fix(embed): stop worker pool from dispatching new slices after a sub-batch failure

Review finding on #3130: when one sub-batch rejected, the surviving pool
workers kept draining ALL remaining slices in the background after
embedBatch had already rejected — real provider spend post-failure,
onBatchComplete firing after the caller handled the error, and stacked
429 pressure when embedBatchWithBackoff retried while the failed run was
still draining. A shared failed flag now stops further dispatch (in-flight
sibling calls still settle, bounded by concurrency-1) and suppresses
post-failure progress callbacks. Pinned by a new test: 10 slices /
concurrency 2 / first call fails → no calls after rejection, no
completions reported.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/cli.ts

@@ -54,7 +54,7 @@ export function bigintToStringReplacer(_key: string, value: unknown): unknown {
 }
 
 // CLI-only commands that bypass the operation layer
-export const CLI_ONLY = new Set(['init', 'reinit-pglite', 'upgrade', 'post-upgrade', 'check-update', 'integrations', 'publish', 'check-backlinks', 'lint', 'report', 'import', 'export', 'files', 'embed', 'serve', 'call', 'config', 'doctor', 'migrate', 'eval', 'sync', 'extract', 'extract-conversation-facts', 'enrich', 'features', 'autopilot', 'graph-query', 'jobs', 'agent', 'apply-migrations', 'skillpack-check', 'skillpack', 'resolvers', 'integrity', 'repair-jsonb', 'orphans', 'sources', 'mounts', 'dream', 'check-resolvable', 'routing-eval', 'skillify', 'smoke-test', 'providers', 'storage', 'repos', 'code-def', 'code-refs', 'reindex', 'reindex-code', 'reindex-frontmatter', 'code-callers', 'code-callees', 'reconcile-links', 'frontmatter', 'auth', 'friction', 'claw-test', 'book-mirror', 'takes', 'think', 'salience', 'anomalies', 'calibration', 'transcripts', 'models', 'remote', 'recall', 'forget', 'edges-backfill', 'facts', 'cache', 'ze-switch', 'founder', 'brainstorm', 'lsd', 'schema', 'capture', 'onboard', 'conversation-parser', 'status', 'connect', 'skillopt', 'quarantine', 'self-upgrade', 'advisor', 'watch', 'reindex-search-vector']);
+export const CLI_ONLY = new Set(['init', 'reinit-pglite', 'upgrade', 'post-upgrade', 'check-update', 'integrations', 'publish', 'check-backlinks', 'lint', 'report', 'import', 'export', 'files', 'embed', 'serve', 'call', 'config', 'doctor', 'migrate', 'eval', 'sync', 'extract', 'extract-conversation-facts', 'enrich', 'features', 'autopilot', 'graph-query', 'jobs', 'agent', 'apply-migrations', 'skillpack-check', 'skillpack', 'resolvers', 'integrity', 'repair-jsonb', 'orphans', 'sources', 'mounts', 'dream', 'check-resolvable', 'routing-eval', 'skillify', 'smoke-test', 'providers', 'storage', 'repos', 'code-def', 'code-refs', 'reindex', 'reindex-code', 'reindex-frontmatter', 'code-callers', 'code-callees', 'reconcile-links', 'frontmatter', 'auth', 'friction', 'claw-test', 'book-mirror', 'takes', 'think', 'salience', 'anomalies', 'calibration', 'transcripts', 'models', 'remote', 'recall', 'forget', 'edges-backfill', 'cache', 'ze-switch', 'founder', 'brainstorm', 'lsd', 'schema', 'capture', 'onboard', 'conversation-parser', 'status', 'connect', 'skillopt', 'quarantine', 'self-upgrade', 'advisor', 'watch', 'reindex-search-vector']);
 // CLI-only commands whose handlers print their own --help text. These are
 // excluded from the generic short-circuit so detailed per-command and
 // per-subcommand usage stays reachable.
@@ -991,8 +991,6 @@ const THIN_CLIENT_REFUSED_COMMANDS = new Set([
   // hint pointing at the routable MCP tools; per-subcommand splits are
   // a v0.31.x follow-up TODO.
   'takes', 'sources',
-  // #1867: fence-backfill edits local .md fences + stamps the local DB.
-  'facts',
   // v0.32 thin-client routing audit (Codex round 2 findings #2, #4):
   // - `pages` purge-deleted is admin+localOnly (operations.ts:856-864)
   // - `files` list / file_url MCP ops are localOnly (operations.ts:1769-1879)
@@ -1028,7 +1026,6 @@ const THIN_CLIENT_REFUSE_HINTS: Record<string, string> = {
   migrate: "migrate runs on the host's local engine. Run on the host machine.",
   'apply-migrations': 'schema migrations run on the host. SSH and run there.',
   'repair-jsonb': 'repair-jsonb operates on the local DB only.',
-  facts: 'facts fence-backfill edits local entity-page fences. Run on the host machine.',
   integrity: 'integrity scans local files. Run on the host machine.',
   serve: 'serve starts a server. Run on the host, not the thin client.',
   dream: 'dream runs the autopilot cycle on the host. `gbrain remote ping` queues one. (Native `gbrain dream` thin-client routing planned for v0.31.2.)',
@@ -1858,13 +1855,6 @@ async function handleCliOnly(command: string, args: string[]) {
         await runEdgesBackfill(engine, args);
         break;
       }
-      case 'facts': {
-        // #1867 — re-runnable fence-backfill for row_num-NULL legacy fact
-        // rows (idempotent v0_32_2 phase B, exposed as an operator command).
-        const { runFactsCommand } = await import('./commands/facts.ts');
-        await runFactsCommand(engine, args);
-        break;
-      }
       case 'whoknows': {
         // v0.33 (Issue #?): expertise + relationship-proximity routing.
         // MCP op `find_experts` (read-scoped) backs the same code path; CLI
@@ -2345,7 +2335,6 @@ TOOLS
   check-backlinks <check|fix> [dir]  Find/fix missing back-links across brain
   lint <dir|file> [--fix]            Catch LLM artifacts, placeholder dates, bad frontmatter
   orphans [--json] [--count]         Find pages with no inbound wikilinks
-  facts fence-backfill [--dry-run]   Fence legacy fact rows (row_num NULL) onto entity pages
   salience [--days N] [--kind P]     v0.29: pages ranked by emotional + activity salience
   anomalies [--since D] [--sigma N]  v0.29: cohort-based statistical anomalies (tag, type)
   transcripts recent [--days N]      v0.29: recent raw .txt transcripts (local-only)

src/commands/facts.ts

@@ -1,48 +0,0 @@
-/**
- * gbrain facts — fact-store maintenance surface (#1867).
- *
- * `fence-backfill` re-runs the v0_32_2 fence-backfill phase on demand.
- * Remote `extract_facts` deposits that predate the fence-write backstop
- * (and any legacy DB-only insert) leave `row_num IS NULL` rows that the
- * cycle extract_facts guard refuses to reconcile past — previously the
- * only remedy was the one-shot v0_32_2 migration, which the ledger marks
- * complete and never re-runs. The phase is idempotent (only touches
- * `row_num IS NULL` rows), so exposing it as a command is safe to re-run
- * any time the backlog reappears.
- */
-import type { BrainEngine } from '../core/engine.ts';
-import { setCliExitVerdict } from '../core/cli-force-exit.ts';
-import { phaseBFenceFacts } from './migrations/v0_32_2.ts';
-
-function printHelp(): void {
-  process.stderr.write(
-    `Usage: gbrain facts fence-backfill [--dry-run]\n\n` +
-      `Fence-backfill: appends every legacy fact row (row_num IS NULL) to its\n` +
-      `entity page's \`## Facts\` fence and stamps row_num + source_markdown_slug\n` +
-      `back onto the DB row. Idempotent — re-runs only pick up rows still\n` +
-      `missing a fence assignment. Clears the backlog that makes the cycle's\n` +
-      `extract_facts phase skip fence→DB reconciliation.\n\n` +
-      `  --dry-run   report what would be fenced; no FS or DB writes\n`,
-  );
-}
-
-export async function runFactsCommand(engine: BrainEngine, args: string[]): Promise<void> {
-  const sub = args[0];
-  if (!sub || sub === '--help' || sub === '-h') {
-    printHelp();
-    return;
-  }
-  if (sub !== 'fence-backfill') {
-    process.stderr.write(`Unknown facts subcommand: ${sub}\n`);
-    printHelp();
-    setCliExitVerdict(1);
-    return;
-  }
-
-  const dryRun = args.includes('--dry-run');
-  const result = await phaseBFenceFacts(engine, { dryRun });
-  process.stderr.write(
-    `fence-backfill: ${result.status}${result.detail ? ` — ${result.detail}` : ''}\n`,
-  );
-  if (result.status === 'failed') setCliExitVerdict(1);
-}

