const targetUnion: HNSWVerbWithMetadata[] = []
-        const verbsByTarget = this.applyVerbMetadataFilters(
+        const seenTargetVerbIds = new Set<string>()
-          await this.getVerbsByTarget_internal(targetId),
+        for (const oneTarget of onlyTargetIds) {
-          options.filter
+          for (const v of await this.getVerbsByTarget_internal(oneTarget)) {
-        )
+            if (!seenTargetVerbIds.has(v.id)) {
+              seenTargetVerbIds.add(v.id)
+              targetUnion.push(v)
+            }
+          }
+        }
+        const verbsByTarget = this.applyVerbMetadataFilters(targetUnion, options.filter)
 
         // Apply pagination
         const paginatedVerbs = verbsByTarget.slice(offset, offset + limit)
@@ -3061,16 +3087,25 @@ export abstract class BaseStorage extends BaseStorageAdapter {
         !options.filter.service &&
         !options.filter.metadata
       ) {
-        const verbType = Array.isArray(options.filter.verbType)
+        // EVERY requested verb type is honoured — an array used to collapse to
-          ? options.filter.verbType[0]
+        // its first element here, silently dropping the rest of the ask.
-          : options.filter.verbType
+        const verbTypes = Array.isArray(options.filter.verbType)
+          ? options.filter.verbType
+          : [options.filter.verbType]
 
-        // Get verbs by type directly (hydrated with metadata), then apply the
+        // Get verbs by each requested type (hydrated with metadata), deduped by
-        // subtype / visibility metadata filters on the candidate set.
+        // id, then apply the subtype / visibility metadata filters on the set.
-        const verbsByType = this.applyVerbMetadataFilters(
+        const byType: HNSWVerbWithMetadata[] = []
-          await this.getVerbsByType_internal(verbType),
+        const seenTypeVerbIds = new Set<string>()
-          options.filter
+        for (const oneType of verbTypes) {
-        )
+          for (const v of await this.getVerbsByType_internal(oneType)) {
+            if (!seenTypeVerbIds.has(v.id)) {
+              seenTypeVerbIds.add(v.id)
+              byType.push(v)
+            }
+          }
+        }
+        const verbsByType = this.applyVerbMetadataFilters(byType, options.filter)
 
         // Apply pagination
         const paginatedVerbs = verbsByType.slice(offset, offset + limit)

src/transaction/operations/IndexOperations.ts

@@ -14,6 +14,7 @@ import type { MetadataIndexManager } from '../../utils/metadataIndex.js'
 import type { GraphVerb } from '../../coreTypes.js'
 import type { Operation, RollbackAction } from '../types.js'
 import { isZeroNormVector } from '../../utils/distance.js'
+import { jsonSafeIndexMetadata } from '../../utils/jsonSafeIndexMetadata.js'
 import { prodLog } from '../../utils/logger.js'
 
 /**
@@ -390,13 +391,21 @@ export class AddToMetadataIndexOperation implements Operation {
     // rollback so add + undo reference the same watermark.
     const generation = this.generationFn?.()
 
-    // Add to metadata index (skipFlush=true for transaction atomicity)
+    // The JSON-safe view is taken HERE, per crossing, never at construction:
-    await this.index.addToIndex(this.id, this.entity, true, false, generation)
+    // the entity reference this op holds can be mutated between plan and
+    // execute (a graph op's execute-time endpoint-int resolution mirrors
+    // BigInts onto a shared verb object) — see jsonSafeIndexMetadata's
+    // module doc.
+    await this.index.addToIndex(
+      this.id, jsonSafeIndexMetadata(this.entity), true, false, generation
+    )
 
     // Return rollback action
     return async () => {
       // Remove from metadata index
-      await this.index.removeFromIndex(this.id, this.entity, generation)
+      await this.index.removeFromIndex(
+        this.id, jsonSafeIndexMetadata(this.entity), generation
+      )
     }
   }
 }
@@ -432,13 +441,21 @@ export class RemoveFromMetadataIndexOperation implements Operation {
     // Resolve the removal generation once; reuse it for the rollback re-add.
     const generation = this.generationFn?.()
 
-    // Remove from metadata index
+    // Sanitized per crossing, never at construction — transact()'s delete
-    await this.index.removeFromIndex(this.id, this.entity, generation)
+    // legs hand this op the SAME verb object the graph-retraction op's
+    // execute-time endpoint resolution mutates (BigInt sourceInt/targetInt),
+    // so a plan-time view aliases the pollution. See jsonSafeIndexMetadata's
+    // module doc.
+    await this.index.removeFromIndex(
+      this.id, jsonSafeIndexMetadata(this.entity), generation
+    )
 
