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>
test: fix CI red on #3130 — withEnv for batch-concurrency env + close resetGateway shard-order poison window
2026-08-17 10:22:34 +00:00 · 2026-07-22 11:52:30 -07:00 · 2026-07-22 11:08:07 -07:00 · 2026-07-21 14:39:33 -07:00
31 changed files with 549 additions and 766 deletions
@@ -527,9 +527,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
  process.on('SIGINT',  () => { void shutdown('SIGINT'); });

  let consecutiveErrors = 0;
-  // Parser-probe fixture warning is once-per-process, not once-per-cycle
-  // (compiled-binary installs have no source tree; don't spam the log).
-  let parserProbeFixtureWarned = false;
  // v0.37.7.0 #1162 — counter for consecutive reconnect failures.
  // Reset on every successful health probe or reconnect. Threshold
  // controlled by GBRAIN_AUTOPILOT_MAX_RECONNECT_FAILS env (default 30).
@@ -1076,36 +1073,17 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
    // loop. Probe runs even when cycleOk=false (probe may surface signal
    // explaining why the cycle is failing).
    try {
-      const { resolveProbeEnabled, resolveProbeMaxUsd, runNightlyQualityProbe } = await import('../core/cycle/nightly-quality-probe.ts');
-      // Dual-plane read: `gbrain config set` (what the doctor enable hint
-      // prints) writes the DB plane; ~/.gbrain/config.json is the fallback.
-      let dbEnabled: string | null = null;
-      let dbMaxUsd: string | null = null;
-      try {
-        dbEnabled = await engine.getConfig('autopilot.nightly_quality_probe.enabled');
-        dbMaxUsd = await engine.getConfig('autopilot.nightly_quality_probe.max_usd');
-      } catch { /* DB unavailable → file plane only */ }
-      const probeEnabled = resolveProbeEnabled(dbEnabled, cfg?.autopilot?.nightly_quality_probe?.enabled);
+      const probeEnabled = cfg?.autopilot?.nightly_quality_probe?.enabled === true;
      if (probeEnabled) {
+        const { runNightlyQualityProbe } = await import('../core/cycle/nightly-quality-probe.ts');
        const { runLongMemEvalForProbe, runCrossModalBatchForProbe } = await import('../core/cycle/nightly-probe-adapters.ts');
        const { isAvailable } = await import('../core/ai/gateway.ts');
-        const { existsSync } = await import('node:fs');
-        const { fileURLToPath } = await import('node:url');
-        const { join } = await import('node:path');
-        const maxUsd = resolveProbeMaxUsd(dbMaxUsd, cfg?.autopilot?.nightly_quality_probe?.max_usd);
-        // The committed fixture (test/fixtures/longmemeval-nightly.jsonl)
-        // lives in the gbrain PACKAGE, not the brain repo — repoPath is
-        // sync.repo_path (the user's brain), where the fixture never
-        // exists, so the probe error'd on every real install. Resolve the
-        // package root from the module location; keep repoPath as the
-        // fallback for setups that vendor the fixture into the brain repo.
-        const pkgRoot = fileURLToPath(new URL('../..', import.meta.url));
-        const fixtureAtPkgRoot = existsSync(join(pkgRoot, 'test', 'fixtures', 'longmemeval-nightly.jsonl'));
+        const maxUsd = Number(cfg?.autopilot?.nightly_quality_probe?.max_usd ?? 5);
        await runNightlyQualityProbe({
          isEnabled: () => true, // already gated above; phase re-checks for defense-in-depth
          hasEmbeddingProvider: () => isAvailable('embedding'),
          resolveMaxUsd: () => maxUsd,
-          resolveRepoRoot: () => (fixtureAtPkgRoot ? pkgRoot : repoPath ?? gbrainHomePath('.')),
+          resolveRepoRoot: () => repoPath ?? gbrainHomePath('.'),
          runLongMemEval: runLongMemEvalForProbe,
          runCrossModalBatch: runCrossModalBatchForProbe,
          now: () => new Date(),
@@ -1117,62 +1095,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
      // informational; autopilot loop continues.