src/commands/import.ts

@@ -170,10 +170,14 @@ export async function runImport(
   // v0.22.13 (PR #490 Q2): shared parseWorkers helper rejects bad input
   // (--workers 0, -3, "foo") with a loud error instead of silently falling
   // through to 1. Mirrors sync.ts's flag handling.
-  const { parseWorkers } = await import('../core/sync-concurrency.ts');
-  let workerCount: number;
+  const { parseWorkers, autoConcurrency } = await import('../core/sync-concurrency.ts');
+  // #1207: undefined (no --workers flag) defers to autoConcurrency below —
+  // the shared sync/import policy (PGLite → 1, >100 files → 4) — instead of
+  // hardcoding serial. Large Postgres imports stop paying one embedding
+  // round-trip per file in sequence.
+  let workerCount: number | undefined;
   try {
-    workerCount = parseWorkers(workersArg ?? undefined) ?? 1;
+    workerCount = parseWorkers(workersArg ?? undefined);
   } catch (e) {
     console.error(e instanceof Error ? e.message : String(e));
     process.exit(1);
@@ -252,8 +256,9 @@ export async function runImport(
   }
   const files = resumeFilter(allFiles, dir, completed);
 
-  // Determine actual worker count
-  const actualWorkers = workerCount > 1 ? workerCount : 1;
+  // Determine actual worker count. Explicit --workers wins; otherwise the
+  // shared autoConcurrency policy decides from engine kind + file count.
+  const actualWorkers = autoConcurrency(engine, files.length, workerCount);
   if (actualWorkers > 1) {
     console.log(`Using ${actualWorkers} parallel workers`);
   }

src/commands/migrations/v0_32_2.ts

@@ -39,7 +39,6 @@ import type { BrainEngine } from '../../core/engine.ts';
 import { loadConfig, toEngineConfig } from '../../core/config.ts';
 import { createEngine } from '../../core/engine-factory.ts';
 import { upsertFactRow, parseFactsFence } from '../../core/facts-fence.ts';
-import { resolvePageFilePath } from '../../core/markdown.ts';
 
 let testEngineOverride: BrainEngine | null = null;
 export function __setTestEngineOverride(engine: BrainEngine | null): void {
@@ -149,16 +148,9 @@ function isLocalPathDirty(localPath: string): boolean {
   }
 }
 
-/**
- * Exported (not just via `__testing`) because `gbrain facts fence-backfill`
- * (#1867) re-runs this phase on demand: remote `extract_facts` deposits that
- * predate the fence-write backstop leave row_num-NULL rows the cycle guard
- * refuses to reconcile past. The phase is idempotent (only touches
- * `row_num IS NULL` rows), so re-running is always safe.
- */
-export async function phaseBFenceFacts(
+async function phaseBFenceFacts(
   engine: BrainEngine | null,
-  opts: Pick<OrchestratorOpts, 'dryRun'>,
+  opts: OrchestratorOpts,
 ): Promise<OrchestratorPhaseResult> {
   if (opts.dryRun) {
     // Dry-run: report what WOULD happen without touching FS or DB.
@@ -246,11 +238,7 @@ export async function phaseBFenceFacts(
     for (const [key, group] of groups) {
       const [sourceId, entitySlug] = key.split('\0');
       const localPath = localPathById.get(sourceId)!;
-      // resolvePageFilePath, NOT a bare join — non-default sources fence
-      // into `<local_path>/.sources/<id>/<slug>.md`, the same path the
-      // fence-write backstop and put_page write-through compute. A bare
-      // join here diverges fence and DB for non-default sources (#2044).
-      const filePath = resolvePageFilePath(localPath, entitySlug, sourceId);
+      const filePath = join(localPath, `${entitySlug}.md`);
       const tmpPath = `${filePath}.tmp`;
 
       try {
@@ -281,21 +269,6 @@ export async function phaseBFenceFacts(
         const existingFence = parseFactsFence(body);
         const existingKeySet = new Set(existingFence.facts.map(f => `${f.claim}\0${f.source ?? ''}`));
 
-        // Seed appended row_nums from MAX(fence max, DB max) — same #2044
-        // divergence guard as writeFactsToFence. When the fence at the
-        // resolved path is missing/behind but the DB already holds stamped
-        // rows for this slug (legacy wrong-path fence writes), fence-max+1
-        // collides with idx_facts_fence_key on the post-rename UPDATE,
-        // failing the page and leaving fence and DB disagreeing.
-        const dbMaxRows = await engine.executeRaw<{ max: number | string | null }>(
-          `SELECT MAX(row_num) AS max FROM facts
-            WHERE source_id = $1 AND source_markdown_slug = $2`,
-          [sourceId, entitySlug],
-        );
-        const dbMaxRowNum = Number(dbMaxRows[0]?.max ?? 0) || 0;
-        const fenceMaxRowNum = existingFence.facts.reduce((m, f) => Math.max(m, f.rowNum), 0);
-        let nextRowNum = Math.max(fenceMaxRowNum, dbMaxRowNum) + 1;
-
         const assignments: Array<{ id: string; row_num: number }> = [];
         for (const row of group) {
           const key = `${row.fact}\0${row.source ?? ''}`;
@@ -318,7 +291,6 @@ export async function phaseBFenceFacts(
                 .toISOString().slice(0, 10)
             : undefined;
           const { body: updated, rowNum } = upsertFactRow(body, {
-            rowNum:     nextRowNum++,
             claim:      row.fact,
             kind:       row.kind,
             confidence: row.confidence,
@@ -409,7 +381,7 @@ async function phaseCVerify(
     for (const g of groups) {
       const localPath = localPathById.get(g.source_id);
       if (!localPath) continue;
-      const filePath = resolvePageFilePath(localPath, g.source_markdown_slug, g.source_id);
+      const filePath = join(localPath, `${g.source_markdown_slug}.md`);
       if (!existsSync(filePath)) {
         mismatches.push(`${g.source_markdown_slug} (file missing)`);
         continue;

src/core/ai/gateway.ts

@@ -1513,12 +1513,21 @@ export async function embed(texts: string[], opts?: EmbedOpts): Promise<Float32A
 
   const embedding = recipe.touchpoints?.embedding;
   const maxBatchTokens = embedding?.max_batch_tokens;
+  const maxBatchCount = embedding?.max_batch_count;
   const charsPerToken = embedding?.chars_per_token ?? DEFAULT_CHARS_PER_TOKEN;
 
-  // Pre-split is gated on max_batch_tokens. Recipes without it (e.g. OpenAI)
-  // ride the fast path: one embedMany call, no recursion safety net.
-  const batches = maxBatchTokens
-    ? splitByTokenBudget(truncated, Math.floor(maxBatchTokens * effectiveSafetyFactor(recipe)), charsPerToken)
+  // Pre-split is gated on max_batch_tokens / max_batch_count. Recipes with
+  // neither (e.g. OpenAI) ride the fast path: one embedMany call, no
+  // recursion safety net.
+  const batches = (maxBatchTokens || maxBatchCount)
+    ? splitByTokenBudget(
+        truncated,
+        maxBatchTokens
+          ? Math.floor(maxBatchTokens * effectiveSafetyFactor(recipe))
+          : Number.MAX_SAFE_INTEGER,
+        charsPerToken,
+        maxBatchCount,
+      )
     : [truncated];
 
   const allEmbeddings: Float32Array[] = [];
@@ -1568,6 +1577,9 @@ export async function embed(texts: string[], opts?: EmbedOpts): Promise<Float32A
  *   responsible for applying any safety-factor shrink before passing in.
  * @param charsPerToken - Provider-specific character density. Defaults to
  *   `DEFAULT_CHARS_PER_TOKEN` (4) when omitted, matching OpenAI tiktoken.
+ * @param maxBatchCount - #1199: optional cap on INPUTS per sub-batch, for
+ *   providers that reject batches by count (DashScope: 10). When omitted,
+ *   only the token budget governs.
  *
  * @internal exported for tests; not part of the public gateway API.
  */
@@ -1575,15 +1587,17 @@ export function splitByTokenBudget(
   texts: string[],
   budgetTokens: number,
   charsPerToken: number = DEFAULT_CHARS_PER_TOKEN,
+  maxBatchCount?: number,
 ): string[][] {
   const ratio = charsPerToken > 0 ? charsPerToken : DEFAULT_CHARS_PER_TOKEN;
+  const maxCount = maxBatchCount !== undefined && maxBatchCount > 0 ? maxBatchCount : Infinity;
   const batches: string[][] = [];
   let current: string[] = [];
   let currentTokens = 0;
 
