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

@@ -466,11 +466,6 @@ async function main() {
     const result = JSON.parse(JSON.stringify(rawResult, bigintToStringReplacer));
     const output = formatResult(op.name, result);
     if (output) process.stdout.write(output);
-    // #1484 — invisible-miss hint: a bare query/search that hit zero results
-    // on a multi-source brain tells the user (stderr) which source it
-    // actually searched and how to widen the scope.
-    const hint = await sourceScopeHint(op.name, params, ctx.sourceId, engine, result);
-    if (hint) console.error(hint);
   } catch (e: unknown) {
     // v0.42.20.0 (codex D4): on error, set exitCode + return so the `finally`
     // STILL runs (drains every background-work sink + disconnects). A bare
@@ -842,44 +837,6 @@ async function makeContext(engine: BrainEngine, params: Record<string, unknown>)
   };
 }
 
-/**
- * #1484 — a bare `gbrain query`/`search` silently scopes to the resolved
- * source (usually 'default'); on a multi-source brain a zero-hit run looks
- * identical to "the brain doesn't know this" even when the answer lives in
- * another source. Returns a stderr hint when (a) the op is query/search,
- * (b) it returned zero results, (c) the caller did NOT scope explicitly
- * (--source / --source-id / --all-sources), and (d) the brain has >1
- * registered source. Best-effort: any lookup failure returns null.
- *
- * Exported for tests (same import-safety contract as formatResult).
- */
-export async function sourceScopeHint(
-  opName: string,
-  params: Record<string, unknown>,
-  sourceId: string,
-  engine: BrainEngine,
-  result: unknown,
-): Promise<string | null> {
-  if (opName !== 'query' && opName !== 'search') return null;
-  if (!Array.isArray(result) || result.length > 0) return null;
-  // Explicit scoping (flag tier) = user intent; don't second-guess it.
-  if (params.source || params.source_id || params.all_sources) return null;
-  if (sourceId === '__all__') return null;
-  try {
-    const rows = await engine.executeRaw<{ n: number }>(
-      `SELECT count(*)::int AS n FROM sources`,
-    );
-    const n = Number(rows[0]?.n ?? 0);
-    if (n <= 1) return null;
-    return (
-      `Hint: this brain has ${n} sources; you searched only "${sourceId}". ` +
-      `Retry with --source-id __all__ (all sources) or --source-id <id>.`
-    );
-  } catch {
-    return null; // hint is best-effort; never fail the query over it
-  }
-}
-
 // Exported for tests (same import-safety contract as cliAliases/printOpHelp).
 export function formatResult(opName: string, result: unknown): string {
   switch (opName) {

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/core/ai/gateway.ts

@@ -1000,24 +1000,9 @@ const voyageCompatFetch = (async (input: RequestInfo | URL, init?: RequestInit)
   // Voyage diverges from OpenAI in two places that break the parser:
   //   - `embedding` is a base64 string (SDK schema expects `number[]`)
   //   - `usage` lacks `prompt_tokens` (SDK schema requires it when usage present)
-  //
-  // #1610: read the body ONCE via text() and JSON.parse it. The pre-fix
-  // `await resp.clone().json()` truncated large bodies on bun < 1.1.27
-  // (oven-sh/bun#6348) — the parse threw, the catch fell back to the raw
-  // response, and multi-chunk pages died with "Invalid JSON response".
-  // Every JSON return path below rebuilds the Response so a stale
-  // Content-Length/Content-Encoding header from the original can't lie
-  // about the rewritten body.
-  const bodyText = await resp.text();
-  const rebuild = (body: string) => {
-    const headers = new Headers(resp.headers);
-    headers.delete('content-length');
-    headers.delete('content-encoding');
-    return new Response(body, { status: resp.status, statusText: resp.statusText, headers });
-  };
   try {
-    const json: any = JSON.parse(bodyText);
-    if (!json || typeof json !== 'object') return rebuild(bodyText);
+    const json: any = await resp.clone().json();
+    if (!json || typeof json !== 'object') return resp;
     let modified = false;
     if (Array.isArray(json.data)) {
       for (const item of json.data) {
@@ -1052,19 +1037,22 @@ const voyageCompatFetch = (async (input: RequestInfo | URL, init?: RequestInit)
         : 0;
       modified = true;
     }
-    if (!modified) return rebuild(bodyText);
-    return rebuild(JSON.stringify(json));
+    if (!modified) return resp;
+    return new Response(JSON.stringify(json), {
+      status: resp.status,
+      statusText: resp.statusText,
+      headers: resp.headers,
+    });
   } catch (err) {
     // OOM-cap throws MUST propagate. The catch is here for "Voyage returned
     // JSON I can't reshape" (parse error, unexpected schema) — falling back
-    // to the original body is correct in that case. Letting the
+    // to the original response is correct in that case. Letting the
     // too-large response through here would defeat the entire purpose of
     // Layer 2 (the per-embedding cap that fires when Content-Length wasn't
     // available to Layer 1).
     if (err instanceof VoyageResponseTooLargeError) throw err;
-    // If parsing/transformation fails, pass the original body through
-    // (rebuilt — resp's body stream is already consumed by text()).
-    return rebuild(bodyText);
+    // If parsing/transformation fails, fall back to the original response.
+    return resp;
   }
 }) as unknown as typeof fetch;
 