    }

-    // 4.6 — Nightly conversation-parser probe (v0.41.16.0 phase module;
-    // the scheduler wire-up was deferred at ship and is added here). Same
-    // posture as 4.5: the phase owns its gates (enabled/mode-gate, LLM
-    // key), the wiring owns invocation + the audit row, and a probe
-    // failure NEVER crashes the autopilot loop. Per D10 the probe is
-    // default-ON for search.mode=tokenmax, opt-in otherwise.
-    try {
-      const { runConversationParserNightlyProbe } = await import('../core/conversation-parser/nightly-probe.ts');
-      const { logParserProbeEvent, parserProbeRanWithin } = await import('../core/audit-parser-probe.ts');
-      const { isAvailable } = await import('../core/ai/gateway.ts');
-      const { existsSync } = await import('node:fs');
-      const { fileURLToPath } = await import('node:url');
-      const { join } = await import('node:path');
-      // Flag reads dual-plane: the DB row (`gbrain config set …`) wins,
-      // ~/.gbrain/config.json is the fallback. search.mode lives on the
-      // DB plane only (mode.ts owns it).
-      let parserDbEnabled: string | null = null;
-      let dbSearchMode: string | null = null;
-      try {
-        parserDbEnabled = await engine.getConfig('autopilot.conversation_parser_probe.enabled');
-        dbSearchMode = await engine.getConfig('search.mode');
-      } catch { /* DB unavailable → file plane only */ }
-      const parserEnabled = parserDbEnabled != null
-        ? parserDbEnabled === 'true'
-        : cfg?.autopilot?.conversation_parser_probe?.enabled === true;
-      const searchMode = dbSearchMode ?? '';
-      // Fixtures are committed in the gbrain package (test/fixtures/…),
-      // NOT the brain repo — resolve from the module location. Compiled
-      // binaries carry no source tree: skip quietly instead of writing
-      // failure rows that would flip doctor to WARN on every binary install.
-      const pkgRoot = fileURLToPath(new URL('../..', import.meta.url));
-      const fixturePath = join(pkgRoot, 'test', 'fixtures', 'conversation-formats', 'all.jsonl');
-      const adversarialPath = join(pkgRoot, 'test', 'fixtures', 'conversation-formats', 'adversarial.jsonl');
-      const shouldInvoke = parserEnabled || searchMode === 'tokenmax';
-      if (shouldInvoke && existsSync(fixturePath) && existsSync(adversarialPath)) {
-        const result = await runConversationParserNightlyProbe({
-          isEnabled: () => parserEnabled,
-          searchMode: () => searchMode,
-          hasLlmKey: () => isAvailable('chat'),
-          resolveFixturePath: () => fixturePath,
-          resolveAdversarialPath: () => adversarialPath,
-          now: () => new Date(),
-          shouldSkipForRateLimit: () => parserProbeRanWithin(24 * 60 * 60 * 1000),
-        });
-        // rate_limited is a non-run: the loop ticks every few minutes, so
-        // logging every skip would flood the audit file with no-signal rows.
-        if (result.outcome !== 'rate_limited') logParserProbeEvent(result);
-      } else if (shouldInvoke && !parserProbeFixtureWarned) {
-        parserProbeFixtureWarned = true;
-        console.error(`[parser-probe] fixtures not found under ${pkgRoot}; skipping (probe needs a source-checkout install)`);
-      }
-    } catch (e) {
-      logError('autopilot.parser_probe', e);
-      // Informational, like 4.5: do NOT bump consecutiveErrors.
-    }
-
    // Wait for next cycle
    await new Promise(r => setTimeout(r, interval * 1000));
  }
@@ -2960,54 +2960,6 @@ function _resolveSyncFreshnessHours(varName: string, fallback: number): number {
 * branch (disabled / enabled-no-events / enabled-all-pass / enabled-with-failures)
 * without spinning up the audit JSONL or a real config file.
 */
-/**
- * Pure function form of the conversation_parser_probe_health check.
- * Mirrors computeNightlyQualityProbeHealthCheck: skip-with-hint when the
- * probe is off and silent, surface the last 7 days of audit events when
- * it has run, WARN on any non-pass outcome.