     // Return rollback action
     return async () => {
       // Re-add with original metadata (skipFlush=true)
-      await this.index.addToIndex(this.id, this.entity, true, false, generation)
+      await this.index.addToIndex(
+        this.id, jsonSafeIndexMetadata(this.entity), true, false, generation
+      )
     }
   }
 }

src/utils/jsonSafeIndexMetadata.ts

@@ -0,0 +1,47 @@
+/**
+ * @module utils/jsonSafeIndexMetadata
+ * @description The metadata-index crossing's JSON-safety law, as a leaf
+ * function both the coordinator and the transaction operations share.
+ *
+ * The seam's metadata is JSON-safe BY CONTRACT (a native provider serializes
+ * it; u64 ints as Number corrupt above 2^53) — but `resolveVerbEndpointInts`
+ * MIRRORS the resolved endpoint ints onto the verb object itself as BigInt
+ * (`verb.sourceInt`/`targetInt`), so a verb object reused as index metadata
+ * carries BigInts into JSON.stringify, which throws, aborting the whole
+ * transaction. Endpoint ints ride their OWN op params on the graph legs — the
+ * metadata crossing drops every BigInt-valued top-level key instead of
+ * guessing at a lossy numeric encoding.
+ *
+ * WHY THIS IS A LEAF MODULE, ENFORCED AT THE CROSSING: sanitizing only at
+ * operation-construction time is not enough. `transact()`'s delete legs pass
+ * the SAME verb object to both the graph-retraction op (whose endpoint-int
+ * thunk deliberately resolves at EXECUTE time, for same-batch forward refs)
+ * and the metadata-retraction op. At plan time the verb is still clean, so a
+ * plan-time sanitize returns the same reference — then the graph op executes
+ * first, mirrors the BigInt ints onto the shared object, and the metadata op
+ * crosses the seam with them (found by the first fleet adoption of the native
+ * pair: every transact-wrapped edge delete aborted). The crossing itself is
+ * the only place ordering cannot bypass.
+ */
+
+/**
+ * A JSON-safe view of a record bound for the metadata-index crossing.
+ *
+ * @param metadata - The candidate index-metadata record.
+ * @returns The same object when already JSON-safe, else a shallow copy
+ *   without the BigInt-valued keys.
+ */
+export function jsonSafeIndexMetadata(metadata: unknown): unknown {
+  if (metadata === null || typeof metadata !== 'object') return metadata
+  const rec = metadata as Record<string, unknown>
+  let hasBigint = false
+  for (const k in rec) {
+    if (typeof rec[k] === 'bigint') { hasBigint = true; break }
+  }
+  if (!hasBigint) return metadata
+  const out: Record<string, unknown> = {}
+  for (const k in rec) {
+    if (typeof rec[k] !== 'bigint') out[k] = rec[k]
+  }
+  return out
+}

src/utils/metadataIndex.ts

@@ -2575,6 +2575,19 @@ export class MetadataIndexManager implements MetadataIndexProvider {
   /** Once-per-field flag for the fallback-degradation announcement. */
   private static announcedFallbackSorts = new Set<string>()
 
+  /**
+   * Evaluate `filter` over `ids` only — the graph-first find's door (the
+   * neighbour set filtered by id, never the store filtered and then
+   * intersected). This index answers from its own `getIdsForFilter`, so the
+   * two doors cannot disagree; the cost is that of the filter over this
+   * in-memory index, and the answer keeps the caller's order.
+   */
+  async filterIdsWithin(filter: any, ids: readonly string[]): Promise<string[]> {
+    if (ids.length === 0) return []
+    const matched = new Set(await this.getIdsForFilter(filter))
+    return ids.filter((id) => matched.has(id))
+  }
+
   async getSortedIdsForFilter(
     filter: any,
     orderBy: string,

