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>

skills/article-enrichment/SKILL.md

@@ -85,10 +85,6 @@ gbrain get media/articles/<slug>
 
 # 5. Cross-link entities
 #    For every person/company mentioned, add a timeline back-link.
-#    Mark NEW pages minted for article subjects the reader only reads ABOUT
-#    (not personal contacts) as reference: `gbrain reference <slug>`
-#    (or reference: true in frontmatter). Exempts them from coverage nudges;
-#    they stay searchable. Default for real contacts: do NOT set it.
 ```
 
 ## Quality bar

skills/book-mirror/SKILL.md

@@ -273,13 +273,6 @@ Cross-link entities mentioned in the analysis:
 - For every person the right column references with a brain page, add a
   back-link from `people/<slug>` to the new `media/books/<slug>-personalized`
   page (per `conventions/quality.md` Iron Law).
-- **Mark book figures as reference entities.** Any NEW person/company page you
-  mint for a figure from the book (an author, a historical figure, a company
-  the book discusses) — someone the reader reads ABOUT but doesn't personally
-  interact with — should be flagged: `gbrain reference <slug>` (or
-  `reference: true` in frontmatter). They stay fully searchable/linkable but
-  are exempt from coverage nudges (timeline/links). Skip this for anyone the
-  reader actually knows. Default for normal contacts: do NOT set it.
 
 ## Quality bar (the bar)
 

skills/conventions/reference-entities.md

@@ -1,62 +0,0 @@
-# Convention: Reference entities (canon figures)
-
-Some person/company pages are people/orgs the user **reads about** but does not
-personally interact with — a book's author, a historical figure, a company an
-article discusses (Andy Grove, Kleiner Perkins, Intel…). They are real
-knowledge, worth a page, but they have **no dated history in the user's own
-life**, so the entity coverage metrics (`timeline_coverage`,
-`entity_link_coverage`) flag them as permanently incomplete with no honest fix.
-
-The `reference: true` frontmatter flag resolves this.
-
-## What it does
-
-- A page with `reference: true` is **exempt from the entity coverage metrics
-  only** (`timeline_coverage`, `entity_link_coverage`, and their onboard
-  nudges).
-- It keeps its real `type` (`person` / `company`), so it stays **fully
-  searchable, enrichable, linkable, and edge-resolvable**. NOTHING about
-  retrieval changes — this is the whole reason it's a flag, not a new `type`.
-- It is **opt-in**. Absent / `false` / anything-but-`true` = a normal entity
-  that DOES count toward coverage. **This is the default — do not set it on real
-  contacts.**
-
-## When to set it
-
-Set `reference: true` when the entity is a figure/org the user reads ABOUT, not
-someone they deal with:
-
-- authors and figures discussed in a book (book-mirror) or article
-  (article-enrichment)
-- historical / canon figures imported as reference knowledge
-- companies named only as examples in source material
-
-Do NOT set it for people the user actually meets, emails, or works with — those
-are normal entities whose missing timeline/links is a real, actionable gap.
-
-## How to set it
-
-```bash
-gbrain reference <slug>            # mark as reference
-gbrain reference <slug> --unset    # back to a normal entity
-```
-
-The command writes the flag to BOTH the markdown frontmatter (durable; survives
-re-ingest / engine rebuild — markdown is the source of truth) AND the engine
-JSONB (so coverage reflects it immediately, no re-sync). It's idempotent. You
-can also hand-edit frontmatter (`reference: true`) and re-ingest.
-
-## Why a flag, not a type
-
-A new `type: reference-person` would drop the page out of every
-`type IN ('person','company')` filter — search, enrichment, whoknows, link
-inference — so you'd lose the figure everywhere, not just the metric. The flag
-narrows the change to exactly the coverage denominators and nothing else.
-
-## Implementation
-
-`src/core/reference-flag.ts` — `referenceExclusionSql(alias?)` is the single
-source of truth for the predicate `(frontmatter->>'reference') IS DISTINCT FROM
-'true'`, ANDed into both numerator and denominator at every coverage site
-(getHealth in both engines, onboard/checks.ts, init-nudge.ts). Backed by the GIN
-index on `pages.frontmatter`.

skills/enrich/SKILL.md

@@ -268,18 +268,6 @@ Active items, pending decisions, things to track.
 - **YYYY-MM-DD** | Event description [Source: ...]
 ```
 