- *
- * `effectiveEnabled` folds the D10 mode-gate in: explicitly enabled OR
- * search.mode=tokenmax (where the probe is default-on).
- */
-export function computeConversationParserProbeHealthCheck(
-  effectiveEnabled: boolean,
-  events: ReadonlyArray<{ outcome: string; ts: string; reason?: string }>,
-): Check {
-  const name = 'conversation_parser_probe_health';
-  if (!effectiveEnabled && events.length === 0) {
-    return {
-      name,
-      status: 'ok',
-      message:
-        'disabled (opt-in; default-on only for search.mode=tokenmax). Enable with: ' +
-        '`gbrain config set autopilot.conversation_parser_probe.enabled true`',
-    };
-  }
-  if (events.length === 0) {
-    return {
-      name,
-      status: 'ok',
-      message: 'enabled but no probe events in the last 7 days (next run by autopilot; fixtures require a source-checkout install).',
-    };
-  }
-  const bad = events.filter(e => e.outcome !== 'pass');
-  const latest = events[events.length - 1]!;
-  if (bad.length > 0) {
-    return {
-      name,
-      status: 'warn',
-      message:
-        `${bad.length}/${events.length} probe run(s) in the last 7 days did not pass; ` +
-        `latest: ${latest.outcome}${latest.reason ? ` (${latest.reason})` : ''}`,
-    };
-  }
-  return {
-    name,
-    status: 'ok',
-    message: `${events.length} probe run(s) in the last 7 days, all pass (latest ${latest.ts}).`,
-  };
-}
-
 export function computeNightlyQualityProbeHealthCheck(
  probeEnabled: boolean,
  events: ReadonlyArray<{ outcome: string; ts: string; detail?: string }>,
@@ -4891,17 +4843,10 @@ export async function buildChecks(
  try {
    const { readRecentQualityProbeEvents } = await import('../core/audit-quality-probe.ts');
    const { loadConfig } = await import('../core/config.ts');
-    const { resolveProbeEnabled } = await import('../core/cycle/nightly-quality-probe.ts');
    let probeEnabled = false;
    try {
-      // Dual-plane read, matching the autopilot gate: the DB row (what the
-      // enable hint's `gbrain config set` writes) wins; file plane fallback.
-      let dbVal: string | null = null;
-      try {
-        dbVal = engine ? await engine.getConfig('autopilot.nightly_quality_probe.enabled') : null;
-      } catch { /* DB unavailable → file plane only */ }
      const cfg = loadConfig();
-      probeEnabled = resolveProbeEnabled(dbVal, (cfg as any)?.autopilot?.nightly_quality_probe?.enabled);
+      probeEnabled = Boolean((cfg as any)?.autopilot?.nightly_quality_probe?.enabled);
    } catch { /* config unavailable → treat as disabled */ }
    const events = readRecentQualityProbeEvents(7);
    const check = computeNightlyQualityProbeHealthCheck(probeEnabled, events);
@@ -5085,29 +5030,19 @@ export async function buildChecks(

  // 3d.5 v0.41.13.0 — conversation_parser_probe_health. Mode-gated
  // per D10: ON when search.mode=tokenmax, opt-in for other modes.
-  // Surfaces the last 7 days of nightly-probe audit events; warn on any
-  // non-pass outcome (fail / budget_exceeded / adversarial_false_positive).
-  // (Until the autopilot wire-up this was a hardcoded "Skipped" stub.)
-  try {
-    const { readRecentParserProbeEvents } = await import('../core/audit-parser-probe.ts');
-    let parserProbeEnabled = false;
-    try {
-      let dbVal: string | null = null;
-      let dbMode: string | null = null;
-      try {
-        dbVal = engine ? await engine.getConfig('autopilot.conversation_parser_probe.enabled') : null;
-        dbMode = engine ? await engine.getConfig('search.mode') : null;
-      } catch { /* DB unavailable → file plane only */ }
-      const { loadConfig } = await import('../core/config.ts');
-      const fileVal = (loadConfig() as any)?.autopilot?.conversation_parser_probe?.enabled;
-      const flagOn = dbVal != null ? dbVal === 'true' : fileVal === true;
-      parserProbeEnabled = flagOn || dbMode === 'tokenmax';
-    } catch { /* config unavailable → treat as disabled */ }
-    const parserEvents = readRecentParserProbeEvents(7);
-    checks.push(computeConversationParserProbeHealthCheck(parserProbeEnabled, parserEvents));
-  } catch {
-    // Best-effort; audit-log read failure shouldn't stop doctor.