@@ -1204,21 +1192,9 @@ const zeroEntropyCompatFetch = (async (input: RequestInfo | URL, init?: RequestI
   // validates. Also map usage.total_tokens → prompt_tokens (SDK requires
   // prompt_tokens when `usage` is present — same divergence Voyage hit at
   // gateway.ts:655).
-  //
-  // #1610: read the body ONCE via text() + JSON.parse — `resp.clone().json()`
-  // truncated large bodies on bun < 1.1.27 (oven-sh/bun#6348), so the parse
-  // threw and the catch fell back to the RAW ZE `{results: ...}` shape, which
-  // the AI SDK schema rejects → "Invalid JSON response" on multi-chunk pages.
-  const bodyText = await resp.text();
-  const rebuild = (body: string) => {
-    const headers = new Headers(resp.headers);
-    headers.delete('content-length');
-    headers.delete('content-encoding');
-    return new Response(body, { status: resp.status, statusText: resp.statusText, headers });
-  };
   try {
-    const json: any = JSON.parse(bodyText);
-    if (!json || typeof json !== 'object') return rebuild(bodyText);
+    const json: any = await resp.clone().json();
+    if (!json || typeof json !== 'object') return resp;
     let modified = false;
     if (Array.isArray(json.results) && !Array.isArray(json.data)) {
       // Layer 2 OOM cap — per-embedding size. ZE returns float[] arrays,
@@ -1252,25 +1228,20 @@ const zeroEntropyCompatFetch = (async (input: RequestInfo | URL, init?: RequestI
       // SDK also expects total_tokens; ZE provides it directly.
       modified = true;
     }
-    if (!modified) return rebuild(bodyText);
-    return rebuild(JSON.stringify(json));
+    if (!modified) return resp;
+    return new Response(JSON.stringify(json), {
+      status: resp.status,
+      statusText: resp.statusText,
+      headers: resp.headers,
+    });
   } catch (err) {
     // OOM-cap throws MUST propagate. Voyage's pattern: instanceof check on
     // its own tagged class. Same here — only rethrow our own cap class.
     if (err instanceof ZeroEntropyResponseTooLargeError) throw err;
-    return rebuild(bodyText);
+    return resp;
   }
 }) as unknown as typeof fetch;
 
-/**
- * Test-only seams (#1610): the compat shims are module-private closures;
- * exporting them lets tests drive the response-rewrite paths behaviorally
- * (truncating clone(), stale Content-Length) without a live provider.
- * Same pattern as __getShrinkStateForTests.
- */
-export const __voyageCompatFetchForTests = voyageCompatFetch;
-export const __zeroEntropyCompatFetchForTests = zeroEntropyCompatFetch;
-
 /**
  * Generic asymmetric-embedding shim for openai-compatible recipes that
  * ship no compat fetch of their own (llama-server, litellm, ollama, ...).
@@ -1542,12 +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[] = [];
@@ -1597,6 +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.
  */
@@ -1604,15 +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;
@@ -1638,7 +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/operations.ts