   for (const text of texts) {
     const estTokens = Math.ceil(text.length / ratio);
-    if (current.length > 0 && currentTokens + estTokens > budgetTokens) {
+    if (current.length > 0 && (currentTokens + estTokens > budgetTokens || current.length >= maxCount)) {
       batches.push(current);
       current = [];
       currentTokens = 0;
@@ -1609,7 +1623,11 @@ export function isTokenLimitError(err: unknown): boolean {
     /token.*limit.*exceeded/i.test(msg) ||
     // OpenAI embeddings: "Invalid 'input': maximum request size is 300000 tokens per request."
     /maximum request size.*tokens/i.test(msg) ||
-    /max.*tokens.*per.*request/i.test(msg)
+    /max.*tokens.*per.*request/i.test(msg) ||
+    // DashScope: "batch size is invalid, it should not be larger than 10." (#1199)
+    // Count-cap error, but recursive halving shrinks count too, so the same
+    // safety net converges.
+    /batch size is invalid/i.test(msg)
   );
 }
 

src/core/ai/recipes/dashscope.ts

@@ -31,6 +31,10 @@ export const dashscope: Recipe = {
       // path. Conservative declaration so the gateway pre-splits before
       // hitting whatever undocumented server-side limit exists.
       max_batch_tokens: 8192,
+      // #1199: DashScope hard-caps embeddings at 10 inputs per request
+      // ("batch size is invalid, it should not be larger than 10"). The
+      // token budget alone admits far more than 10 short chunks per batch.
+      max_batch_count: 10,
       // text-embedding-v3 mixes English + CJK heavily; the tokenizer is
       // closer to Voyage density than OpenAI tiktoken for CJK-dominant
       // content. Conservative chars_per_token=2 leaves headroom.

src/core/ai/recipes/google.ts

@@ -16,6 +16,15 @@ export const google: Recipe = {
       dims_options: [768, 1536, 3072],
       cost_per_1m_tokens_usd: 0.15,
       price_last_verified: '2026-04-20',
+      // #970: Gemini's documented limits are per-INPUT (2048 tokens,
+      // silently truncated beyond) and per-REQUEST count (batchEmbedContents
+      // caps at 100 inputs). There is no separate per-request token cap, so
+      // the token budget is derived: 100 inputs × 2048 tokens. The count cap
+      // binds first for typical chunk sizes. Do NOT copy the 2048 per-input
+      // limit into max_batch_tokens — that would over-split 50×.
+      max_batch_tokens: 204_800,
+      chars_per_token: 4,
+      max_batch_count: 100,
     },
     expansion: {
       models: ['gemini-2.0-flash', 'gemini-2.0-flash-lite'],

src/core/ai/recipes/index.ts

@@ -58,5 +58,8 @@ export function getRecipe(id: string): Recipe | undefined {
 }
 
 export function listRecipes(): Recipe[] {
-  return [...ALL];
+  // Read the map (not ALL) so there is one source of truth — getRecipe,
+  // model-resolver, and listRecipes all see the same registry, and tests
+  // can inject a synthetic recipe via RECIPES to exercise registry walks.
+  return [...RECIPES.values()];
 }

src/core/ai/types.ts

@@ -46,6 +46,16 @@ export interface EmbeddingTouchpoint {
    * Only consulted when `max_batch_tokens` is also set.
    */
   chars_per_token?: number;
+  /**
+   * #1199: maximum number of INPUTS per embedding request, for providers
+   * that hard-cap batch size by count rather than (or in addition to)
+   * tokens — DashScope text-embedding-v3 rejects batches > 10 with
+   * `InvalidParameter`, Gemini batchEmbedContents caps at 100 requests.
+   * When set, the gateway's pre-split flushes a sub-batch at this count
+   * even if the token budget still has room. Independent of
+   * `max_batch_tokens`; either alone triggers the pre-split.
+   */
+  max_batch_count?: number;
   /**
    * Budget-utilization ceiling in (0, 1]. The gateway pre-splits at
    * `safety_factor × max_batch_tokens` to leave headroom for tokenizer

src/core/cycle/extract-facts.ts

@@ -176,11 +176,9 @@ export async function runExtractFacts(
   if (legacyCount > 0) {
     result.guardTriggered = true;
     result.warnings.push(
-      `extract_facts: ${legacyCount} legacy fact rows pending fence backfill ` +
-      `(row_num IS NULL — v0.31 rows or remote extract_facts deposits that ` +
-      `predate the fence backstop). Run \`gbrain facts fence-backfill\` ` +
-      `(idempotent, re-runnable) before this phase can safely reconcile ` +
-      `fence → DB.`,
+      `extract_facts: ${legacyCount} legacy v0.31 fact rows pending fence backfill. ` +
+      `Run \`gbrain apply-migrations --yes\` to complete v0_32_2 before this phase ` +
+      `can safely reconcile fence → DB.`,
     );
     return result;
   }

src/core/embedding.ts

@@ -79,15 +79,34 @@ export interface EmbedBatchOptions {
    * and amplify rate-limit pressure.
    */
   maxRetries?: number;
+  /**
+   * #1818: bounded parallelism across BATCH_SIZE sub-batches. Defaults to
+   * `GBRAIN_EMBED_BATCH_CONCURRENCY` env, else 4. Results are
+   * index-addressed so output order always matches input order. Set 1 to
+   * force the pre-v0.42 serial dispatch.
+   */
+  concurrency?: number;
 }
 
 /**
  * Embed a batch of texts via the gateway. Sub-batches of 100 so upstream
  * progress callbacks fire incrementally on large imports. The gateway owns
  * adaptive batch splitting and per-recipe token-budget logic; this paginator
- * is purely about progress-callback granularity.
+ * owns progress-callback granularity and (#1818) bounded parallel dispatch
+ * of the sub-batches — the embed-stale.ts worker-pool pattern, scoped down.
  */
 const BATCH_SIZE = 100;
+const DEFAULT_EMBED_BATCH_CONCURRENCY = 4;
+
+function resolveEmbedBatchConcurrency(options: EmbedBatchOptions): number {
+  if (options.concurrency !== undefined) {
+    return Math.max(1, Math.floor(options.concurrency));
+  }
+  const env = Number(process.env.GBRAIN_EMBED_BATCH_CONCURRENCY);
+  if (Number.isFinite(env) && env >= 1) return Math.floor(env);
+  return DEFAULT_EMBED_BATCH_CONCURRENCY;
+}
+
 export async function embedBatch(
   texts: string[],
   options: EmbedBatchOptions = {},
@@ -103,13 +122,44 @@ export async function embedBatch(
   if (texts.length <= BATCH_SIZE && !options.onBatchComplete) {
     return gatewayEmbed(texts, gwOpts);
   }
-  const results: Float32Array[] = [];
+  // #1818: dispatch sub-batches through a bounded worker pool instead of a
+  // serial loop. Results are written into a preallocated index-addressed
+  // array so output order matches input order regardless of completion
+  // order; onBatchComplete reports a monotonic completed-embedding count.
+  const slices: Array<{ start: number; texts: string[] }> = [];
   for (let i = 0; i < texts.length; i += BATCH_SIZE) {
-    const slice = texts.slice(i, i + BATCH_SIZE);
-    const out = await gatewayEmbed(slice, gwOpts);
-    results.push(...out);
-    options.onBatchComplete?.(results.length, texts.length);
+    slices.push({ start: i, texts: texts.slice(i, i + BATCH_SIZE) });
   }
+  const results = new Array<Float32Array>(texts.length);
+  let next = 0;
+  let done = 0;
+  const numWorkers = Math.min(resolveEmbedBatchConcurrency(options), slices.length);
+  // Once any sub-batch fails, `failed` stops the surviving workers from
+  // dispatching FURTHER slices — the whole call is rejecting anyway, so
+  // continuing would burn real provider spend in the background and fire
+  // onBatchComplete after the caller already saw the failure (worst with
+  // embedBatchWithBackoff, whose 429 backoff assumes nothing is in flight).
+  // In-flight sibling calls still run to completion (bounded by numWorkers-1).
+  let failed = false;
+  const worker = async (): Promise<void> => {
+    while (!failed && next < slices.length) {
+      // NOTE: no local aborted-check here — an aborted signal makes the next
+      // gatewayEmbed call throw (SDK-side), which rejects the pool. Returning
+      // silently instead would resolve with holes in `results`.
+      const slice = slices[next++];
+      let out: Float32Array[];
+      try {
+        out = await gatewayEmbed(slice.texts, gwOpts);
+      } catch (err) {
+        failed = true;
+        throw err;
+      }
+      for (let j = 0; j < out.length; j++) results[slice.start + j] = out[j];
+      done += out.length;
+      if (!failed) options.onBatchComplete?.(done, texts.length);
+    }
+  };
+  await Promise.all(Array.from({ length: numWorkers }, () => worker()));
   return results;
 }
 