-  }
+  // Surface the last 7 days of nightly-probe events; warn on FAIL /
+  // BUDGET_EXCEEDED / adversarial_false_positive.
+  //
+  // v0.41.13.0 ships the probe as opt-in (autopilot wiring deferred
+  // to T7 in the cathedral plan); this check skips with an enable
+  // hint until the probe has at least one audit event written.
+  checks.push({
+    name: 'conversation_parser_probe_health',
+    status: 'ok',
+    message:
+      'Skipped (nightly probe is opt-in; enable with ' +
+      '`gbrain config set autopilot.conversation_parser_probe.enabled true`)',
+  });

  // 3e. home_dir_in_worktree (v0.35.8.0). Walks up from `gbrainPath()`
  // looking for a `.git` directory OR file. If found, warns: `~/.gbrain/`
@@ -76,7 +76,7 @@ FLAGS:
                           dimensions (goal, depth, sourcing, specificity, useful).
  --cycles N               1-3. Default: 3 in TTY, 1 in non-TTY (T11). Each
                           cycle is 3 model calls; verdict aggregates over them.
-  --slot-a-model <id>      Override default 'openai:gpt-5.2'.
+  --slot-a-model <id>      Override default 'openai:gpt-4o'.
  --slot-b-model <id>      Override default 'anthropic:claude-opus-4-7'.
  --slot-c-model <id>      Override default 'google:gemini-1.5-pro'.
  --receipt-dir <path>     Default: gbrainPath('eval-receipts').
@@ -468,14 +468,6 @@ interface BatchRow {
  question_id: string;
  question: string;
  hypothesis: string;
-  /**
-   * Gold answer from the benchmark dataset, when the upstream eval emits
-   * it (eval-longmemeval does). Folded into the judge task so CORRECTNESS
-   * is verifiable — without it a judge panel that sees only
-   * {question, hypothesis} cannot validate a terse factual answer against
-   * a haystack it never saw.
-   */
-  answer?: string;
 }

 /**
@@ -589,7 +581,6 @@ function readBatchRows(path: string): BatchReadResult {
      question_id: typeof obj.question_id === 'string' ? obj.question_id : `line-${lineNo}`,
      question: obj.question,
      hypothesis: obj.hypothesis,
-      ...(typeof obj.answer === 'string' && obj.answer.length > 0 ? { answer: obj.answer } : {}),
    });
  }
  if (summarySkipped > 0) {
@@ -706,11 +697,7 @@ async function runBatchMode(parsed: ParsedArgs, opts: RunCrossModalOpts): Promis
      fn: async (row, idx) => {
        process.stderr.write(`[eval cross-modal batch] ${idx + 1}/${rows.length} ${row.question_id} starting...\n`);
        return await runEvalFn({
-          // With a gold answer the judges can actually verify correctness;
-          // without one they see only {question, hypothesis} and cannot.
-          task: row.answer
-            ? `${row.question}\n\nExpected answer (gold label from the benchmark dataset): ${row.answer}`
-            : row.question,
+          task: row.question,
          output: row.hypothesis,
          slug: row.question_id,
          dimensions,
@@ -33,7 +33,6 @@ import {
  type AliasMap,
 } from '../eval/longmemeval/extract.ts';
 import { extractCandidateEntities } from '../core/think/entity-extract.ts';
-import { splitProviderModelId } from '../core/model-id.ts';
 import { resolveEntitySlugWithSource, type ResolutionSource } from '../core/entities/resolve.ts';
 import { formatTrajectoryBlock } from '../core/trajectory-format.ts';

@@ -470,22 +469,14 @@ export async function runEvalLongMemEval(args: string[], runOpts: RunOpts = {}):
  });

  // Wrap Anthropic SDK so its `.messages.create` shape matches ThinkLLMClient.