@@ -499,7 +499,7 @@ export function linkReadScopeOpts(ctx: OperationContext): { sourceId?: string; s
  * FAIL-CLOSED: anything not strictly `ctx.remote === false` is untrusted.
  *
  * This is the SINGLE resolver for every read op that accepts a per-call
- * `source_id` / `all_sources` parameter (query, search, code_callers, code_callees,
+ * `source_id` / `all_sources` parameter (query, code_callers, code_callees,
  * get_page, search_by_image, code_blast, code_flow). Inlining the `__all__`
  * branch per handler is the bug class that leaked cross-source reads (#1924,
  * #1371): a remote client could pass `source_id: '__all__'` to opt out of its
@@ -1442,24 +1442,13 @@ const search: Operation = {
     limit: { type: 'number', description: 'Max results (default 20)' },
     offset: { type: 'number', description: 'Skip first N results (for pagination)' },
     mode: { type: 'string', description: 'Search mode (conservative|balanced|tokenmax). Local callers only.' },
-    source_id: {
-      type: 'string',
-      description:
-        "Scope search to a single source. Defaults to OperationContext.sourceId. Pass '__all__' to span every source for trusted local callers; for remote callers '__all__' spans only your granted sources.",
-    },
-    all_sources: { type: 'boolean', description: "Span sources (equivalent to source_id=__all__): every source locally, your grant remotely." },
   },
   handler: async (ctx, p) => {
     const startedAt = Date.now();
     const queryText = p.query as string;
     const limit = (p.limit as number) || 20;
     const offset = (p.offset as number) || 0;
-    // #1484 follow-up: route through the canonical fail-closed resolver so
-    // `--source-id __all__` / `all_sources` behave the same as on `query`
-    // (the zero-hit CLI hint advises exactly that retry). Without a per-call
-    // param, `search` silently ignored --source-id — the retry looked like
-    // a genuine miss.
-    const scope = resolveRequestedScope(ctx, p.source_id as string | undefined, p.all_sources === true);
+    const scope = sourceScopeOpts(ctx);
 
     // T4/D5 — per-call mode honored ONLY for trusted/local callers so a remote
     // OAuth client can't escalate to the costly tokenmax bundle. Local + unknown

src/core/search/hybrid.ts

@@ -1323,18 +1323,8 @@ export async function hybridSearch(
       if (effectiveModality === 'both' && imageVectorList !== null) {
         vectorLists = [...vectorLists, imageVectorList];
       }
-    } catch (err) {
-      // Embedding/vector failure is non-fatal — fall back to keyword-only —
-      // but say WHY (#1626): this arm only runs when the embedding provider
-      // probed available, so a throw here is a real failure (embed timeout,
-      // transient pooler error on the searchVector fan-out). Pre-fix the bare
-      // catch made a cross-source `--source __all__` run silently collapse to
-      // keyword-only/"No results" with zero diagnostics.
-      warnOncePerProcess(
-        'hybrid-vector-arm-failed',
-        `[gbrain] vector arm failed (fail-open, keyword-only fallback): ` +
-          `${err instanceof Error ? err.message : String(err)}`,
-      );
+    } catch {
+      // Embedding failure is non-fatal, fall back to keyword-only
     }
   }
 

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/compat-fetch-body-read.serial.test.ts