tests/integration/counts-persist-single-flight.test.ts

@@ -0,0 +1,111 @@
+/**
+ * @module tests/integration/counts-persist-single-flight
+ * @description Regression for a production race in FileSystemStorage's
+ * counts ledger: `persistCounts()` was write-through on every count change
+ * with no serialization, and the atomic writer named its temp file with
+ * millisecond granularity (`.tmp-<pid>-<ms>`). Two persists inside one
+ * millisecond shared the temp path — both wrote it, the first rename
+ * consumed it, the second rename found nothing: ENOENT, ~1,500 times a day
+ * on a busy production brain, with a full ledger write per change behind it.
+ *
+ * Under pin: persists are single-flight and coalesced — one in flight, at
+ * most one trailing pass carrying the burst's final state — and every atomic
+ * write owns a unique temp path. A burst of N count changes costs at most
+ * two ledger writes, never errors, and leaves a ledger equal to memory.
+ */
+import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest'
+import * as fs from 'node:fs'
+import * as os from 'node:os'
+import * as path from 'node:path'
+import { Brainy } from '../../src/brainy.js'
+import { NounType } from '../../src/types/graphTypes.js'
+
+describe('counts persistence is single-flight, coalesced, and never races its own temp file', () => {
+  let dir: string
+  let brain: any
+
+  beforeEach(async () => {
+    process.env.BRAINY_DETERMINISTIC_EMBEDDINGS = 'true'
+    dir = fs.mkdtempSync(path.join(os.tmpdir(), 'brainy-counts-race-'))
+    brain = new Brainy({
+      requireSubtype: false,
+      storage: { type: 'filesystem', path: dir },
+      dimensions: 384,
+      silent: true
+    })
+    await brain.init()
+  })
+
+  afterEach(async () => {
+    vi.restoreAllMocks()
+    await brain.close()
+    fs.rmSync(dir, { recursive: true, force: true })
+  })
+
+  it('a burst of concurrent count changes → at most two ledger writes, zero errors, ledger == memory', async () => {
+    const storage = brain.storage
+    const countsPath: string = storage.countsFilePath
+    expect(countsPath, 'the filesystem adapter persists a counts ledger').toBeTruthy()
+
+    // Let init's own persists settle so the burst is measured alone.
+    await storage.flushCounts?.()
+
+    const renameSpy = vi.spyOn(fs.promises, 'rename')
+    const errorSpy = vi.spyOn(console, 'error')
+
+    // Twenty-five concurrent count changes — the shape of a write burst; each
+    // used to launch its own persist.
+    const BURST = 25
+    await Promise.all(
+      Array.from({ length: BURST }, () => storage.scheduleCountPersist())
+    )
+
+    const ledgerRenames = renameSpy.mock.calls.filter(([, to]) => String(to) === countsPath)
+    expect(ledgerRenames.length, 'single-flight + one trailing pass').toBeLessThanOrEqual(2)
+    expect(ledgerRenames.length, 'the burst was persisted at all').toBeGreaterThanOrEqual(1)
+
+    const persistErrors = errorSpy.mock.calls.filter((args) => String(args[0]).includes('persisting counts'))
+    expect(persistErrors).toEqual([])
+
+    const ledger = JSON.parse(fs.readFileSync(countsPath, 'utf-8'))
+    expect(ledger.totalNounCount).toBe(storage.totalNounCount)
+    expect(ledger.totalVerbCount).toBe(storage.totalVerbCount)
+  })
+
+  it('real writes in parallel: the ledger lands complete and no persist error is logged', async () => {
+    const storage = brain.storage
+    const countsPath: string = storage.countsFilePath
+    const errorSpy = vi.spyOn(console, 'error')
+
+    await Promise.all(
+      Array.from({ length: 12 }, (_, i) =>
+        brain.add({ data: `burst row ${i}`, type: NounType.Thing })
+      )
+    )
+    await storage.flushCounts?.()
+
+    const persistErrors = errorSpy.mock.calls.filter((args) => String(args[0]).includes('persisting counts'))
+    expect(persistErrors).toEqual([])
+    const ledger = JSON.parse(fs.readFileSync(countsPath, 'utf-8'))
+    expect(ledger.totalNounCount).toBe(storage.totalNounCount)
+    expect(await brain.getNounCount()).toBe(ledger.totalNounCount)
+  })
+
+  it('every atomic write owns its own temp path — two writes in one millisecond never collide', async () => {
+    const storage = brain.storage
+    const tmpNames: string[] = []
+    vi.spyOn(fs.promises, 'writeFile').mockImplementation(async (p: any) => {
+      tmpNames.push(String(p))
+    })
+    vi.spyOn(fs.promises, 'rename').mockImplementation(async () => undefined)
+    const target = path.join(dir, 'probe.json')
+    await Promise.all([
+      storage.writeFileAtomic(target, '{"a":1}'),
+      storage.writeFileAtomic(target, '{"a":2}'),
+      storage.writeFileAtomic(target, '{"a":3}')
+    ])
+    const probeTmps = tmpNames.filter((n) => n.startsWith(`${target}.tmp-`))
+    expect(probeTmps.length).toBe(3)
+    expect(new Set(probeTmps).size, 'no two writes shared a temp path').toBe(3)
+  })
+})