-  // Same pattern as src/core/think/index.ts:247-249 — EXCEPT think's default
-  // client routes through the gateway, which parses `provider:model` recipe
-  // ids. This eval's client is a raw SDK by design (hermetic, no gateway
-  // dependency), and resolveModel returns RECIPE ids (`anthropic:claude-…`);
-  // passing one through unstripped 404s every answer/extractor call, which
-  // surfaces downstream as all-upstream_error batches in the nightly probe.
-  const toSdkModel = (m: string): string => splitProviderModelId(m).model || m;
+  // Same pattern as src/core/think/index.ts:247-249.
  const realClient = new Anthropic();
  const client: ThinkLLMClient = runOpts.client ?? {
-    create: (params, callOpts) =>
-      realClient.messages.create({ ...params, model: toSdkModel(params.model) }, callOpts),
+    create: (params, callOpts) => realClient.messages.create(params, callOpts),
  };
  // v0.40.2.0 — separate extractor client (defaults to same SDK).
  const extractorClient: ThinkLLMClient = runOpts.extractorClient ?? {
-    create: (params, callOpts) =>
-      realClient.messages.create({ ...params, model: toSdkModel(params.model) }, callOpts),
+    create: (params, callOpts) => realClient.messages.create(params, callOpts),
  };
  const trajectoryEnabled = !opts.noTrajectory;
  const extractorModel = trajectoryEnabled
@@ -760,11 +751,6 @@ async function runOneQuestion(
    // v0.40.1.0 (Track D / T2) — copy question_type into the row so the
    // by_type_summary can be rebuilt from the file on resume runs.
    question_type: q.question_type,
-    // Gold answer for downstream consumers that verify correctness (the
-    // cross-modal --batch judge folds it into the task; evaluate_qa.py
-    // ignores unknown fields). Without it a judge can't validate a terse
-    // factual hypothesis against a haystack it never saw.
-    ...(q.answer !== undefined ? { answer: q.answer } : {}),
    hypothesis,
    retrieved_session_ids: retrievedSessionIds,
    ...(recallHit !== undefined ? { recall_hit: recallHit } : {}),
@@ -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`);
  }
@@ -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)
  );
 }

@@ -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.
@@ -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'],
@@ -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()];
 }
@@ -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
@@ -1,63 +0,0 @@
-/**
- * Nightly conversation-parser probe audit trail.
- *
- * One event per REAL probe run lands in
- * `~/.gbrain/audit/parser-probe-YYYY-Www.jsonl` (ISO-week rotation via the
- * shared audit-writer primitive; honors `GBRAIN_AUDIT_DIR`).
- * Scheduler-cadence skips (`rate_limited`) are NOT logged — the autopilot
- * loop ticks every few minutes, so logging every skip would flood the
- * audit file with rows that carry no signal.
- *
- * Read by `gbrain doctor`'s `conversation_parser_probe_health` check and
- * by the autopilot wiring's 24h rate-limit gate (`parserProbeRanWithin`).
- */
-
-import { createAuditWriter } from './audit/audit-writer.ts';
-import type { NightlyProbeResult } from './conversation-parser/nightly-probe.ts';
-
-export type ParserProbeAuditEvent = NightlyProbeResult;
-
-const writer = createAuditWriter<ParserProbeAuditEvent>({
-  featureName: 'parser-probe',
-  errorLabel: 'gbrain',
-  errorMessagePrefix: 'parser-probe audit ',
-  errorTrailer: '; probe continues',
-});
-
-/** Append one parser-probe event. Best-effort; never throws. */
-export function logParserProbeEvent(event: ParserProbeAuditEvent): void {
-  writer.log(event);
-}
-
-/**
- * Read recent parser-probe events (current + previous ISO week, filtered
- * to the window). Missing files and corrupt rows are skipped silently.