-### Reference entities (canon figures)
-
-If the entity is someone/something the user reads ABOUT but does not personally
-interact with — a book author, a historical figure, a company discussed in an
-article — mark the page as reference: `gbrain reference <slug>` (or
-`reference: true` in frontmatter). Reference pages keep their `person`/`company`
-type and stay fully searchable, enrichable, and linkable; they are only exempt
-from the entity coverage nudges (timeline/links) that don't apply to figures
-with no dated history in the user's own life. **Default: do NOT set it** — real
-people and companies the user deals with are normal entities. Full convention:
-`conventions/reference-entities.md`.
-
 ### Step 7: Cross-reference
 
 - Update company pages from person enrichment (and vice versa)

skills/maintain/SKILL.md

@@ -270,11 +270,6 @@ Populate them periodically or after major imports:
 - `gbrain stats` — verify `link_count > 0` and `timeline_entry_count > 0` after extraction.
 - `gbrain health` — review `link_coverage` and `timeline_coverage` percentages
   on entity pages (person/company). Below 50% means more extraction is needed.
-  Note: pages flagged `reference: true` (canon/reference figures the user only
-  reads about) are EXEMPT from these two metrics — if coverage looks stuck
-  because of book/article-imported figures with no real history, mark them with
-  `gbrain reference <slug>` rather than chasing the percentage. See
-  `conventions/reference-entities.md`.
 
 Available link types (use with `gbrain graph-query --type`):
 `attended`, `works_at`, `invested_in`, `founded`, `advises`, `mentions`, `source`.

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', 'reference', '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']);
+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.
@@ -1652,11 +1652,6 @@ async function handleCliOnly(command: string, args: string[]) {
         await runFiles(engine, args);
         break;
       }
-      case 'reference': {
-        const { runReference } = await import('./commands/reference.ts');
-        await runReference(engine, args);
-        break;
-      }
       case 'embed': {
         const { runEmbed } = await import('./commands/embed.ts');
         await runEmbed(engine, args);