tests/integration/find-connected-order.test.ts

@@ -0,0 +1,165 @@
+/**
+ * @module tests/integration/find-connected-order
+ * @description The graph-first law for `find({ connected })` (10.4.8).
+ *
+ * With `connected` present the neighbour set is the candidate universe: it is
+ * resolved from the adjacency first, the metadata filter is evaluated over
+ * those ids only, and the page is cut last. The earlier order materialized the
+ * whole-store filtered id list, paged it, hydrated the page, and only then
+ * intersected with the neighbours — so a neighbour outside the first page of
+ * the filtered STORE was silently dropped, and every call paid O(store).
+ *
+ * These pins hold both halves. The answer: every matching neighbour is
+ * reachable by paging, a non-neighbour never appears, a negation (`missing`)
+ * is evaluated over the neighbours, `orderBy` sorts the whole neighbour set
+ * before the page is cut, and the vector leg walks the neighbours only. The
+ * cost shape: the metadata index is asked about the neighbour ids only, and
+ * hydration is one page — never the store.
+ */
+import { describe, it, expect, beforeAll, afterAll, vi } from 'vitest'
+import { Brainy } from '../../src/brainy'
+import { NounType, VerbType } from '../../src/types/graphTypes'
+import { v5 } from '../../src/universal/uuid'
+import { generateTestVector } from '../helpers/test-factory'
+
+/** Matching rows that are NOT neighbours — added FIRST, so the whole-store filtered list leads with them. */
+const NOISE = 120
+/** Matching rows that ARE neighbours of the anchor. */
+const NEIGHBOURS = 30
+/** Neighbours carrying `retracted: true` — excluded by the `missing` negation. */
+const RETRACTED = 4
+
+describe('find({ connected }) is graph-first: neighbours → filter → page', () => {
+  let brain: Brainy<any>
+  const anchor = 'anchor'
+  const sharedVector = generateTestVector()
+  const neighbourIds = new Set(Array.from({ length: NEIGHBOURS }, (_, i) => v5(`nb-${i}`)))
+
+  beforeAll(async () => {
+    brain = new Brainy({ requireSubtype: false, storage: { type: 'memory' } })
+    await brain.init()
+    await brain.add({
+      id: anchor,
+      data: 'the anchor',
+      type: NounType.Person,
+      metadata: { kind: 'anchor' },
+      vector: generateTestVector()
+    })
+    for (let i = 0; i < NOISE; i++) {
+      await brain.add({
+        id: `noise-${i}`,
+        data: `noise ${i}`,
+        type: NounType.Person,
+        metadata: { kind: 'note', rank: 1000 + i },
+        vector: sharedVector
+      })
+    }
+    for (let i = 0; i < NEIGHBOURS; i++) {
+      await brain.add({
+        id: `nb-${i}`,
+        data: `neighbour ${i}`,
+        type: NounType.Person,
+        metadata: { kind: 'note', rank: i + 1, ...(i < RETRACTED ? { retracted: true } : {}) },
+        vector: sharedVector
+      })
+      await brain.relate({ from: anchor, to: `nb-${i}`, type: VerbType.Knows })
+    }
+  })
+
+  afterAll(async () => {
+    brain = null as any
+  })
+
+  it('returns the matching neighbours page by page — none dropped, never a non-neighbour', async () => {
+    const seen = new Set<string>()
+    for (let offset = 0; offset <= NEIGHBOURS; offset += 10) {
+      const page = await brain.find({
+        connected: { from: anchor, direction: 'out' },
+        where: { kind: 'note' },
+        limit: 10,
+        offset
+      })
+      expect(page).toHaveLength(offset < NEIGHBOURS ? 10 : 0)
+      for (const r of page) {
+        expect(neighbourIds.has(r.entity.id)).toBe(true)
+        expect(seen.has(r.entity.id)).toBe(false)
+        seen.add(r.entity.id)
+      }
+    }
+    expect(seen.size).toBe(NEIGHBOURS)
+  })
+
+  it('evaluates a negation (`missing`) over the neighbour set, not the store', async () => {
+    const results = await brain.find({
+      connected: { from: anchor, direction: 'out' },
+      where: { kind: 'note', retracted: { missing: true } },
+      limit: 100
+    })
+    expect(results).toHaveLength(NEIGHBOURS - RETRACTED)
+    for (const r of results) {
+      expect(neighbourIds.has(r.entity.id)).toBe(true)
+      expect(r.entity.metadata.retracted).toBeUndefined()
+    }
+  })
+
+  it('asks the metadata index about the neighbour ids only, and hydrates one page', async () => {
+    const index = (brain as any).metadataIndex
+    const within = vi.spyOn(index, 'filterIdsWithin')
+    const hydrate = vi.spyOn(brain as any, 'batchGet')
+    try {
+      const results = await brain.find({
+        connected: { from: anchor, direction: 'out' },
+        where: { kind: 'note' },
+        limit: 10
+      })
+      expect(results).toHaveLength(10)
+      expect(within).toHaveBeenCalledTimes(1)
+      const askedIds = within.mock.calls[0][1] as string[]
+      expect(askedIds).toHaveLength(NEIGHBOURS)
+      for (const id of askedIds) expect(neighbourIds.has(id)).toBe(true)
+      expect(hydrate).toHaveBeenCalledTimes(1)
+      expect(hydrate.mock.calls[0][0]).toHaveLength(10)
+    } finally {
+      within.mockRestore()
+      hydrate.mockRestore()
+    }
+  })
+
+  it('orders the WHOLE neighbour set before cutting the page', async () => {
+    const results = await brain.find({
+      connected: { from: anchor, direction: 'out' },
+      where: { kind: 'note' },
+      orderBy: 'rank',
+      order: 'desc',
+      limit: 5
+    })
+    expect(results.map((r) => r.entity.metadata.rank)).toEqual([30, 29, 28, 27, 26])
+  })
+
+  it('walks the vector leg over the neighbours only', async () => {
+    const results = await brain.find({
+      vector: sharedVector,
+      connected: { from: anchor, direction: 'out' },
+      where: { kind: 'note' },
+      limit: 5
+    })
+    expect(results).toHaveLength(5)
+    for (const r of results) expect(neighbourIds.has(r.entity.id)).toBe(true)
+  })
+
+  it('an anchor without neighbours answers [] before the filter is asked', async () => {
+    const index = (brain as any).metadataIndex
+    const within = vi.spyOn(index, 'filterIdsWithin')
+    try {
+      const results = await brain.find({
+        connected: { from: 'noise-0', direction: 'out' },
+        where: { kind: 'note' },
+        limit: 10
+      })
+      expect(results).toEqual([])
+      expect(within).not.toHaveBeenCalled()
+    } finally {
+      within.mockRestore()
+    }
+  })
+})