- */
-export function readRecentParserProbeEvents(
-  days = 7,
-  now: Date = new Date(),
-): ParserProbeAuditEvent[] {
-  return writer.readRecent(days, now);
-}
-
-/** Exposed for tests pinning the rotation edge cases. */
-export function computeParserProbeAuditFilename(now: Date = new Date()): string {
-  return writer.computeFilename(now);
-}
-
-/**
- * 24h rate-limit gate for the autopilot wiring: true when any audited run
- * happened within `windowMs` of `now`. Only REAL outcomes are audited (see
- * module header), so a pass/fail today blocks re-runs until tomorrow while
- * scheduler-cadence skips never extend the window.
- */
-export function parserProbeRanWithin(
-  windowMs: number,
-  now: Date = new Date(),
-): boolean {
-  const cutoff = now.getTime() - windowMs;
-  return readRecentParserProbeEvents(2, now).some((ev) => {
-    const ts = Date.parse(ev.ts);
-    return Number.isFinite(ts) && ts >= cutoff;
-  });
-}
@@ -105,16 +105,6 @@ export interface GBrainConfig {
       */
      max_usd?: number;
    };
-    /**
-     * v0.41.16.0 — nightly conversation-parser probe. Per D10: default ON
-     * for `search.mode=tokenmax` brains, opt-in for conservative/balanced.
-     * ~$0.05/night with the committed fixtures × Haiku polish. Gated
-     * INSIDE the autopilot tick body, like nightly_quality_probe.
-     */
-    conversation_parser_probe?: {
-      /** Enable for non-tokenmax modes. Defaults to false. */
-      enabled?: boolean;
-    };
    /**
     * v0.42.x (#1685 GAP D) — extract_atoms backlog auto-drain. Default ON so a
     * pack-gated silent backlog never piles up unseen; daily-spend-capped so the
@@ -17,11 +17,11 @@
 * Cost: ~$0.05/night with default fixtures × Haiku polish. Bounded
 * by the active BudgetTracker the autopilot loop creates per-tick.
 *
- * Wired into the autopilot loop (step 4.6 in autopilot.ts), following
- * the same shape as `src/core/cycle/nightly-quality-probe.ts`
- * (v0.40.1.0 Track D / T6): the wiring resolves fixtures from the
- * gbrain package root, writes real outcomes to the parser-probe audit
- * trail (`audit-parser-probe.ts`), and never crashes the loop.
+ * **Wiring into the autopilot loop is deferred to a follow-up**
+ * (filed in TODOS.md). v0.41.16.0 ships the phase as a callable
+ * module so doctor + future cron drivers can invoke it; the
+ * scheduler wire-up follows the same shape as
+ * `src/core/cycle/nightly-quality-probe.ts` (v0.40.1.0 Track D / T6).
 *
 * Test seam: all dependencies are injected via NightlyProbeDeps so
 * unit tests don't touch real LLMs or real fixtures.
@@ -44,12 +44,7 @@ export const DEFAULT_DIMENSIONS: string[] = [
 * `--slot-a-model`, `--slot-b-model`, `--slot-c-model` on the CLI.
 */
 export const DEFAULT_SLOTS: SlotConfig[] = [
-  // Every default MUST be listed in its recipe's chat touchpoint (pinned by
-  // test/cross-modal-default-slots.test.ts) — `openai:gpt-4o` sat here after
-  // the OpenAI recipe dropped it, so slot A errored "not listed for OpenAI
-  // chat" on every install and the 3-slot panel could never reach its
-  // 2-model quorum without a Google key (verdict: permanently inconclusive).
-  { id: 'A', model: 'openai:gpt-5.2' },
+  { id: 'A', model: 'openai:gpt-4o' },
  { id: 'B', model: 'anthropic:claude-opus-4-7' },
  { id: 'C', model: 'google:gemini-1.5-pro' },
 ];
@@ -70,28 +70,6 @@ export async function runLongMemEvalForProbe(args: LongMemEvalProbeArgs): Promis
 * the batch input) or unparseable (cross-modal wrote garbage). Both
 * cases are paste-ready in the error message.