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/reference.ts

@@ -1,130 +0,0 @@
-// gbrain reference <slug> [--unset] [--brain <dir>] [--json]
-//
-// Mark (or unmark) a page as a reference-only entity. A reference page keeps its
-// real type (person/company) — fully searchable/enrichable/linkable — but is
-// exempt from the entity coverage metrics (timeline_coverage,
-// entity_link_coverage). See src/core/reference-flag.ts for the rationale.
-//
-// Durability: the flag is written to BOTH the markdown frontmatter (source of
-// truth; survives re-ingest / engine rebuild) AND the engine `pages.frontmatter`
-// JSONB (so the metric reflects it immediately, no re-sync needed).
-
-import { existsSync, readFileSync, writeFileSync } from 'node:fs';
-import { isAbsolute, join, resolve } from 'node:path';
-import { setCliExitVerdict } from '../core/cli-force-exit.ts';
-import type { BrainEngine } from '../core/engine.ts';
-import { REFERENCE_FRONTMATTER_KEY } from '../core/reference-flag.ts';
-import { resolveSourceId } from '../core/source-resolver.ts';
-
-/** Insert/replace/remove `reference: true` in a markdown frontmatter block.
- *  Minimal-diff: only the one line changes; key order is otherwise preserved. */
-export function applyReferenceFrontmatter(content: string, on: boolean): string {
-  const keyLine = `${REFERENCE_FRONTMATTER_KEY}: true`;
-  const block = content.match(/^---\n([\s\S]*?)\n---/);
-
-  if (!block) {
-    // No frontmatter. Nothing to remove; if setting, prepend a block.
-    if (!on) return content;
-    return `---\n${keyLine}\n---\n\n${content}`;
-  }
-
-  let fm = block[1];
-  const hasKey = new RegExp(`^${REFERENCE_FRONTMATTER_KEY}:.*$`, 'm').test(fm);
-
-  if (on) {
-    fm = hasKey
-      ? fm.replace(new RegExp(`^${REFERENCE_FRONTMATTER_KEY}:.*$`, 'm'), keyLine)
-      : `${fm}\n${keyLine}`;
-  } else {
-    if (!hasKey) return content;
-    fm = fm.replace(new RegExp(`^${REFERENCE_FRONTMATTER_KEY}:.*$\\n?`, 'm'), '');
-  }
-
-  // Function replacement so YAML chars ($, & …) in fm aren't treated as
-  // replacement patterns.
-  return content.replace(/^---\n[\s\S]*?\n---/, () => `---\n${fm}\n---`);
-}
-
-function parseArgs(args: string[]): { slug?: string; unset: boolean; json: boolean; brain?: string; source?: string } {
-  let slug: string | undefined;
-  let unset = false;
-  let json = false;
-  let brain: string | undefined;
-  let source: string | undefined;
-  for (let i = 0; i < args.length; i++) {
-    const a = args[i];
-    if (a === '--unset') unset = true;
-    else if (a === '--json') json = true;
-    else if (a === '--brain' || a === '--dir') brain = args[++i];
-    else if (a === '--source') source = args[++i];
-    else if (!a.startsWith('--') && !slug) slug = a;
-  }
-  return { slug, unset, json, brain, source };
-}
-
-async function resolveBrainDir(engine: BrainEngine, explicit?: string): Promise<string | null> {
-  if (explicit) return resolve(explicit);
-  const configured = await engine.getConfig('sync.repo_path');
-  if (configured && existsSync(configured)) return resolve(configured);
-  return null;
-}
-
-export async function runReference(engine: BrainEngine, args: string[]): Promise<void> {
-  const { slug, unset, json, brain, source } = parseArgs(args);
-  if (!slug) {
-    console.error('Usage: gbrain reference <slug> [--unset] [--source <id>] [--brain <dir>] [--json]');
-    setCliExitVerdict(2);
-    return;
-  }
-
-  const brainDir = await resolveBrainDir(engine, brain);
-  if (!brainDir) {
-    console.error('reference: could not resolve brain dir. Pass --brain <dir> or set sync.repo_path.');
-    setCliExitVerdict(1);
-    return;
-  }
-
-  const rel = slug.endsWith('.md') ? slug : `${slug}.md`;
-  const filePath = isAbsolute(rel) ? rel : join(brainDir, rel);
-  if (!existsSync(filePath)) {
-    console.error(`reference: page not found on disk: ${filePath}`);
-    setCliExitVerdict(1);
-    return;
-  }
-
-  // 1) Durable: edit the markdown frontmatter.
-  const before = readFileSync(filePath, 'utf8');
-  const after = applyReferenceFrontmatter(before, !unset);
-  const fileChanged = after !== before;
-  if (fileChanged) writeFileSync(filePath, after, 'utf8');
-
-  // 2) Immediate: update the engine frontmatter JSONB so the metric reflects it
-  //    without waiting for a re-sync. Scoped to (source_id, slug) — slug is only
-  //    unique per source, so a bare-slug UPDATE would clobber same-named pages
-  //    in every other source. Resolve the source from the brain dir (tier-4
-  //    local_path match) unless --source / GBRAIN_SOURCE overrides.
-  const cleanSlug = slug.replace(/\.md$/, '');
-  const sourceId = await resolveSourceId(engine, source ?? null, brainDir);
-  if (unset) {
-    await engine.executeRaw(
-      `UPDATE pages SET frontmatter = frontmatter - '${REFERENCE_FRONTMATTER_KEY}' WHERE slug = $1 AND source_id = $2`,
-      [cleanSlug, sourceId],
-    );
-  } else {
-    await engine.executeRaw(
-      `UPDATE pages SET frontmatter = jsonb_set(COALESCE(frontmatter, '{}'::jsonb), '{${REFERENCE_FRONTMATTER_KEY}}', 'true'::jsonb) WHERE slug = $1 AND source_id = $2`,
-      [cleanSlug, sourceId],
-    );
-  }
-
-  const result = { slug: cleanSlug, source_id: sourceId, reference: !unset, file_changed: fileChanged, file: filePath };
-  if (json) {
-    console.log(JSON.stringify(result));
-  } else {
-    const verb = unset ? 'unmarked' : 'marked';
-    console.log(`${verb} ${cleanSlug} as reference=${!unset}${fileChanged ? '' : ' (frontmatter already current)'}`);
-    if (!unset) {
-      console.log('  → exempt from timeline_coverage / entity_link_coverage; still fully searchable & linkable.');
-    }
-  }
-}