src/core/engine.ts

@@ -1739,14 +1739,13 @@ export interface BrainEngine {
    * single-row supersede flow because fence reconciliation is the canonical
    * source-of-truth direction, not the consolidator path.
    *
-   * Insertion runs in a single transaction. A collision on the v51
-   * partial UNIQUE index `(source_id, source_markdown_slug, row_num)`
-   * skips ONLY that row (ON CONFLICT DO NOTHING, #2044) — the rest of
-   * the batch still commits, so a redundant deposit against an
-   * already-indexed fence row is idempotent instead of a hard failure.
+   * Insertion is atomic per call: all rows commit in a single transaction
+   * or none commit (the transaction rolls back on any constraint
+   * violation, e.g. the v51 partial UNIQUE index on
+   * `(source_id, source_markdown_slug, row_num)`).
    *
-   * Returns the inserted ids in input-order (colliding rows omitted) so
-   * callers can correlate fence-row → DB-id without a separate lookup.
+   * Returns the inserted ids in input-order so callers can correlate
+   * fence-row → DB-id without a separate lookup.
    */
   insertFacts(
     rows: Array<NewFact & { row_num: number; source_markdown_slug: string }>,

src/core/facts/fence-write.ts

@@ -218,27 +218,11 @@ export async function writeFactsToFence(
       }
 
       // 2. Upsert each fact onto the fence in input order. row_num
-      //    monotonically increases, append-only, seeded from the MAX of
-      //    the fence and the DB index (#2044): when fence and DB have
-      //    diverged (e.g. legacy writes that stamped DB rows against a
-      //    fence at a path this code no longer reads), fence-max+1 can
-      //    collide with an existing DB row_num, tripping
-      //    idx_facts_fence_key and rolling back the whole batch.
-      const dbMaxRows = await engine.executeRaw<{ max: number | string | null }>(
-        `SELECT MAX(row_num) AS max FROM facts
-          WHERE source_id = $1 AND source_markdown_slug = $2`,
-        [target.sourceId, target.slug],
-      );
-      const dbMaxRowNum = Number(dbMaxRows[0]?.max ?? 0) || 0;
-      const fenceMaxRowNum = parseFactsFence(body).facts
-        .reduce((m, f) => Math.max(m, f.rowNum), 0);
-      let nextRowNum = Math.max(fenceMaxRowNum, dbMaxRowNum) + 1;
-
+      //    monotonically increases (max-existing + 1 per call, append-only).
       const assignedRowNums: number[] = [];
       for (const f of facts) {
         const validFromStr = (f.validFrom ?? new Date()).toISOString().slice(0, 10);
         const { body: updated, rowNum } = upsertFactRow(body, {
-          rowNum:      nextRowNum++,
           claim:       f.fact,
           kind:        (f.kind ?? 'fact') as 'fact' | 'event' | 'preference' | 'commitment' | 'belief',
           confidence:  f.confidence ?? 1.0,

src/core/pglite-engine.ts

@@ -4102,13 +4102,11 @@ export class PGLiteEngine implements BrainEngine {
   ): Promise<{ inserted: number; ids: number[] }> {
     if (rows.length === 0) return { inserted: 0, ids: [] };
 
-    // Single transaction; per-row INSERTs (not multi-row VALUES) keep the
-    // embedding-vs-no-embedding branching readable; batch sizes are small
-    // (5-30 rows per page in practice) so the loop overhead is negligible
-    // vs the embedding compute cost. #2044: ON CONFLICT DO NOTHING on the
-    // v51 partial UNIQUE index makes a residual fence/DB row_num collision
-    // skip that row instead of rolling back the whole batch (parity with
-    // postgres-engine.ts).
+    // Single transaction so the v51 partial UNIQUE index can roll back the
+    // whole batch on constraint violation. Per-row INSERTs (not multi-row
+    // VALUES) keep the embedding-vs-no-embedding branching readable; batch
+    // sizes are small (5-30 rows per page in practice) so the loop overhead
+    // is negligible vs the embedding compute cost.
     const ids = await this.db.transaction(async (tx) => {
       const out: number[] = [];
       for (const input of rows) {
@@ -4151,11 +4149,7 @@ export class PGLiteEngine implements BrainEngine {
                  $14, $15,
                  $16, $17, $18, $19,
                  $20
-               )
-               ON CONFLICT (source_id, source_markdown_slug, row_num)
-                 WHERE row_num IS NOT NULL
-                 DO NOTHING
-               RETURNING id`
+               ) RETURNING id`
             : `INSERT INTO facts (
                  source_id, entity_slug, fact, kind, visibility, notability, context,
                  valid_from, valid_until, source, source_session, confidence,
@@ -4169,16 +4163,12 @@ export class PGLiteEngine implements BrainEngine {
                  $15, $16,
                  $17, $18, $19, $20,
                  $21
-               )
-               ON CONFLICT (source_id, source_markdown_slug, row_num)
-                 WHERE row_num IS NOT NULL
-                 DO NOTHING
-               RETURNING id`,
+               ) RETURNING id`,
           embedStr === null
             ? [ctx.source_id, entitySlug, input.fact, kind, visibility, notability, context, validFrom, validUntil, input.source, sourceSession, confidence, embeddedAt, input.row_num, input.source_markdown_slug, claimMetric, claimValue, claimUnit, claimPeriod, eventType]
             : [ctx.source_id, entitySlug, input.fact, kind, visibility, notability, context, validFrom, validUntil, input.source, sourceSession, confidence, embedStr, embeddedAt, input.row_num, input.source_markdown_slug, claimMetric, claimValue, claimUnit, claimPeriod, eventType],
         );
-        if (ins.rows[0]) out.push(ins.rows[0].id);
+        out.push(ins.rows[0].id);
       }
       return out;
     });

src/core/postgres-engine.ts

@@ -4302,12 +4302,10 @@ export class PostgresEngine implements BrainEngine {
     // ONCE per process so the cast matches the actual column type
     // (halfvec vs vector). The probe is cached after first call.
     const castSuffix = await this.resolveFactsEmbeddingCast();
-    // Single transaction; per-row INSERTs (not multi-row VALUES) keep the
-    // embedding-vs-no-embedding branching readable; batch sizes are small
-    // (5-30 rows per page in practice). #2044: ON CONFLICT DO NOTHING on
-    // the v51 partial UNIQUE index makes a residual fence/DB row_num
-    // collision skip that row instead of rolling back the whole batch —
-    // the fence stays system-of-record and reconciliation catches up.
+    // Single transaction so the v51 partial UNIQUE index can roll back
+    // the whole batch on constraint violation. Per-row INSERTs (not
+    // multi-row VALUES) keep the embedding-vs-no-embedding branching
+    // readable; batch sizes are small (5-30 rows per page in practice).
     // No supersede flow in this path — fence reconciliation is the
     // canonical source-of-truth direction, not the consolidator path.
     const ids = await sql.begin(async (tx) => {
@@ -4348,13 +4346,9 @@ export class PostgresEngine implements BrainEngine {
             ${input.row_num}, ${input.source_markdown_slug},
             ${claimMetric}, ${claimValue}, ${claimUnit}, ${claimPeriod},
             ${eventType}
-          )
-          ON CONFLICT (source_id, source_markdown_slug, row_num)
-            WHERE row_num IS NOT NULL
-            DO NOTHING
-          RETURNING id
+          ) RETURNING id
         `;
-        if (ins[0]) out.push(Number(ins[0].id));
+        out.push(Number(ins[0].id));
       }
       return out;
     });

test/ai/adaptive-embed-batch.test.ts

@@ -39,6 +39,8 @@ import {
   __getShrinkStateForTests,
 } from '../../src/core/ai/gateway.ts';
 import { AIConfigError, AITransientError } from '../../src/core/ai/errors.ts';
+import { RECIPES } from '../../src/core/ai/recipes/index.ts';
+import type { Recipe } from '../../src/core/ai/types.ts';
 