 */
-/**
- * QA-shaped judge dimensions for the nightly probe. The batch judge's
- * DEFAULT_DIMENSIONS rubric (DEPTH / SOURCING / SPECIFICITY / …) is built
- * for rich agent responses; LongMemEval hypotheses are deliberately terse
- * factual answers ("in widget-co") that can never score ≥7 on DEPTH or
- * SOURCING — so with the default rubric the probe FAILs every night even
- * when retrieval + answering are perfectly healthy. The probe owns its
- * invocation of the eval tool and passes dimensions matching the
- * fixture's QA shape instead.
- *
- * NOTE: the `--dimensions` CLI flag splits on commas, so these dimension
- * descriptions must stay comma-free.
- */
-export const PROBE_QA_DIMENSIONS: string[] = [
-  // No faithfulness/grounding dimension on purpose: the judge never sees
-  // the haystack, so any accurate detail beyond the terse gold label reads
-  // as "invented" and correct answers fail (verified empirically — a
-  // correct "before + dates" answer scored 4/10 on such a dimension).
-  'CORRECTNESS — Does the hypothesis state the same fact as the expected answer? A terse direct answer is ideal.',
-  'DIRECTNESS — Does it answer THIS question without hedging or padding or answering something else?',
-];
-
 export async function runCrossModalBatchForProbe(
  args: CrossModalProbeArgs,
 ): Promise<{ exitCode: number; summary: CrossModalBatchSummary }> {
@@ -103,8 +81,6 @@ export async function runCrossModalBatchForProbe(
    args.summaryPath,
    '--max-usd',
    String(args.maxUsd),
-    '--dimensions',
-    PROBE_QA_DIMENSIONS.join(','),
    '--yes',
    '--json',
  ]);
@@ -62,42 +62,6 @@ export interface NightlyProbeDeps {
  now: () => Date;
 }

-/**
- * Dual-plane flag resolution (same precedent as `mcp.publish_skills` in
- * serve-http.ts): the DB config row — what `gbrain config set` writes —
- * wins when present; the file plane (~/.gbrain/config.json) is the
- * fallback. Doctor's paste-ready enable hint says `gbrain config set
- * autopilot.nightly_quality_probe.enabled true`, so the gate MUST read
- * the DB plane — a file-only read turns that hint into a silent no-op.
- */
-export function resolveProbeEnabled(
-  dbVal: string | null | undefined,
-  fileVal: unknown,
-): boolean {
-  if (dbVal != null) return dbVal === 'true';
-  return fileVal === true;
-}
-
-/**
- * Same dual-plane rule for the per-run cost cap. Malformed or negative
- * values on either plane fall through to the next plane / the default.
- */
-export function resolveProbeMaxUsd(
-  dbVal: string | null | undefined,
-  fileVal: unknown,
-  fallback: number = DEFAULT_MAX_USD,
-): number {
-  if (dbVal != null) {
-    const n = Number(dbVal);
-    if (Number.isFinite(n) && n >= 0) return n;
-  }
-  if (fileVal != null) {
-    const n = Number(fileVal);
-    if (Number.isFinite(n) && n >= 0) return n;
-  }
-  return fallback;
-}
-
 /**
 * Pure function: decide whether the probe should run given the audit
 * history. Returns reason when skipping.
@@ -137,17 +101,21 @@ export async function runNightlyQualityProbe(deps: NightlyProbeDeps): Promise<Ni
    return { outcome: 'disabled', exit_code: 0, detail: 'feature flag off' };
  }

-  // 24h rate limit — skip WITHOUT an audit row. The autopilot loop invokes
-  // the probe every cycle (~5-10 min), so all but one invocation per day
-  // lands here; logging each skip floods the audit file (~hundreds of
-  // rows/day) and — because doctor treats any non-pass outcome as bad
-  // signal — flips nightly_quality_probe_health to a permanent WARN the
-  // moment the probe is enabled. A skip is a non-event: the real runs are
-  // the signal, and their rows are what gates the next 24h window.