@@ -1,115 +0,0 @@
-/**
- * #1610 — Voyage/ZeroEntropy compat shims must read the response body ONCE
- * via text() instead of `resp.clone().json()`.
- *
- * On bun < 1.1.27, Response.clone() truncates large bodies (oven-sh/bun#6348):
- * the clone().json() parse threw, the shim's catch fell back to the ORIGINAL
- * response — whose wire shape (ZE `{results: ...}`, Voyage base64 embeddings)
- * the AI SDK's openai-compatible Zod schema rejects — and multi-chunk pages
- * failed with "Invalid JSON response".
- *
- * These tests simulate the truncating clone() and assert the shims still
- * return the fully rewritten body. They also pin that the rewritten Response
- * does NOT carry the original (now stale) Content-Length header, which lied
- * about the rewritten body's size (gateway.ts previously copied
- * `headers: resp.headers` verbatim).
- */
-
-import { afterEach, describe, expect, test } from 'bun:test';
-import {
-  __voyageCompatFetchForTests,
-  __zeroEntropyCompatFetchForTests,
-} from '../../src/core/ai/gateway.ts';
-
-const origFetch = globalThis.fetch;
-afterEach(() => {
-  globalThis.fetch = origFetch;
-});
-
-/** Build a Response whose clone() truncates the body (bun < 1.1.27 behavior). */
-function truncatingCloneResponse(body: string): Response {
-  const headers = {
-    'content-type': 'application/json',
-    // Deliberately stale after any rewrite: the original wire body's length.
-    'content-length': String(Buffer.byteLength(body)),
-  };
-  const resp = new Response(body, { status: 200, headers });
-  (resp as any).clone = () =>
-    new Response(body.slice(0, 32), { status: 200, headers });
-  return resp;
-}
-
-describe('voyageCompatFetch — single body read (#1610)', () => {
-  test('rewrites base64 embeddings even when clone() truncates the body', async () => {
-    const floats = new Float32Array([0.5, 0.25, -1]);
-    const b64 = Buffer.from(floats.buffer).toString('base64');
-    const wireBody = JSON.stringify({
-      object: 'list',
-      data: [{ object: 'embedding', embedding: b64, index: 0 }],
-      model: 'voyage-3',
-      usage: { total_tokens: 7 },
-    });
-    globalThis.fetch = (async () => truncatingCloneResponse(wireBody)) as unknown as typeof fetch;
-
-    const out = await __voyageCompatFetchForTests('https://api.voyageai.com/v1/embeddings', {
-      method: 'POST',
-      body: JSON.stringify({ input: ['hello'], model: 'voyage-3' }),
-      headers: { 'content-type': 'application/json' },
-    });
-
-    const json: any = await out.json();
-    expect(Array.from(json.data[0].embedding)).toEqual([0.5, 0.25, -1]);
-    expect(json.usage.prompt_tokens).toBe(7);
-    // Stale Content-Length from the wire body must not survive the rewrite.
-    expect(out.headers.get('content-length')).toBeNull();
-    expect(out.headers.get('content-encoding')).toBeNull();
-  });
-});
-
-describe('zeroEntropyCompatFetch — single body read (#1610)', () => {
-  test('rewrites {results} → {data} even when clone() truncates the body', async () => {
-    const wireBody = JSON.stringify({
-      results: [{ embedding: [0.1, 0.2] }, { embedding: [0.3, 0.4] }],
-      usage: { total_bytes: 42, total_tokens: 9 },
-    });
-    let fetchedUrl = '';
-    globalThis.fetch = (async (url: string | URL | Request) => {
-      fetchedUrl = String(url);
-      return truncatingCloneResponse(wireBody);
-    }) as unknown as typeof fetch;
-
-    const out = await __zeroEntropyCompatFetchForTests('https://api.zeroentropy.dev/v1/embeddings', {
-      method: 'POST',
-      body: JSON.stringify({ input: ['hello'], model: 'zembed-1' }),
-      headers: { 'content-type': 'application/json' },
-    });
-
-    expect(fetchedUrl.endsWith('/v1/models/embed')).toBe(true);
-    const json: any = await out.json();
-    // The AI SDK schema requires {data: [{embedding, index}]} — the raw ZE
-    // {results} fallback is exactly the pre-fix "Invalid JSON response".
-    expect(json.results).toBeUndefined();
-    expect(json.data).toHaveLength(2);
-    expect(json.data[0]).toEqual({ object: 'embedding', embedding: [0.1, 0.2], index: 0 });
-    expect(json.data[1].index).toBe(1);
-    expect(json.usage.prompt_tokens).toBe(9);
-    expect(out.headers.get('content-length')).toBeNull();
-  });
-
-  test('non-JSON body falls back to the original bytes (rebuilt, still readable)', async () => {
-    const wireBody = 'plain text, not json';
-    globalThis.fetch = (async () =>
-      new Response(wireBody, {
-        status: 200,
-        headers: { 'content-type': 'application/json' },
-      })) as unknown as typeof fetch;
-
-    const out = await __zeroEntropyCompatFetchForTests('https://api.zeroentropy.dev/v1/embeddings', {
-      method: 'POST',
-      body: JSON.stringify({ input: ['hello'] }),
-    });
-    // Body was consumed by the shim's single read; the fallback must
-    // rebuild a readable Response rather than return the drained original.
-    expect(await out.text()).toBe(wireBody);
-  });
-});

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/ai/zeroentropy-compat-fetch.test.ts

@@ -98,18 +98,16 @@ describe('zeroEntropyCompatFetch — OOM caps', () => {
     expect(src).toMatch(/MAX_ZEROENTROPY_RESPONSE_BYTES\s*=\s*256\s*\*\s*1024\s*\*\s*1024/);
   });
 