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/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/onboard/checks.ts

@@ -18,7 +18,6 @@
 import type { BrainEngine } from '../engine.ts';
 import type { RemediationStep } from '../remediation-step.ts';
 import { makeRemediationStep } from '../remediation-step.ts';
-import { referenceExclusionSql } from '../reference-flag.ts';
 
 /** Shared shape returned by all four checks. */
 export interface OnboardCheckResult {
@@ -121,8 +120,7 @@ export async function checkEntityLinkCoverage(
     engine,
     `SELECT COUNT(*) AS count FROM pages
        WHERE type IN ('person', 'company', 'organization', 'entity')
-         AND deleted_at IS NULL
-         AND ${referenceExclusionSql()}`,
+         AND deleted_at IS NULL`,
   );
 
   if (totalEntities === 0) {
@@ -146,7 +144,6 @@ export async function checkEntityLinkCoverage(
        SELECT p.id FROM pages p ${sampleClause}
        WHERE p.type IN ('person', 'company', 'organization', 'entity')
          AND p.deleted_at IS NULL
-         AND ${referenceExclusionSql('p')}
          AND EXISTS (SELECT 1 FROM links l WHERE l.to_page_id = p.id)
      ) sub`,
   );
@@ -218,8 +215,7 @@ export async function checkTimelineCoverage(
     engine,
     `SELECT COUNT(*) AS count FROM pages
        WHERE type IN ('person', 'company', 'organization', 'entity')
-         AND deleted_at IS NULL
-         AND ${referenceExclusionSql()}`,
+         AND deleted_at IS NULL`,
   );
 
   if (totalEntities === 0) {
@@ -241,7 +237,6 @@ export async function checkTimelineCoverage(
        SELECT p.id FROM pages p ${sampleClause}
        WHERE p.type IN ('person', 'company', 'organization', 'entity')
          AND p.deleted_at IS NULL
-         AND ${referenceExclusionSql('p')}
          AND EXISTS (SELECT 1 FROM timeline_entries t WHERE t.page_id = p.id)
      ) sub`,
   );

src/core/onboard/init-nudge.ts

@@ -14,7 +14,6 @@
 // also short-circuits (CI/scripted callers see nothing).
 
 import type { BrainEngine } from '../engine.ts';
-import { referenceExclusionSql } from '../reference-flag.ts';
 
 const NUDGE_BUDGET_MS = 3000;
 
@@ -58,8 +57,7 @@ export async function runInitNudge(engine: BrainEngine): Promise<void> {
       engine.executeRaw<{ count: string | number }>(
         `SELECT COUNT(*) AS count FROM pages
            WHERE type IN ('person', 'company', 'organization', 'entity')
-             AND deleted_at IS NULL
-             AND ${referenceExclusionSql()}`,
+             AND deleted_at IS NULL`,
         [],
         { signal: controller.signal },
       ),
@@ -67,7 +65,6 @@ export async function runInitNudge(engine: BrainEngine): Promise<void> {
         `SELECT COUNT(*) AS count FROM pages p
            WHERE p.type IN ('person', 'company', 'organization', 'entity')
              AND p.deleted_at IS NULL
-             AND ${referenceExclusionSql('p')}
              AND EXISTS (SELECT 1 FROM links l WHERE l.to_page_id = p.id)`,
         [],
         { signal: controller.signal },
@@ -76,7 +73,6 @@ export async function runInitNudge(engine: BrainEngine): Promise<void> {
         `SELECT COUNT(*) AS count FROM pages p
            WHERE p.type IN ('person', 'company', 'organization', 'entity')
              AND p.deleted_at IS NULL
-             AND ${referenceExclusionSql('p')}
              AND EXISTS (SELECT 1 FROM timeline_entries t WHERE t.page_id = p.id)`,
         [],
         { signal: controller.signal },