tests/integration/pending-embed-low-water.test.ts

@@ -0,0 +1,141 @@
+/**
+ * @module tests/integration/pending-embed-low-water
+ * @description The pending-embed recovery fold is bounded and background (10.4.9).
+ *
+ * The fold used to scan the generation log from generation 1 at EVERY open,
+ * on the open's foreground — O(whole history) per open on long-lived brains.
+ * Now: an advisory low-water mark (`_system/pending_embeds_lowwater.json`)
+ * records the committed generation whenever the pending set drains to empty,
+ * recovery scans from `mark + 1` on the open's foreground — the crash-recovery
+ * contract keeps markers re-armed when open() returns. The mark is advisory: stale-low costs a longer scan, never a
+ * marker — a pending embed enqueued before a crash is still recovered.
+ */
+import { describe, it, expect, afterEach, vi } from 'vitest'
+import { mkdtempSync, rmSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { Brainy } from '../../src/brainy'
+import { NounType } from '../../src/types/graphTypes'
+
+const LOWWATER_PATH = '_system/pending_embeds_lowwater.json'
+
+describe('pending-embed recovery: bounded by the low-water mark', () => {
+  const roots: string[] = []
+  const dir = (): string => {
+    const d = mkdtempSync(join(tmpdir(), 'brainy-lowwater-'))
+    roots.push(d)
+    return d
+  }
+  const open = async (root: string): Promise<Brainy<any>> => {
+    const brain = new Brainy<any>({
+      requireSubtype: false,
+      storage: { type: 'filesystem', path: root }
+    })
+    await brain.init()
+    return brain
+  }
+
+  afterEach(() => {
+    for (const d of roots.splice(0)) rmSync(d, { recursive: true, force: true })
+  })
+
+  it('drain-to-empty writes the mark, and the next open scans from mark + 1', async () => {
+    const root = dir()
+    const brain = await open(root)
+    // Hold the worker so the pending state is observable, then release it.
+    const realKick = (brain as any).kickEmbedWorker.bind(brain)
+    ;(brain as any).kickEmbedWorker = () => {}
+    await brain.add({
+      id: 'row-1',
+      data: 'the first deferred row',
+      type: NounType.Thing,
+      deferEmbedding: true
+    })
+    expect(brain.pendingEmbedCount()).toBeGreaterThan(0)
+    ;(brain as any).kickEmbedWorker = realKick
+    await brain.awaitPendingEmbeds()
+    // The drain wrote the advisory mark (fire-and-forget: settle the microtask).
+    await new Promise((r) => setTimeout(r, 50))
+    const mark = (await (brain as any).storage.readRawObject(LOWWATER_PATH)) as {
+      generation: number
+    } | null
+    expect(mark).not.toBeNull()
+    expect(mark!.generation).toBeGreaterThan(0)
+    await brain.close()
+
+    const brain2 = await open(root)
+    const log = (brain2 as any).generationStore.getFactLog()
+    const scanSpy = vi.spyOn(log, 'scanFacts')
+    try {
+      await (brain2 as any).recoverPendingEmbedsFromLog()
+      expect(scanSpy).toHaveBeenCalledTimes(1)
+      const opts = scanSpy.mock.calls[0][0] as { fromGeneration?: number }
+      expect(opts.fromGeneration).toBeGreaterThanOrEqual(mark!.generation + 1)
+    } finally {
+      scanSpy.mockRestore()
+      await brain2.close()
+    }
+  })
+
+  it('a pending embed enqueued after the mark survives an unclean stop', async () => {
+    const root = dir()
+    const brain = await open(root)
+    await brain.add({ id: 'settled', data: 'lands before the mark', type: NounType.Thing })
+    await brain.awaitPendingEmbeds()
+    await new Promise((r) => setTimeout(r, 50))
+
+    // A deferred write whose embed never lands: block the worker, then drop
+    // the instance without close() — the unclean-stop shape.
+    ;(brain as any).kickEmbedWorker = () => {}
+    await brain.add({
+      id: 'orphan',
+      data: 'enqueued then abandoned',
+      type: NounType.Thing,
+      deferEmbedding: true
+    })
+    expect(brain.pendingEmbedCount()).toBeGreaterThan(0)
+    // No close(): simulate the crash by releasing only the writer lock so the
+    // next open can proceed.
+    await (brain as any).storage.releaseWriterLock()
+
+    const brain2 = await open(root)
+    expect(brain2.pendingEmbedCount()).toBeGreaterThan(0)
+    await brain2.awaitPendingEmbeds()
+    expect(brain2.pendingEmbedCount()).toBe(0)
+    await brain2.close()
+    // Reap the crashed instance: its fence is gone, so close() fails loudly —
+    // swallow that here; the point is clearing its watchers and registry entry.
+    await brain.close().catch(() => undefined)
+  })
+
+  it('a reopened brain has its pending set settled when open() returns', async () => {
+    const root = dir()
+    const brain = await open(root)
+    await brain.add({ id: 'a-row', data: 'some data', type: NounType.Thing })
+    await brain.awaitPendingEmbeds()
+    await brain.close()
+
+    const brain2 = await open(root)
+    // The crash-recovery contract: markers are re-armed by open itself —
+    // no latch, no background race. (Here the drain landed, so zero.)
+    expect(brain2.pendingEmbedCount()).toBe(0)
+    await brain2.close()
+  })
+
+  it('a clean close with an empty set writes the mark even if no drain happened', async () => {
+    const root = dir()
+    const brain = await open(root)
+    await brain.add({ id: 'r1', data: 'row one', type: NounType.Thing })
+    await brain.awaitPendingEmbeds()
+    await brain.close()
+    // Read the mark back through the storage door (the adapter owns the
+    // on-disk encoding), on a fresh instance.
+    const brain2 = await open(root)
+    const mark = (await (brain2 as any).storage.readRawObject(LOWWATER_PATH)) as {
+      generation: number
+    } | null
+    expect(mark).not.toBeNull()
+    expect(mark!.generation).toBeGreaterThan(0)
+    await brain2.close()
+  })
+})