 // The last test in this file leaves the gateway configured with a remote
 // provider + fake key and a REAL embed transport. Without a final reset,
@@ -93,6 +95,14 @@ function configureGoogle(): void {
   });
 }
 
+function configureDashscope(): void {
+  configureGateway({
+    embedding_model: 'dashscope:text-embedding-v3',
+    embedding_dimensions: 1024,
+    env: { DASHSCOPE_API_KEY: 'sk-fake' },
+  });
+}
+
 // --------- 1. Pure helpers ---------
 
 describe('splitByTokenBudget (pure helper)', () => {
@@ -149,6 +159,27 @@ describe('splitByTokenBudget (pure helper)', () => {
     expect(splitByTokenBudget(texts, 96_000, 0)).toEqual(splitByTokenBudget(texts, 96_000, 4));
     expect(splitByTokenBudget(texts, 96_000, -1)).toEqual(splitByTokenBudget(texts, 96_000, 4));
   });
+
+  // #1199: count cap for providers that reject batches by input count.
+  test('max_batch_count flushes even when token budget has room', () => {
+    const texts = Array.from({ length: 25 }, (_, i) => `t${i}`);
+    const result = splitByTokenBudget(texts, 1_000_000, 4, 10);
+    expect(result.map(b => b.length)).toEqual([10, 10, 5]);
+    expect(result.flat()).toEqual(texts);
+  });
+
+  test('token budget still governs alongside max_batch_count', () => {
+    const texts = ['a'.repeat(50_000), 'b'.repeat(50_000), 'c'.repeat(50_000)];
+    const result = splitByTokenBudget(texts, 96_000, 1, 10);
+    expect(result).toHaveLength(3);
+  });
+
+  test('undefined / zero / negative max_batch_count is ignored', () => {
+    const texts = Array.from({ length: 25 }, () => 'x');
+    expect(splitByTokenBudget(texts, 1_000_000, 4, undefined)).toHaveLength(1);
+    expect(splitByTokenBudget(texts, 1_000_000, 4, 0)).toHaveLength(1);
+    expect(splitByTokenBudget(texts, 1_000_000, 4, -5)).toHaveLength(1);
+  });
 });
 
 describe('isTokenLimitError (pure helper)', () => {
@@ -179,6 +210,12 @@ describe('isTokenLimitError (pure helper)', () => {
     expect(isTokenLimitError(new Error('Exceeded 300000 max tokens per request'))).toBe(true);
   });
 
+  test('matches DashScope batch-count error (#1199)', () => {
+    expect(isTokenLimitError(new Error(
+      'InvalidParameter: batch size is invalid, it should not be larger than 10.',
+    ))).toBe(true);
+  });
+
   test('does not match unrelated errors', () => {
     expect(isTokenLimitError(new Error('Connection refused'))).toBe(false);
     expect(isTokenLimitError(new Error('Invalid API key'))).toBe(false);
@@ -387,26 +424,92 @@ describe('shrink-on-miss adaptive cache', () => {
   });
 });
 
+// --------- 8. Pre-split count cap through public embed() (#1199 / #970) ---------
+
+describe('embed() pre-split honors max_batch_count', () => {
+  beforeEach(() => resetGateway());
+  afterEach(() => __setEmbedTransportForTests(null));
+
+  test('dashscope never dispatches more than 10 inputs per call (#1199)', async () => {
+    configureDashscope();
+    const stub = mock(async ({ values }: { values: string[] }) => fakeEmbeddings(values, 1024));
+    __setEmbedTransportForTests(stub as any);
+
+    // 25 short texts fit trivially in the 8192-token budget; without the
+    // count cap they'd ship as ONE batch and DashScope would reject it.
+    const texts = Array.from({ length: 25 }, (_, i) => `short-${i}`);
+    const result = await embed(texts);
+
+    expect(result).toHaveLength(25);
+    const callLengths = stub.mock.calls.map(([arg]) => (arg as { values: string[] }).values.length);
+    expect(Math.max(...callLengths)).toBeLessThanOrEqual(10);
+    expect(callLengths.reduce((a, b) => a + b, 0)).toBe(25);
+    // Order preserved across sub-batches.
+    expect((stub.mock.calls[0][0] as { values: string[] }).values[0]).toBe('short-0');
+  });
+
+  test('google pre-splits at 100 inputs per batchEmbedContents call (#970)', async () => {
+    configureGoogle();
+    const stub = mock(async ({ values }: { values: string[] }) => fakeEmbeddings(values, 768));
+    __setEmbedTransportForTests(stub as any);
+
+    const texts = Array.from({ length: 250 }, (_, i) => `g${i}`);
+    const result = await embed(texts);
+
+    expect(result).toHaveLength(250);
+    const callLengths = stub.mock.calls.map(([arg]) => (arg as { values: string[] }).values.length);
+    expect(callLengths).toEqual([100, 100, 50]);
+  });
+});
+
 // --------- 7. Startup warning (D9-B) ---------
 
 describe('startup warning for recipes missing max_batch_tokens', () => {
   beforeEach(() => resetGateway());
 
+  // #970 closed google's missing cap, so no registered recipe is capless
+  // anymore. Inject a synthetic capless recipe to keep the warning path
+  // covered for the NEXT recipe that forgets the field.
+  const caplessRecipe: Recipe = {
+    id: 'capless-test',
+    name: 'Capless Test Provider',
+    tier: 'openai-compat',
+    implementation: 'openai-compatible',
+    base_url_default: 'https://example.invalid/v1',
+    auth_env: { required: [] },
+    touchpoints: {
+      embedding: { models: ['capless-embed-1'], default_dims: 768 },
+    },
+  };
+
+  function configureCapless(): void {
+    configureGateway({
+      embedding_model: 'capless-test:capless-embed-1',
+      embedding_dimensions: 768,
+      env: {},
+    });
+  }
+
   test('configured missing-cap recipe warns once; unrelated recipes stay quiet', () => {
     const warnings: string[] = [];
     const original = console.warn;
     console.warn = (msg: string) => warnings.push(String(msg));
+    RECIPES.set(caplessRecipe.id, caplessRecipe);
     try {
       configureOpenAI();
       expect(warnings.length).toBe(0);
+      // #970 regression: google now declares max_batch_tokens → quiet.
       configureGoogle();
+      expect(warnings.length).toBe(0);
+      configureCapless();
       const firstCallCount = warnings.length;
       // Reconfigure: the warning should NOT re-fire for the same recipes
       // within one process (we already told the operator).
-      configureGoogle();
+      configureCapless();
       expect(warnings.length).toBe(firstCallCount);
     } finally {
       console.warn = original;
+      RECIPES.delete(caplessRecipe.id);
     }
 
     // The warning text should match the documented contract.
@@ -415,11 +518,12 @@ describe('startup warning for recipes missing max_batch_tokens', () => {
     );
     expect(contractMatch.length).toBe(1);
 
-    // Voyage declares max_batch_tokens → suppressed. OpenAI is the
-    // canonical fast-path recipe → also suppressed by id. Both must be
-    // absent from the warnings.
+    // Voyage + google declare max_batch_tokens → suppressed. OpenAI is the
+    // canonical fast-path recipe → also suppressed by id. All must be
+    // absent from the warnings; only the synthetic capless recipe fires.
     expect(warnings.find(w => w.includes('"voyage"'))).toBeUndefined();
     expect(warnings.find(w => w.includes('"openai"'))).toBeUndefined();
-    expect(warnings.find(w => w.includes('"google"'))).toBeDefined();
+    expect(warnings.find(w => w.includes('"google"'))).toBeUndefined();
+    expect(warnings.find(w => w.includes('"capless-test"'))).toBeDefined();
   });
 });

test/ai/no-batch-cap-suppression.serial.test.ts

@@ -52,16 +52,7 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
     }
   });
 