-  test('Layer 1: Content-Length pre-check before the body is read', async () => {
+  test('Layer 1: Content-Length pre-check before resp.clone().json()', async () => {
     const src = await Bun.file(GATEWAY_PATH).text();
     // Find the zeroEntropyCompatFetch block bounds, then assert ordering
-    // within it (mirroring the voyage cap test pattern). #1610 moved the
-    // body read from `resp.clone().json()` to a single `resp.text()` (bun
-    // < 1.1.27 truncates clone()d bodies, oven-sh/bun#6348).
+    // within it (mirroring the voyage cap test pattern).
     const zeFetchStart = src.indexOf('const zeroEntropyCompatFetch');
     expect(zeFetchStart).toBeGreaterThan(0);
-    const block = src.slice(zeFetchStart, zeFetchStart + 9000);
+    const block = src.slice(zeFetchStart, zeFetchStart + 8000);
 
     const preCheckIdx = block.indexOf("resp.headers.get('content-length')");
-    const jsonParseIdx = block.indexOf('const bodyText = await resp.text()');
+    const jsonParseIdx = block.indexOf('await resp.clone().json()');
     expect(preCheckIdx).toBeGreaterThan(0);
     expect(jsonParseIdx).toBeGreaterThan(0);
     // The pre-check MUST appear before the JSON parse — Voyage's lesson

test/cli-source-scope-hint.test.ts

@@ -1,61 +0,0 @@
-/**
- * #1484 — invisible-miss hint. A bare `gbrain query` resolves to a single
- * source (usually 'default'); on a multi-source brain a zero-hit run gave no
- * signal that the answer might live in another source. sourceScopeHint
- * returns the stderr hint exactly when: query/search op + zero results +
- * no explicit scoping param + >1 registered source.
- */
-
-import { describe, expect, test } from 'bun:test';
-import { sourceScopeHint } from '../src/cli.ts';
-import type { BrainEngine } from '../src/core/engine.ts';
-
-function fakeEngine(sourceCount: number, fail = false): BrainEngine {
-  return {
-    executeRaw: async () => {
-      if (fail) throw new Error('sources table missing');
-      return [{ n: sourceCount }];
-    },
-  } as unknown as BrainEngine;
-}
-
-describe('sourceScopeHint (#1484)', () => {
-  test('fires on a bare zero-hit query against a multi-source brain', async () => {
-    const hint = await sourceScopeHint('query', {}, 'default', fakeEngine(3), []);
-    expect(hint).toContain('3 sources');
-    expect(hint).toContain('"default"');
-    expect(hint).toContain('--source-id __all__');
-  });
-
-  test('fires for search too', async () => {
-    const hint = await sourceScopeHint('search', {}, 'wiki', fakeEngine(2), []);
-    expect(hint).toContain('"wiki"');
-  });
-
-  test('silent when results were found', async () => {
-    expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3), [{ slug: 'a' }])).toBeNull();
-  });
-
-  test('silent when the caller scoped explicitly', async () => {
-    expect(await sourceScopeHint('query', { source_id: 'wiki' }, 'wiki', fakeEngine(3), [])).toBeNull();
-    expect(await sourceScopeHint('query', { source: 'wiki' }, 'wiki', fakeEngine(3), [])).toBeNull();
-    expect(await sourceScopeHint('query', { all_sources: true }, '__all__', fakeEngine(3), [])).toBeNull();
-  });
-
-  test('silent when the resolved scope is already __all__', async () => {
-    expect(await sourceScopeHint('query', {}, '__all__', fakeEngine(3), [])).toBeNull();
-  });
-
-  test('silent on a single-source brain', async () => {
-    expect(await sourceScopeHint('query', {}, 'default', fakeEngine(1), [])).toBeNull();
-  });
-
-  test('silent for non-search ops and non-array results', async () => {
-    expect(await sourceScopeHint('get_stats', {}, 'default', fakeEngine(3), [])).toBeNull();
-    expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3), { rows: [] })).toBeNull();
-  });
-
-  test('best-effort: sources lookup failure returns null, never throws', async () => {
-    expect(await sourceScopeHint('query', {}, 'default', fakeEngine(3, true), [])).toBeNull();
-  });
-});

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/hybrid-vector-arm-warn.serial.test.ts