src/core/pglite-engine.ts

@@ -23,7 +23,6 @@ import { runMigrations } from './migrate.ts';
 import { PGLITE_SCHEMA_SQL, getPGLiteSchema } from './pglite-schema.ts';
 import { DEFAULT_EMBEDDING_MODEL, DEFAULT_EMBEDDING_DIMENSIONS } from './ai/defaults.ts';
 import { DELETE_BATCH_SIZE } from './engine-constants.ts';
-import { referenceExclusionSql } from './reference-flag.ts';
 import { MARKDOWN_CHUNKER_VERSION } from './chunkers/recursive.ts';
 import { acquireLock, releaseLock, type LockHandle } from './pglite-lock.ts';
 import { getFtsLanguage } from './fts-language.ts';
@@ -5203,7 +5202,6 @@ export class PGLiteEngine implements BrainEngine {
     const { rows: [h] } = await this.db.query(`
       WITH entity_pages AS (
         SELECT id, slug FROM pages WHERE type IN ('person', 'company')
-          AND ${referenceExclusionSql()}
       )
       SELECT
         (SELECT count(*) FROM pages) as page_count,

src/core/postgres-engine.ts

@@ -5320,12 +5320,7 @@ export class PostgresEngine implements BrainEngine {
     // is working as intended, not an orphan.
     const [h] = await sql`
       WITH entity_pages AS (
-        -- reference:true pages are exempt from coverage metrics.
-        -- Inlined (postgres.js tagged-template interpolation = bound param,
-        -- not raw SQL); keep in sync with referenceExclusionSql() in
-        -- reference-flag.ts.
         SELECT id, slug FROM pages WHERE type IN ('person', 'company')
-          AND (frontmatter->>'reference') IS DISTINCT FROM 'true'
       )
       SELECT
         (SELECT count(*) FROM pages) as page_count,

src/core/reference-flag.ts

@@ -1,37 +0,0 @@
-// Reference / canon entities.
-//
-// A page with frontmatter `reference: true` is a reference-only entity — a
-// figure or organization imported from a book / article / external source that
-// the user reads ABOUT but does not actively interact with (e.g. Andy Grove,
-// Kleiner Perkins). It keeps its real `type` (`person` / `company`), so it stays
-// fully searchable, enrichable, linkable, and edge-resolvable — NOTHING about
-// retrieval changes.
-//
-// The ONLY behavior it opts out of is the entity *coverage* metrics
-// (`timeline_coverage`, `entity_link_coverage` + their onboard-nudge mirrors).
-// Those metrics nudge "this entity should accumulate dated history / inbound
-// links"; that assumption is right for people you actually deal with and wrong
-// for canon imports, which have no dated events in the user's life. Excluding
-// them keeps the metric honest and actionable instead of permanently red.
-//
-// Opt-in: absent / false / anything-but-true = normal entity (the default).
-// Set via `gbrain reference <slug>` (or hand-edit frontmatter).
-
-export const REFERENCE_FRONTMATTER_KEY = 'reference';
-
-/**
- * SQL predicate (true for NON-reference pages) to AND into entity-coverage
- * denominators AND numerators so the ratio stays consistent. JSONB `->>` yields
- * text; `IS DISTINCT FROM 'true'` treats absent (NULL) and every non-true value
- * as a normal counted entity. Backed by the GIN index on `pages.frontmatter`.
- *
- * @param alias optional table alias (e.g. 'p' for `pages p`); omit for bare `pages`.
- *
- * NOTE: postgres-engine.ts uses a postgres.js tagged template where `${}` is a
- * bound parameter, not raw SQL, so it inlines this predicate literally — keep
- * the two in sync.
- */
-export function referenceExclusionSql(alias = ''): string {
-  const col = alias ? `${alias}.frontmatter` : 'frontmatter';
-  return `(${col}->>'${REFERENCE_FRONTMATTER_KEY}') IS DISTINCT FROM 'true'`;
-}