-  test('configureGateway warns for google only when google embedding is configured', () => {
-    warnSpy.mockClear();
-    resetGateway();
-    configureGateway({ env: {} });
-    let messages = warnSpy.mock.calls.map(c => String(c[0] ?? ''));
-    expect(
-      messages.some(m => m.includes('"google"') && m.includes('without max_batch_tokens')),
-      'google should not warn while OpenAI default is configured',
-    ).toBe(false);
-
+  test('configureGateway does NOT warn for google now that it declares batch caps (#970)', () => {
     warnSpy.mockClear();
     resetGateway();
     configureGateway({
@@ -69,11 +60,20 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
       embedding_dimensions: 768,
       env: { GOOGLE_GENERATIVE_AI_API_KEY: 'fake' },
     });
-    messages = warnSpy.mock.calls.map(c => String(c[0] ?? ''));
+    const messages = warnSpy.mock.calls.map(c => String(c[0] ?? ''));
     expect(
       messages.some(m => m.includes('"google"') && m.includes('without max_batch_tokens')),
-      'google should warn when configured because it has fixed-cap models',
-    ).toBe(true);
+      'google declares max_batch_tokens/max_batch_count since #970 — no warning',
+    ).toBe(false);
+  });
+
+  test('google recipe declares its derived batch caps (#970)', () => {
+    const e = getRecipe('google')!.touchpoints.embedding!;
+    // Count cap is the REAL Gemini limit (batchEmbedContents: 100 inputs);
+    // the token budget is derived (100 × 2048 per-input tokens), NOT the
+    // 2048 per-input limit — copying that verbatim would over-split 50×.
+    expect(e.max_batch_count).toBe(100);
+    expect(e.max_batch_tokens).toBe(204_800);
   });
 
   test('every recipe with empty models[] declares user_provided_models OR has openai-fast-path', () => {

test/ai/recipe-dashscope.test.ts

@@ -55,6 +55,11 @@ describe('recipe: dashscope', () => {
     expect(r.touchpoints.embedding!.chars_per_token).toBeGreaterThan(0);
   });
 
+  test('declares max_batch_count: 10 — DashScope rejects larger batches (#1199)', () => {
+    const r = getRecipe('dashscope')!;
+    expect(r.touchpoints.embedding!.max_batch_count).toBe(10);
+  });
+
   test('dimsProviderOptions threads dimensions for text-embedding-v3 (Matryoshka)', async () => {
     // Codex finding #1: DashScope text-embedding-v3 is Matryoshka 64-1024.
     // Without `dimensions` on the wire, user-selected non-default dims are

test/embed-batch-concurrency.test.ts

@@ -0,0 +1,161 @@
+/**
+ * #1818: embedBatch dispatches its 100-input sub-batches through a bounded
+ * worker pool (the embed-stale.ts concurrency pattern) instead of a serial
+ * `for` loop. This file pins:
+ *
+ *   - output order matches input order regardless of completion order
+ *     (index-addressed results)
+ *   - parallelism actually happens (max in-flight > 1) and stays bounded
+ *     (max in-flight <= configured concurrency)
+ *   - concurrency: 1 restores the serial pre-#1818 dispatch
+ *   - GBRAIN_EMBED_BATCH_CONCURRENCY env is honored when the option is unset
+ *   - onBatchComplete reports a monotonic completed count ending at total
+ *
+ * Transport is stubbed via the gateway's __setEmbedTransportForTests seam
+ * (same pattern as test/ai/adaptive-embed-batch.test.ts). OpenAI recipe =
+ * fast path (no pre-split), so each embedBatch sub-batch is exactly one
+ * transport call.
+ */
+
+import { afterAll, afterEach, beforeEach, describe, expect, test } from 'bun:test';
+import {
+  configureGateway,
+  resetGateway,
+  __setEmbedTransportForTests,
+} from '../src/core/ai/gateway.ts';
+import { embedBatch } from '../src/core/embedding.ts';
+import { withEnv } from './helpers/with-env.ts';
+
+const DIMS = 1536;
+
+function configureOpenAI(): void {
+  configureGateway({
+    embedding_model: 'openai:text-embedding-3-large',
+    embedding_dimensions: DIMS,
+    env: { OPENAI_API_KEY: 'sk-fake' },
+  });
+}
+
+/**
+ * Install a transport whose returned embedding encodes the GLOBAL input
+ * index in dim 0 (texts are `t<N>`), so order can be asserted end-to-end.
+ * Tracks the max number of concurrently in-flight transport calls.
+ */
+function installTrackingTransport(delayMs = 5): { maxInFlight: () => number } {
+  let inFlight = 0;
+  let maxInFlight = 0;
+  __setEmbedTransportForTests((async ({ values }: { values: string[] }) => {
+    inFlight++;
+    maxInFlight = Math.max(maxInFlight, inFlight);
+    await new Promise(r => setTimeout(r, delayMs));
+    inFlight--;
+    return {
+      embeddings: values.map(v => {
+        const idx = Number(v.slice(1));
+        return Array.from({ length: DIMS }, (_, j) => (j === 0 ? idx : 0.1));
+      }),
+    };
+  }) as any);
+  return { maxInFlight: () => maxInFlight };
+}
+
+const texts = Array.from({ length: 250 }, (_, i) => `t${i}`);
+
+afterAll(() => resetGateway());
+
+describe('embedBatch bounded parallelism (#1818)', () => {
+  beforeEach(() => {
+    resetGateway();
+    configureOpenAI();
+  });
+  afterEach(() => {
+    __setEmbedTransportForTests(null);
+  });
+
+  test('default pool dispatches sub-batches in parallel, order preserved', async () => {
+    const tracker = installTrackingTransport();
+    const result = await embedBatch(texts, { onBatchComplete: () => {} });
+    expect(result).toHaveLength(250);
+    for (let i = 0; i < 250; i++) {
+      expect(result[i][0]).toBe(i);
+    }
+    // 250 texts → 3 sub-batches; default concurrency 4 → all 3 in flight.
+    expect(tracker.maxInFlight()).toBeGreaterThan(1);
+    expect(tracker.maxInFlight()).toBeLessThanOrEqual(4);
+  });
+
+  test('concurrency: 1 keeps the serial dispatch', async () => {
+    const tracker = installTrackingTransport();
+    const result = await embedBatch(texts, { concurrency: 1, onBatchComplete: () => {} });
+    expect(result).toHaveLength(250);
+    expect(tracker.maxInFlight()).toBe(1);
+  });
+
+  test('GBRAIN_EMBED_BATCH_CONCURRENCY env bounds the pool when option unset', async () => {
+    const tracker = installTrackingTransport();
+    await withEnv({ GBRAIN_EMBED_BATCH_CONCURRENCY: '2' }, async () => {
+      await embedBatch(texts, { onBatchComplete: () => {} });
+    });
+    expect(tracker.maxInFlight()).toBeGreaterThan(1);
+    expect(tracker.maxInFlight()).toBeLessThanOrEqual(2);
+  });
+
+  test('onBatchComplete reports a monotonic count ending at total', async () => {
+    installTrackingTransport();
+    const seen: number[] = [];
+    await embedBatch(texts, {
+      onBatchComplete: (done, total) => {
+        expect(total).toBe(250);
+        seen.push(done);
+      },
+    });
+    expect(seen).toHaveLength(3); // 100 + 100 + 50 sub-batches
+    for (let i = 1; i < seen.length; i++) {
+      expect(seen[i]).toBeGreaterThan(seen[i - 1]);
+    }
+    expect(seen[seen.length - 1]).toBe(250);
+  });
+
+  test('a failing sub-batch rejects the whole call', async () => {
+    let call = 0;
+    __setEmbedTransportForTests((async ({ values }: { values: string[] }) => {
+      call++;
+      if (call === 2) throw new Error('boom');
+      await new Promise(r => setTimeout(r, 2));
+      return { embeddings: values.map(() => Array.from({ length: DIMS }, () => 0.1)) };
+    }) as any);
+    await expect(embedBatch(texts, { onBatchComplete: () => {} })).rejects.toThrow();
+  });
+
+  test('after a failure, surviving workers stop dispatching new slices', async () => {
+    // 1000 texts → 10 slices, concurrency 2. First call fails immediately;
+    // without the `failed` flag the second worker would keep draining all
+    // 10 slices in the background AFTER embedBatch already rejected —
+    // burning provider spend and firing onBatchComplete post-rejection.
+    let calls = 0;
+    const completions: number[] = [];
+    __setEmbedTransportForTests((async ({ values }: { values: string[] }) => {
+      calls++;
+      if (calls === 1) throw new Error('boom');
+      await new Promise(r => setTimeout(r, 5));
+      return { embeddings: values.map(() => Array.from({ length: DIMS }, () => 0.1)) };
+    }) as any);
+    const many = Array.from({ length: 1000 }, (_, i) => `t${i}`);
+    await expect(
+      embedBatch(many, { concurrency: 2, onBatchComplete: d => completions.push(d) }),
+    ).rejects.toThrow('boom');
+    const callsAtRejection = calls;
+    await new Promise(r => setTimeout(r, 50)); // would-be background drain window
+    expect(calls).toBe(callsAtRejection); // no new dispatch after rejection
+    expect(calls).toBeLessThanOrEqual(2); // only the in-flight sibling ran
+    expect(completions).toHaveLength(0); // no progress reported after failure
+  });
+
+  test('single small batch without callback stays on the one-call fast path', async () => {
+    const tracker = installTrackingTransport(1);
+    const result = await embedBatch(['t0', 't1', 't2']);
+    expect(result).toHaveLength(3);
+    expect(result[1][0]).toBe(1);
+    expect(tracker.maxInFlight()).toBe(1);
+  });
+});