tests/integration/related-verb-array.test.ts

@@ -0,0 +1,89 @@
+/**
+ * @module tests/integration/related-verb-array
+ * @description related() honours EVERY verb type in an array (10.4.9).
+ *
+ * The storage fast paths for `sourceId + verbType` and `verbType` collapsed a
+ * verb-type ARRAY to its first element — `related({ from, type: [a, b] })`
+ * silently returned only `a` edges, whichever order the array came in. The
+ * same quiet-loss class as the graph-first paging defect, one seam over.
+ * These pins seed a store where the SECOND requested type's edge must come
+ * back, on every path the collapse lived in.
+ */
+import { describe, it, expect, beforeAll, afterAll } from 'vitest'
+import { Brainy } from '../../src/brainy'
+import { NounType, VerbType } from '../../src/types/graphTypes'
+import { v5 } from '../../src/universal/uuid'
+
+describe('related() with a verb-type array returns every requested type', () => {
+  let brain: Brainy<any>
+
+  beforeAll(async () => {
+    brain = new Brainy({ requireSubtype: false, storage: { type: 'memory' } })
+    await brain.init()
+    for (const id of ['a', 'b', 'c', 'd']) {
+      await brain.add({ id, data: `node ${id}`, type: NounType.Person })
+    }
+    await brain.relate({ from: 'a', to: 'b', type: VerbType.Supports })
+    await brain.relate({ from: 'a', to: 'c', type: VerbType.RelatedTo })
+    await brain.relate({ from: 'a', to: 'd', type: VerbType.Knows })
+    await brain.relate({ from: 'b', to: 'c', type: VerbType.RelatedTo })
+  })
+
+  afterAll(async () => {
+    brain = null as any
+  })
+
+  it('from + type array: the second type\'s edge comes back, both orders', async () => {
+    for (const types of [
+      [VerbType.Supports, VerbType.RelatedTo],
+      [VerbType.RelatedTo, VerbType.Supports]
+    ]) {
+      const edges = await brain.related({ from: 'a', type: types })
+      const targets = new Set(edges.map((e) => e.to))
+      expect(targets.has(v5('b')), `types [${types}] missing Supports edge`).toBe(true)
+      expect(targets.has(v5('c')), `types [${types}] missing RelatedTo edge`).toBe(true)
+      expect(targets.has(v5('d'))).toBe(false)
+      expect(edges).toHaveLength(2)
+    }
+  })
+
+  it('a single-element array behaves exactly like the scalar', async () => {
+    const scalar = await brain.related({ from: 'a', type: VerbType.Supports })
+    const array = await brain.related({ from: 'a', type: [VerbType.Supports] })
+    expect(array.map((e) => e.id).sort()).toEqual(scalar.map((e) => e.id).sort())
+    expect(array).toHaveLength(1)
+  })
+
+  it('no duplicate edges when types overlap the same edge set', async () => {
+    const edges = await brain.related({
+      from: 'a',
+      type: [VerbType.Supports, VerbType.RelatedTo, VerbType.Knows]
+    })
+    const ids = edges.map((e) => e.id)
+    expect(new Set(ids).size).toBe(ids.length)
+    expect(edges).toHaveLength(3)
+  })
+
+  it('type-only asks (no anchor) honour the whole array too', async () => {
+    const edges = await brain.related({ type: [VerbType.Supports, VerbType.Knows] })
+    const verbs = new Set(edges.map((e) => e.type))
+    expect(verbs.has(VerbType.Supports)).toBe(true)
+    expect(verbs.has(VerbType.Knows)).toBe(true)
+    expect(edges).toHaveLength(2)
+  })
+
+  it('to + type array: the target side honours every type too', async () => {
+    const edges = await brain.related({ to: 'c', type: [VerbType.RelatedTo, VerbType.Supports] })
+    const froms = new Set(edges.map((e) => e.from))
+    expect(froms.has(v5('a'))).toBe(true)
+    expect(froms.has(v5('b'))).toBe(true)
+    expect(edges).toHaveLength(2)
+  })
+
+  it('pagination stays consistent across the union', async () => {
+    const page1 = await brain.related({ from: 'a', type: [VerbType.Supports, VerbType.RelatedTo, VerbType.Knows], limit: 2 })
+    const page2 = await brain.related({ from: 'a', type: [VerbType.Supports, VerbType.RelatedTo, VerbType.Knows], limit: 2, offset: 2 })
+    const all = [...page1, ...page2].map((e) => e.id)
+    expect(new Set(all).size).toBe(3)
+  })
+})