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/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/reference-flag.test.ts

@@ -1,154 +0,0 @@
-import { describe, expect, test } from 'bun:test';
-import { referenceExclusionSql, REFERENCE_FRONTMATTER_KEY } from '../src/core/reference-flag.ts';
-import { applyReferenceFrontmatter } from '../src/commands/reference.ts';
-
-describe('referenceExclusionSql', () => {
-  test('bare pages (no alias)', () => {
-    expect(referenceExclusionSql()).toBe(`(frontmatter->>'reference') IS DISTINCT FROM 'true'`);
-  });
-  test('aliased', () => {
-    expect(referenceExclusionSql('p')).toBe(`(p.frontmatter->>'reference') IS DISTINCT FROM 'true'`);
-  });
-  test('key constant', () => {
-    expect(REFERENCE_FRONTMATTER_KEY).toBe('reference');
-  });
-});
-
-describe('applyReferenceFrontmatter', () => {
-  const page = `---\ntitle: Andy Grove\ntype: person\ntags: []\n---\n\n# Andy Grove\n\nBody.`;
-
-  test('adds reference: true to an existing frontmatter block', () => {
-    const out = applyReferenceFrontmatter(page, true);
-    expect(out).toContain('reference: true');
-    expect(out).toContain('type: person');   // other keys preserved
-    expect(out).toContain('# Andy Grove');    // body preserved
-  });
-
-  test('is idempotent — does not duplicate the key', () => {
-    const once = applyReferenceFrontmatter(page, true);
-    const twice = applyReferenceFrontmatter(once, true);
-    expect(twice).toBe(once);
-    expect(twice.match(/reference: true/g)).toHaveLength(1);
-  });
-
-  test('replaces a stale reference: false with true', () => {
-    const off = `---\ntype: person\nreference: false\n---\n\nBody.`;
-    const out = applyReferenceFrontmatter(off, true);
-    expect(out).toContain('reference: true');
-    expect(out).not.toContain('reference: false');
-  });
-
-  test('--unset removes the key', () => {
-    const on = applyReferenceFrontmatter(page, true);
-    const off = applyReferenceFrontmatter(on, false);
-    expect(off).not.toContain('reference:');
-    expect(off).toContain('type: person');
-  });
-
-  test('--unset on a page without the key is a no-op', () => {
-    expect(applyReferenceFrontmatter(page, false)).toBe(page);
-  });
-
-  test('setting on a frontmatter-less page prepends a block', () => {
-    const raw = '# Just a heading\n\nNo frontmatter here.';
-    const out = applyReferenceFrontmatter(raw, true);
-    expect(out.startsWith('---\nreference: true\n---\n')).toBe(true);
-    expect(out).toContain('# Just a heading');
-  });
-
-  test('preserves body containing YAML-special chars', () => {
-    const tricky = `---\ntype: person\n---\n\nText with $& and $1 literals.`;
-    const out = applyReferenceFrontmatter(tricky, true);
-    expect(out).toContain('Text with $& and $1 literals.');
-    expect(out).toContain('reference: true');
-  });
-});
-
-// ── e2e: getHealth exemption + (source_id, slug)-scoped JSONB write ─────────
-
-import { afterAll, beforeAll } from 'bun:test';
-import * as fs from 'node:fs';
-import * as os from 'node:os';
-import * as path from 'node:path';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { runReference } from '../src/commands/reference.ts';
-import { withEnv } from './helpers/with-env.ts';
-
-describe('reference flag e2e (PGLite)', () => {
-  let engine: PGLiteEngine;