test/extract-facts-phase.test.ts

@@ -304,9 +304,7 @@ describe('runExtractFacts — empty-fence guard (Codex R2-#7)', () => {
     expect(r.legacyRowsPending).toBe(1);
     expect(r.factsInserted).toBe(0);
     expect(r.factsDeleted).toBe(0);
-    // #1867: the remedy hint points at the re-runnable backfill command,
-    // not the one-shot v0_32_2 migration (which the ledger never re-runs).
-    expect(r.warnings.some(w => w.includes('gbrain facts fence-backfill'))).toBe(true);
+    expect(r.warnings.some(w => w.includes('apply-migrations'))).toBe(true);
 
     // Legacy row was NOT touched.
     // eslint-disable-next-line @typescript-eslint/no-explicit-any

test/facts-fence-backfill-command.test.ts

@@ -1,133 +0,0 @@
-/**
- * #1867 — `gbrain facts fence-backfill` command tests.
- *
- * The command re-runs the (idempotent) v0_32_2 phase B on demand so
- * row_num-NULL backlogs — remote extract_facts deposits that predate
- * the fence-write backstop — can be cleared without re-running the
- * one-shot migration. Real PGLite + real tempdir filesystem.
- */
-
-import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:test';
-import { mkdtempSync, rmSync, existsSync, readFileSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { runFactsCommand } from '../src/commands/facts.ts';
-import { phaseBFenceFacts } from '../src/commands/migrations/v0_32_2.ts';
-
-let engine: PGLiteEngine;
-let brainDir: string;
-
-beforeAll(async () => {
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-});
-
-afterAll(async () => {
-  await engine.disconnect();
-  try {
-    if (brainDir) rmSync(brainDir, { recursive: true, force: true });
-  } catch { /* best-effort */ }
-});
-
-beforeEach(async () => {
-  brainDir = mkdtempSync(join(tmpdir(), 'facts-backfill-cmd-test-'));
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query('DELETE FROM facts');
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query(
-    `UPDATE sources SET local_path = $1 WHERE id = 'default'`,
-    [brainDir],
-  );
-});
-
-async function seedLegacyFact(fact: string): Promise<void> {
-  // The row_num-NULL shape a remote extract_facts deposit leaves behind
-  // when it lands via the legacy DB-only path.
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query(
-    `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                        valid_from, source, confidence)
-     VALUES ('default', 'people/alice', $1, 'fact', 'private', 'medium',
-             now(), 'mcp:extract_facts', 1.0)`,
-    [fact],
-  );
-}
-
-describe('gbrain facts fence-backfill', () => {
-  test('fences row_num-NULL rows and stamps the DB', async () => {
-    await seedLegacyFact('Deposited remotely');
-
-    await runFactsCommand(engine, ['fence-backfill']);
-
-    // The fence exists on disk with the claim.
-    const filePath = join(brainDir, 'people/alice.md');
-    expect(existsSync(filePath)).toBe(true);
-    expect(readFileSync(filePath, 'utf-8')).toContain('Deposited remotely');
-
-    // The backlog is cleared: no row_num-NULL rows remain.
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num, source_markdown_slug FROM facts',
-    );
-    expect(rows.rows).toHaveLength(1);
-    expect(rows.rows[0].row_num).toBe(1);
-    expect(rows.rows[0].source_markdown_slug).toBe('people/alice');
-  });
-
-  test('re-run is a no-op (idempotent)', async () => {
-    await seedLegacyFact('Deposited remotely');
-    await runFactsCommand(engine, ['fence-backfill']);
-    await runFactsCommand(engine, ['fence-backfill']);
-
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query('SELECT id FROM facts');
-    expect(rows.rows).toHaveLength(1);
-    const body = readFileSync(join(brainDir, 'people/alice.md'), 'utf-8');
-    expect(body.match(/Deposited remotely/g)).toHaveLength(1);
-  });
-
-  test('--dry-run reports without writing', async () => {
-    await seedLegacyFact('Deposited remotely');
-    await runFactsCommand(engine, ['fence-backfill', '--dry-run']);
-
-    expect(existsSync(join(brainDir, 'people/alice.md'))).toBe(false);
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num FROM facts',
-    );
-    expect(rows.rows[0].row_num).toBeNull();
-  });
-
-  test('diverged page: appends past the DB row_num max instead of colliding (#2044 class)', async () => {
-    // The #2044 divergence shape the backfill must survive: the DB already
-    // holds stamped rows 1..3 for the slug (legacy wrong-path fence write),
-    // but the fence at the resolved path is missing. Fence-max+1 (= 1) would
-    // collide with the stamped rows on the post-rename UPDATE, failing the
-    // page and leaving the renamed fence disagreeing with the DB.
-    for (let n = 1; n <= 3; n++) {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      await (engine as any).db.query(
-        `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                            valid_from, source, confidence, row_num, source_markdown_slug)
-         VALUES ('default', 'people/alice', $1, 'fact', 'private', 'medium',
-                 now(), 'mcp:extract_facts', 1.0, $2, 'people/alice')`,
-        [`stamped ${n}`, n],
-      );
-    }
-    await seedLegacyFact('Deposited remotely');
-
-    const result = await phaseBFenceFacts(engine, { dryRun: false });
-    expect(result.status).toBe('complete');
-
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      `SELECT row_num FROM facts WHERE fact = 'Deposited remotely'`,
-    );
-    expect(rows.rows[0].row_num).toBe(4);
-    const body = readFileSync(join(brainDir, 'people/alice.md'), 'utf-8');
-    expect(body).toContain('| 4 | Deposited remotely |');
-  });
-});

test/fence-write.test.ts

@@ -292,49 +292,6 @@ describe('lookupSourceLocalPath', () => {
   });
 });
 
-describe('writeFactsToFence — fence/DB divergence (#2044)', () => {
-  test('seeds row_num past the DB max when the fence lags the DB', async () => {
-    // Simulate the divergence class from #2044: DB rows were stamped with
-    // row_nums against a fence written at a path this code no longer reads
-    // (e.g. the pre-"Local patch 2026-06-11" wrong-path writes). The page
-    // on disk has NO fence, but the DB already holds row_num 1..3 for the
-    // slug. Pre-fix, the next deposit re-assigned row_num=1 from fence
-    // text alone and the whole insertFacts batch failed on
-    // idx_facts_fence_key.
-    for (let n = 1; n <= 3; n++) {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      await (engine as any).db.query(
-        `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                            valid_from, source, confidence, row_num, source_markdown_slug)
-         VALUES ('default', 'people/dana', $1, 'fact', 'private', 'medium',
-                 now(), 'mcp:extract_facts', 1.0, $2, 'people/dana')`,
-        [`old claim ${n}`, n],
-      );
-    }
-
-    const result = await writeFactsToFence(
-      engine,
-      { sourceId: 'default', localPath: brainDir, slug: 'people/dana' },
-      [baseInput({ fact: 'second deposit' })],
-    );
-
-    expect(result.fenceWriteFailed).toBeUndefined();
-    expect(result.inserted).toBe(1);
-
-    // The new row landed PAST the DB max, not at fence-max+1 (= 1).
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num FROM facts WHERE id = $1',
-      [result.ids[0]],
-    );
-    expect(rows.rows[0].row_num).toBe(4);
-
-    // And the on-disk fence carries the same row_num — fence and DB agree.
-    const body = readFileSync(join(brainDir, 'people/dana.md'), 'utf-8');
-    expect(body).toContain('| 4 | second deposit |');
-  });
-});
-
 // Cleanup any leftover tempdirs after the whole suite.
 afterAll(() => {
   // No-op: each test cleaned up via the beforeEach; this is a safety net.