tests/integration/transact-edge-delete-bigint-aliasing.test.ts

@@ -0,0 +1,184 @@
+/**
+ * @module tests/integration/transact-edge-delete-bigint-aliasing
+ * @description Regression for a fleet-adoption blocker: ANY edge delete
+ * inside `transact()` — a direct unrelate or a noun-remove's cascade —
+ * aborted with the metadata seam's BigInt JSON-guard error on a strict
+ * (native) metadata provider.
+ *
+ * The aliasing chain: `planTxUnrelate`/the remove-cascade pass the SAME verb
+ * object to the graph-retraction op and the metadata-retraction op. The
+ * metadata leg's JSON-safe wrap ran at PLAN time, when the verb was still
+ * clean — so it returned the same reference. At EXECUTE time the graph op
+ * runs first and `resolveVerbEndpointInts` mirrors BigInt
+ * `sourceInt`/`targetInt` onto the shared object (deliberately deferred for
+ * same-batch forward refs — see transact-forward-ref-graph.test.ts); the
+ * metadata op then crossed the seam with the polluted object. Direct
+ * `unrelate()` resolves ints at BUILD time, before its sanitize, which is why
+ * only the transact() shapes ever hit it.
+ *
+ * Fix under pin: the JSON-safe view is taken AT THE CROSSING — inside the
+ * metadata-index operations' execute/rollback — so no plan-vs-execute
+ * ordering can bypass it. The JS baseline index tolerates BigInts (it would
+ * mask the bug), so these pins SPY on the seam and assert what actually
+ * crossed, exactly as a strict native provider would judge it.
+ */
+import { describe, it, expect, beforeEach, afterEach } from 'vitest'
+import * as fs from 'node:fs'
+import * as os from 'node:os'
+import * as path from 'node:path'
+import { Brainy } from '../../src/brainy.js'
+import { NounType, VerbType } from '../../src/types/graphTypes.js'
+import {
+  AddToMetadataIndexOperation,
+  RemoveFromMetadataIndexOperation
+} from '../../src/transaction/operations/index.js'
+
+let seq = 0
+const freshId = (): string =>
+  `00000000-0000-4000-8000-${(++seq).toString(16).padStart(12, '0')}`
+
+/** Top-level BigInt-valued keys of a candidate seam crossing (the guard's law). */
+const bigintKeys = (metadata: unknown): string[] => {
+  if (metadata === null || typeof metadata !== 'object') return []
+  return Object.entries(metadata as Record<string, unknown>)
+    .filter(([, v]) => typeof v === 'bigint')
+    .map(([k]) => k)
+}
+
+describe('transact() edge deletes never carry BigInt across the metadata seam', () => {
+  let dir: string
+  let brain: any
+  let crossings: Array<{ door: string; id: string; keys: string[] }>
+
+  beforeEach(async () => {
+    process.env.BRAINY_DETERMINISTIC_EMBEDDINGS = 'true'
+    dir = fs.mkdtempSync(path.join(os.tmpdir(), 'brainy-tx-bigint-'))
+    brain = new Brainy({
+      requireSubtype: false,
+      storage: { type: 'filesystem', path: dir },
+      dimensions: 384,
+      silent: true
+    })
+    await brain.init()
+
+    // Spy on the seam the way a strict native provider judges it: record the
+    // BigInt-valued top-level keys of every metadata argument that crosses.
+    // The JS baseline index tolerates BigInts, so without this the baseline
+    // run would green a shape the native pair aborts on.
+    crossings = []
+    const index = brain.metadataIndex
+    for (const door of ['addToIndex', 'removeFromIndex'] as const) {
+      const real = index[door].bind(index)
+      index[door] = (id: string, metadata: unknown, ...rest: unknown[]) => {
+        crossings.push({ door, id, keys: bigintKeys(metadata) })
+        return real(id, metadata, ...rest)
+      }
+    }
+  })
+
+  afterEach(async () => {
+    await brain.close()
+    fs.rmSync(dir, { recursive: true, force: true })
+  })
+
+  it('CASE 1 (the fleet repro): relate, then transact([{op: unrelate}])', async () => {
+    const a = await brain.add({ id: freshId(), data: 'a', type: NounType.Thing })
+    const b = await brain.add({ id: freshId(), data: 'b', type: NounType.Thing })
+    const verbId = await brain.relate({ from: a, to: b, type: VerbType.RelatedTo })