@@ -1,77 +0,0 @@
-/**
- * #1626 — hybridSearch's text-vector arm must not fail DARK.
- *
- * The arm only runs when the embedding provider probed available, so a throw
- * inside it (embed timeout, transient pooler error on searchVector) is a real
- * failure. Pre-fix, a bare `catch {}` swallowed it and the run silently
- * collapsed to keyword-only — under `--source __all__` on a strained pooler
- * that read as a non-deterministic "No results". The fix logs the swallowed
- * reason via warnOncePerProcess while keeping the keyword fallback.
- */
-
-import { afterAll, beforeAll, describe, expect, test } from 'bun:test';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { hybridSearch } from '../src/core/search/hybrid.ts';
-import {
-  __setEmbedTransportForTests,
-  configureGateway,
-  resetGateway,
-} from '../src/core/ai/gateway.ts';
-import { _resetWarnOnceForTests } from '../src/core/utils.ts';
-
-let engine: PGLiteEngine;
-const origWarn = console.warn;
-
-beforeAll(async () => {
-  // Pin the gateway to OpenAI with a stub key (put-page-provenance pattern):
-  // embed() runs instantiateEmbedding — which requires OPENAI_API_KEY — BEFORE
-  // the stubbed transport is reached. Without this, a keyless CI environment
-  // throws the config error instead of the transport's, and the assertion on
-  // the swallowed reason fails. The key never leaves the process.
-  configureGateway({
-    embedding_model: 'openai:text-embedding-3-large',
-    embedding_dimensions: 1536,
-    env: { ...process.env, OPENAI_API_KEY: process.env.OPENAI_API_KEY || 'sk-test-stub' },
-  });
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-  await engine.putPage('people/alice-example', {
-    type: 'person',
-    title: 'Alice Example',
-    compiled_truth: 'Alice Example is a test person for the vector-arm warn test.',
-  });
-});
-
-afterAll(async () => {
-  console.warn = origWarn;
-  __setEmbedTransportForTests(null);
-  resetGateway();
-  await engine.disconnect();
-});
-
-describe('hybridSearch vector-arm failure telemetry (#1626)', () => {
-  test('embed failure logs the swallowed reason and falls back to keyword', async () => {
-    _resetWarnOnceForTests();
-    // Installing a transport makes isAvailable('embedding') true (test-seam
-    // fast path), so the vector arm RUNS — and then throws.
-    __setEmbedTransportForTests(() => {
-      throw new Error('pooler exploded mid-fanout');
-    });
-    const warnings: string[] = [];
-    console.warn = (...args: unknown[]) => {
-      warnings.push(args.map(String).join(' '));
-    };
-    try {
-      const results = await hybridSearch(engine, 'alice');
-      // Keyword fallback still returns results — fail-open preserved.
-      expect(results.some((r) => r.slug === 'people/alice-example')).toBe(true);
-    } finally {
-      console.warn = origWarn;
-      __setEmbedTransportForTests(null);
-    }
-    const armWarnings = warnings.filter((w) => w.includes('vector arm failed'));
-    expect(armWarnings).toHaveLength(1);
-    expect(armWarnings[0]).toContain('pooler exploded mid-fanout');
-  });
-});

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/search-op-source-scope.test.ts