test/helpers/legacy-embedding-preload.ts

@@ -19,7 +19,7 @@
  * overwrites this preload.
  */
 import { configureGateway, getEmbeddingDimensions } from '../../src/core/ai/gateway.ts';
-import { beforeEach } from 'bun:test';
+import { afterEach, beforeEach } from 'bun:test';
 
 const LEGACY_CONFIG = {
   embedding_model: 'openai:text-embedding-3-large',
@@ -52,7 +52,7 @@ applyLegacy();
 //   2. file-local beforeAll → may overwrite to ZE/1280
 // Since beforeAll runs once per file BEFORE the first beforeEach,
 // file-local beforeAll wins for that file's tests. ✓
-beforeEach(() => {
+function applyLegacyIfEmpty() {
   try {
     // Only re-apply if the gateway was reset (or never configured).
     // Tests that explicitly configured a different model in their
@@ -62,4 +62,28 @@ beforeEach(() => {
   } catch {
     applyLegacy();
   }
-});
+}
+
+beforeEach(applyLegacyIfEmpty);
+
+// PR #3130 shard-order fix: beforeEach alone leaves ONE window open — a file
+// whose LAST afterEach calls resetGateway() poisons the NEXT file's
+// beforeAll, which runs BEFORE any beforeEach fires. A beforeAll there that
+// does engine.initSchema() then sizes the embedding column from the gateway
+// DEFAULTS (zembed-1/1280d) instead of the pinned legacy 1536, and every
+// 1536-d Float32Array fixture in that file dies with
+// "expected 1280 dimensions, not 1536". Which file pair collides is a
+// function of shard composition, so adding/removing ANY test file can
+// surface it (that is exactly how it bit shard 9).
+//
+// Preload hooks are registered before any file-local hooks, and bun runs
+// after-hooks inside-out (file-local afterEach first, then this one), so
+// this repairs the empty slot immediately after the poisoning reset —
+// before the next file's beforeAll can observe it.
+//
+// Known remaining window: a file whose afterAll() resets the gateway (no
+// hook runs between its afterAll and the next file's beforeAll). Files
+// that reset in afterAll and can precede a schema-creating file should
+// re-apply their own config, or the victim file should configureGateway()
+// explicitly in its beforeAll.
+afterEach(applyLegacyIfEmpty);

test/import-default-workers.test.ts

@@ -0,0 +1,69 @@
+/**
+ * #1207: `gbrain import` without `--workers` used to hardcode workerCount=1,
+ * so a large Postgres import paid one serial embedding round-trip per file.
+ * runImport now routes the default through the shared autoConcurrency policy
+ * (PGLite → 1, >100 files on Postgres → DEFAULT_PARALLEL_WORKERS), while an
+ * explicit `--workers N` still wins.
+ *
+ * The engine here is a minimal postgres-kind stub with no database_url in
+ * config — runImport's parallel branch then falls back to serial processing
+ * (its PR #490 guard) but the WORKER-COUNT DECISION (the thing #1207 fixes)
+ * is still observable via the "Using N parallel workers" log line. Per-file
+ * imports fail against the stub engine and are swallowed by runImport's
+ * per-file catch; that's fine — this test pins the policy, not the import.
+ */
+
+import { afterEach, beforeEach, describe, expect, test } from 'bun:test';
+import { mkdtempSync, writeFileSync, mkdirSync, rmSync, realpathSync } from 'fs';
+import { tmpdir } from 'os';
+import { join } from 'path';
+import { withEnv } from './helpers/with-env.ts';
+import { runImport } from '../src/commands/import.ts';
+
+const fakePostgresEngine = {
+  kind: 'postgres',
+  executeRaw: async () => [],
+  logIngest: async () => {},
+  setConfig: async () => {},
+  getConfig: async () => null,
+} as any;
+
+let workspace: string;
+let brainDir: string;
+let logs: string[];
+const realLog = console.log;
+
+beforeEach(() => {
+  workspace = mkdtempSync(join(tmpdir(), 'gbrain-import-workers-home-'));
+  mkdirSync(join(workspace, '.gbrain'), { recursive: true });
+  brainDir = realpathSync(mkdtempSync(join(tmpdir(), 'gbrain-import-workers-brain-')));
+  // 101 files: one past AUTO_CONCURRENCY_FILE_THRESHOLD (100).
+  for (let i = 0; i < 101; i++) {
+    writeFileSync(join(brainDir, `page-${i}.md`), `# Page ${i}\n\nbody ${i}\n`);
+  }
+  logs = [];
+  console.log = (msg?: unknown) => logs.push(String(msg));
+});
+
+afterEach(() => {
+  console.log = realLog;
+  rmSync(workspace, { recursive: true, force: true });
+  rmSync(brainDir, { recursive: true, force: true });
+});
+
+describe('import default worker count (#1207)', () => {
+  test('no --workers flag → autoConcurrency picks 4 for >100 files on Postgres', async () => {
+    await withEnv({ GBRAIN_HOME: join(workspace, '.gbrain'), GBRAIN_SOURCE: undefined }, async () => {
+      await runImport(fakePostgresEngine, [brainDir, '--no-embed'], { sourceId: 'default' });
+    });
+    expect(logs.some(l => l.includes('Using 4 parallel workers'))).toBe(true);
+  });
+
+  test('explicit --workers 2 still wins over the auto policy', async () => {
+    await withEnv({ GBRAIN_HOME: join(workspace, '.gbrain'), GBRAIN_SOURCE: undefined }, async () => {
+      await runImport(fakePostgresEngine, [brainDir, '--no-embed', '--workers', '2'], { sourceId: 'default' });
+    });
+    expect(logs.some(l => l.includes('Using 2 parallel workers'))).toBe(true);
+    expect(logs.some(l => l.includes('Using 4 parallel workers'))).toBe(false);
+  });
+});

test/insert-facts-batch.test.ts

@@ -5,7 +5,7 @@
  *   - Batch insert N rows persists row_num + source_markdown_slug
  *   - Empty batch is a no-op
  *   - Returns ids in input-order
- *   - v51 partial UNIQUE collision skips only the colliding row (#2044)
+ *   - v51 partial UNIQUE index rolls back the whole batch on a collision
  *   - deleteFactsForPage scopes by (source_id, source_markdown_slug);
  *     never touches other pages or pre-v51 NULL-source_markdown_slug rows
  *   - deleteFactsForPage on an empty page returns deleted:0 (idempotent)
@@ -135,29 +135,30 @@ describe('engine.insertFacts — batch insert', () => {
     });
   });
 
-  test('v51 partial UNIQUE collision skips ONLY the colliding row (#2044)', async () => {
+  test('v51 partial UNIQUE index rolls back the whole batch on collision', async () => {
     // Seed row #1 first.
     await engine.insertFacts([fixtureFact(1, { fact: 'seeded' })], { source_id: 'default' });
 
-    // Batch-insert rows that include a colliding row_num=1. Pre-#2044 this
-    // threw and rolled back the whole batch, making a second remote
-    // extract_facts deposit to an already-fenced page a hard failure. Now
-    // ON CONFLICT DO NOTHING skips the colliding row and keeps the rest.
-    const result = await engine.insertFacts(
-      [
-        fixtureFact(2, { fact: 'second' }),
-        fixtureFact(1, { fact: 'collides' }),  // row_num=1 on same (source_id, source_markdown_slug)
-        fixtureFact(3, { fact: 'third' }),
-      ],
-      { source_id: 'default' },
-    );
-    expect(result.inserted).toBe(2);
-    expect(result.ids).toHaveLength(2);
+    // Now try to batch-insert rows that include a colliding row_num=1.
+    let threw = false;
+    try {
+      await engine.insertFacts(
+        [
+          fixtureFact(2, { fact: 'second' }),
+          fixtureFact(1, { fact: 'collides' }),  // row_num=1 on same (source_id, source_markdown_slug)
+          fixtureFact(3, { fact: 'third' }),
+        ],
+        { source_id: 'default' },
+      );
+    } catch {
+      threw = true;
+    }
+    expect(threw).toBe(true);
 
-    // The seeded row survives untouched; the colliding claim is skipped.
+    // Verify the transaction rolled back — only the seeded row should remain.
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     const rows = await (engine as any).db.query('SELECT fact FROM facts ORDER BY id');
-    expect(rows.rows.map((r: { fact: string }) => r.fact)).toEqual(['seeded', 'second', 'third']);
+    expect(rows.rows.map((r: { fact: string }) => r.fact)).toEqual(['seeded']);
   });
 
   test('different source_markdown_slug values DO NOT collide on the same row_num', async () => {