+
+    crossings.length = 0
+    await brain.transact([{ op: 'unrelate', id: verbId }])
+
+    const polluted = crossings.filter((c) => c.keys.length > 0)
+    expect(polluted).toEqual([])
+    expect(await brain.storage.getVerb(verbId)).toBeFalsy()
+  })
+
+  it('CASE 2 (the cascade shape): transact([{op: remove}]) cascading edge deletes', async () => {
+    const a = await brain.add({ id: freshId(), data: 'a', type: NounType.Thing })
+    const b = await brain.add({ id: freshId(), data: 'b', type: NounType.Thing })
+    const c = await brain.add({ id: freshId(), data: 'c', type: NounType.Thing })
+    const ab = await brain.relate({ from: a, to: b, type: VerbType.RelatedTo })
+    const ca = await brain.relate({ from: c, to: a, type: VerbType.RelatedTo })
+
+    crossings.length = 0
+    await brain.transact([{ op: 'remove', id: a }])
+
+    const polluted = crossings.filter((c2) => c2.keys.length > 0)
+    expect(polluted).toEqual([])
+    expect(await brain.get(a)).toBeFalsy()
+    expect(await brain.storage.getVerb(ab)).toBeFalsy()
+    expect(await brain.storage.getVerb(ca)).toBeFalsy()
+  })
+
+  it('CASE 3 (one batch, both legs): adds + relate + unrelate of a pre-existing edge', async () => {
+    const a = await brain.add({ id: freshId(), data: 'a', type: NounType.Thing })
+    const b = await brain.add({ id: freshId(), data: 'b', type: NounType.Thing })
+    const old = await brain.relate({ from: a, to: b, type: VerbType.RelatedTo })
+
+    const x = freshId()
+    crossings.length = 0
+    await brain.transact([
+      { op: 'add', id: x, data: 'x', type: NounType.Thing },
+      { op: 'relate', from: a, to: x, type: VerbType.RelatedTo },
+      { op: 'unrelate', id: old }
+    ])
+
+    const polluted = crossings.filter((c) => c.keys.length > 0)
+    expect(polluted).toEqual([])
+    expect(await brain.storage.getVerb(old)).toBeFalsy()
+    const edges = await brain.related({ from: a })
+    expect(edges.length).toBe(1)
+    expect(edges[0].id).not.toBe(old)
+  })
+})
+
+describe('the metadata-index operations sanitize at the crossing, not at construction', () => {
+  /** A strict seam: refuses BigInts exactly as the native provider does. */
+  const strictIndex = () => {
+    const seen: Array<{ door: string; keys: string[] }> = []
+    const judge = (door: string, metadata: unknown) => {
+      const keys = bigintKeys(metadata)
+      seen.push({ door, keys })
+      if (keys.length > 0) {
+        throw new Error(
+          `${door}: the metadata object violates the provider seam's JSON ` +
+            `contract — BigInt at ${keys.join(', ')}.`
+        )
+      }
+    }
+    return {
+      seen,
+      addToIndex: async (_id: string, metadata: unknown) => judge('addToIndex', metadata),
+      removeFromIndex: async (_id: string, metadata: unknown) => judge('removeFromIndex', metadata)
+    }
+  }
+
+  it('RemoveFromMetadataIndexOperation: entity mutated AFTER construction still crosses clean', async () => {
+    const index = strictIndex()
+    const verb: Record<string, unknown> = { id: 'v1', sourceId: 'a', targetId: 'b' }
+    const op = new RemoveFromMetadataIndexOperation(index as any, 'v1', verb, () => 7n)
+
+    // The graph leg's execute-time endpoint resolution, simulated: the shared
+    // object is polluted between plan and execute.
+    verb.sourceInt = 800_000n
+    verb.targetInt = 800_001n
+
+    const rollback = await op.execute()
+    await rollback()
+    expect(index.seen.map((s) => s.keys)).toEqual([[], []])
+  })
+
+  it('AddToMetadataIndexOperation: same law on the add leg and its rollback', async () => {
+    const index = strictIndex()
+    const verb: Record<string, unknown> = { id: 'v2', sourceId: 'a', targetId: 'b' }
+    const op = new AddToMetadataIndexOperation(index as any, 'v2', verb, () => 7n)
+
+    verb.sourceInt = 800_000n
+    verb.targetInt = 800_001n
+
+    const rollback = await op.execute()
+    await rollback()
+    expect(index.seen.map((s) => s.keys)).toEqual([[], []])
+  })
+})