@@ -1,87 +0,0 @@
-/**
- * #1484 follow-up — the `search` op must honor per-call `source_id` /
- * `all_sources` through the canonical fail-closed resolver
- * (resolveRequestedScope), exactly like `query` does.
- *
- * Pre-fix, `search` had no source_id param at all: the zero-hit CLI hint
- * advised "retry with --source-id __all__", the flag parsed into params,
- * NOTHING consumed it, and the retry silently re-ran the same single-source
- * search — an invisible false negative (and the retry's params.source_id
- * suppressed the hint, so the user got no second warning).
- */
-
-import { describe, expect, test } from 'bun:test';
-import { operationsByName } from '../src/core/operations.ts';
-import type { OperationContext } from '../src/core/operations.ts';
-import type { BrainEngine } from '../src/core/engine.ts';
-
-const searchOp = operationsByName['search'];
-
-/** Fake engine: keyword-only config so the handler's scope goes straight to
- *  searchKeyword, where we capture the opts it was called with. */
-function makeCtx(remote: boolean, allowedSources?: string[]) {
-  const captured: { opts?: Record<string, unknown> } = {};
-  const engine = {
-    getConfig: async (key: string) => (key === 'search.mcp_keyword_only' ? 'true' : null),
-    searchKeyword: async (_q: string, opts: Record<string, unknown>) => {
-      captured.opts = opts;
-      return [];
-    },
-  } as unknown as BrainEngine;
-  const ctx = {
-    engine,
-    config: { engine: 'pglite' },
-    logger: { info: () => {}, warn: () => {}, error: () => {} },
-    dryRun: false,
-    remote,
-    sourceId: 'default',
-    ...(allowedSources ? { auth: { allowedSources } } : {}),
-  } as unknown as OperationContext;
-  return { ctx, captured };
-}
-
-describe('search op per-call source scope (#1484 follow-up)', () => {
-  test('op declares source_id + all_sources params (the CLI hint advises them)', () => {
-    expect(searchOp.params.source_id).toBeDefined();
-    expect(searchOp.params.all_sources).toBeDefined();
-  });
-
-  test('default: scopes to ctx.sourceId', async () => {
-    const { ctx, captured } = makeCtx(false);
-    await searchOp.handler(ctx, { query: 'x' });
-    expect(captured.opts?.sourceId).toBe('default');
-  });
-
-  test("local + source_id '__all__' spans the whole brain (no source filter)", async () => {
-    const { ctx, captured } = makeCtx(false);
-    await searchOp.handler(ctx, { query: 'x', source_id: '__all__' });
-    expect(captured.opts?.sourceId).toBeUndefined();
-    expect(captured.opts?.sourceIds).toBeUndefined();
-  });
-
-  test('local + all_sources=true spans the whole brain', async () => {
-    const { ctx, captured } = makeCtx(false);
-    await searchOp.handler(ctx, { query: 'x', all_sources: true });
-    expect(captured.opts?.sourceId).toBeUndefined();
-    expect(captured.opts?.sourceIds).toBeUndefined();
-  });
-
-  test('explicit source_id wins over ctx.sourceId', async () => {
-    const { ctx, captured } = makeCtx(false);
-    await searchOp.handler(ctx, { query: 'x', source_id: 'wiki' });
-    expect(captured.opts?.sourceId).toBe('wiki');
-  });
-
-  test("remote + '__all__' collapses to the caller's grant (fail-closed)", async () => {
-    const { ctx, captured } = makeCtx(true, ['wiki', 'essays']);
-    await searchOp.handler(ctx, { query: 'x', source_id: '__all__' });
-    expect(captured.opts?.sourceIds).toEqual(['wiki', 'essays']);
-  });
-
-  test('remote + out-of-grant source_id is denied', async () => {
-    const { ctx } = makeCtx(true, ['wiki']);
-    await expect(searchOp.handler(ctx, { query: 'x', source_id: 'secrets' })).rejects.toThrow(
-      /outside your granted sources/,
-    );
-  });
-});

test/voyage-response-cap.test.ts

@@ -34,7 +34,7 @@ describe('v0.31.8 — voyage Content-Length pre-check + per-item cap', () => {
     expect(source).toMatch(/MAX_VOYAGE_RESPONSE_BYTES\s*=\s*256\s*\*\s*1024\s*\*\s*1024/);
   });
 
-  test('Layer 1: Content-Length pre-check fires BEFORE the body is read (D10 OOM defense)', async () => {
+  test('Layer 1: Content-Length pre-check fires BEFORE resp.clone().json() (D10 OOM defense)', async () => {
     const source = await Bun.file(new URL('../src/core/ai/gateway.ts', import.meta.url)).text();
     // Anchor relative to the post-fetch handler block. The function declaration
     // contains an OUTBOUND request body section earlier; we want to verify
@@ -47,10 +47,8 @@ describe('v0.31.8 — voyage Content-Length pre-check + per-item cap', () => {
     // doesn't pin to comment text.
     const preCheckIdx = inboundBlock.indexOf("resp.headers.get('content-length')");
     // Use the full lvalue assignment so the match doesn't accidentally hit
-    // comment text that mentions the body read for context. (#1610 moved the
-    // read from `resp.clone().json()` to a single `resp.text()` — bun <
-    // 1.1.27 truncates clone()d bodies, oven-sh/bun#6348.)
-    const jsonParseIdx = inboundBlock.indexOf('const bodyText = await resp.text()');
+    // comment text that mentions `await resp.clone().json()` for context.
+    const jsonParseIdx = inboundBlock.indexOf('const json: any = await resp.clone().json()');
     expect(preCheckIdx).toBeGreaterThan(0);
     expect(jsonParseIdx).toBeGreaterThan(0);
     // The pre-check MUST appear before the JSON parse — otherwise the OOM