-  let brainDir: string;
-
-  beforeAll(async () => {
-    engine = new PGLiteEngine();
-    await engine.connect({ database_url: '' });
-    await engine.initSchema();
-
-    brainDir = fs.mkdtempSync(path.join(os.tmpdir(), 'gbrain-ref-'));
-    fs.mkdirSync(path.join(brainDir, 'people'), { recursive: true });
-    fs.writeFileSync(
-      path.join(brainDir, 'people/andy-grove.md'),
-      '---\ntitle: Andy Grove\ntype: person\n---\n\n# Andy Grove\n',
-      'utf8',
-    );
-
-    await engine.executeRaw(
-      `INSERT INTO sources (id, name, local_path, config, created_at)
-       VALUES ('src-a', 'a', $1, '{}'::jsonb, now()), ('src-b', 'b', NULL, '{}'::jsonb, now())`,
-      [brainDir],
-    );
-    // Same slug in BOTH sources — the write must only touch src-a.
-    await engine.putPage('people/andy-grove', {
-      type: 'person', title: 'Andy Grove', compiled_truth: 'canon figure',
-    }, { sourceId: 'src-a' });
-    await engine.putPage('people/andy-grove', {
-      type: 'person', title: 'Andy Grove', compiled_truth: 'other-source twin',
-    }, { sourceId: 'src-b' });
-    // A normal contact WITH a timeline entry, so coverage has a live numerator.
-    const contact = await engine.putPage('people/alice-example', {
-      type: 'person', title: 'Alice Example', compiled_truth: 'real contact',
-    }, { sourceId: 'src-a' });
-    await engine.executeRaw(
-      `INSERT INTO timeline_entries (page_id, date, summary) VALUES ($1, '2026-01-01', 'met')`,
-      [contact.id],
-    );
-  });
-
-  afterAll(async () => {
-    await engine.disconnect();
-    fs.rmSync(brainDir, { recursive: true, force: true });
-  });
-
-  test('runReference scopes the JSONB write to the resolved (source_id, slug)', async () => {
-    await withEnv({ GBRAIN_SOURCE: undefined }, () =>
-      runReference(engine, ['people/andy-grove', '--brain', brainDir]));
-
-    const rows = await engine.executeRaw<{ source_id: string; ref: string | null }>(
-      `SELECT source_id, frontmatter->>'reference' AS ref FROM pages WHERE slug = 'people/andy-grove' ORDER BY source_id`,
-    );
-    expect(rows).toEqual([
-      { source_id: 'src-a', ref: 'true' },
-      { source_id: 'src-b', ref: null },   // twin in the other source untouched
-    ]);
-    // Durable half: frontmatter on disk got the flag too.
-    expect(fs.readFileSync(path.join(brainDir, 'people/andy-grove.md'), 'utf8'))
-      .toContain('reference: true');
-  });
-
-  test('getHealth exempts reference pages from timeline/link coverage', async () => {
-    const health = await engine.getHealth();
-    // 3 person pages; the 2 reference-less twins would drag coverage to 1/3.
-    // With the src-a twin marked reference, denominator = 2 (alice + src-b twin).
-    expect(health.timeline_coverage).toBeCloseTo(0.5);
-  });
-
-  test('unset restores the page to a normal counted entity', async () => {
-    await withEnv({ GBRAIN_SOURCE: undefined }, () =>
-      runReference(engine, ['people/andy-grove', '--unset', '--brain', brainDir]));
-    const rows = await engine.executeRaw<{ ref: string | null }>(
-      `SELECT frontmatter->>'reference' AS ref FROM pages WHERE slug = 'people/andy-grove' AND source_id = 'src-a'`,
-    );
-    expect(rows[0].ref).toBeNull();
-    const health = await engine.getHealth();
-    expect(health.timeline_coverage).toBeCloseTo(1 / 3);
-  });
-});