+  // 24h rate limit — skip + audit "rate_limited".
  const now = deps.now();
  const recent = readRecentQualityProbeEvents(2, now); // 2-day window is enough for 24h check
  const decision = shouldRunNightly(now, recent);
  if (!decision.run) {
+    logQualityProbeEvent({
+      outcome: 'rate_limited',
+      exit_code: 0,
+      pass_count: 0,
+      fail_count: 0,
+      inconclusive_count: 0,
+      error_count: 0,
+      est_cost_usd: 0,
+      detail: 'already ran within 24h window',
+    });
    return { outcome: 'rate_limited', exit_code: 0, detail: 'already ran within 24h' };
  }

@@ -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;
 }

@@ -75,11 +75,6 @@ export const CANONICAL_PRICING: Record<string, ModelPricing> = {
  'openai:gpt-4o':                        { input:  2.50, output: 10.00 },
  'openai:gpt-4o-mini':                   { input:  0.15, output:  0.60 },
  'openai:gpt-5':                         { input:  5.00, output: 20.00 },
-  // gpt-5.2: rates from the OpenAI recipe chat touchpoint (verified
-  // 2026-04-20). Needed here because it's the cross-modal DEFAULT_SLOTS
-  // slot-A model — without a canonical entry estimateCost silently drops
-  // slot A from the --max-usd pre-flight and est_cost_usd audit rows.
-  'openai:gpt-5.2':                       { input:  1.25, output: 10.00 },
  'openai:gpt-5.5':                       { input:  4.00, output: 16.00 },

  // ── Google ─────────────────────────────────────────────────────────────
@@ -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();
  });
 });
@@ -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', () => {
@@ -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
@@ -1,102 +0,0 @@
-/**
- * Tests for the parser-probe audit trail + the 24h rate-limit gate.
- *
- * Uses GBRAIN_AUDIT_DIR override pointed at a tmpdir for hermeticity
- * (same pattern as audit-slug-fallback.serial.test.ts). Serial because
- * the env override is process-global.
- */
-import { afterEach, beforeEach, describe, expect, test } from 'bun:test';
-import { mkdtempSync, rmSync, readdirSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-
-import {
-  computeParserProbeAuditFilename,
-  logParserProbeEvent,
-  parserProbeRanWithin,
-  readRecentParserProbeEvents,
-  type ParserProbeAuditEvent,
-} from '../src/core/audit-parser-probe.ts';
-
-let auditDir: string;
-let savedEnv: string | undefined;
-
-beforeEach(() => {
-  auditDir = mkdtempSync(join(tmpdir(), 'parser-probe-audit-'));
-  savedEnv = process.env.GBRAIN_AUDIT_DIR;
-  process.env.GBRAIN_AUDIT_DIR = auditDir;
-});
-
-afterEach(() => {
-  if (savedEnv === undefined) delete process.env.GBRAIN_AUDIT_DIR;
-  else process.env.GBRAIN_AUDIT_DIR = savedEnv;
-  rmSync(auditDir, { recursive: true, force: true });
-});
-
-function makeEvent(overrides: Partial<ParserProbeAuditEvent> = {}): ParserProbeAuditEvent {
-  return {
-    schema_version: 1,
-    ts: new Date().toISOString(),
-    outcome: 'pass',
-    fixtures_total: 12,
-    fixtures_passed: 12,
-    recall_mean: 0.98,
-    participants_recall_mean: 0.97,
-    adversarial_false_positives: 0,
-    failed_fixture_ids: [],
-    ...overrides,
-  };
-}
-
-describe('parser-probe audit trail', () => {
-  test('log + readRecent round-trip', () => {
-    logParserProbeEvent(makeEvent({ outcome: 'fail', reason: '2 fixture(s) failed' }));
-    const events = readRecentParserProbeEvents(7);
-    expect(events.length).toBe(1);
-    expect(events[0]!.outcome).toBe('fail');
-    expect(events[0]!.reason).toBe('2 fixture(s) failed');
-    const files = readdirSync(auditDir);
-    expect(files.length).toBe(1);
-    expect(files[0]).toMatch(/^parser-probe-\d{4}-W\d{2}\.jsonl$/);
-  });
-
-  test('filename uses ISO-week rotation with the parser-probe prefix', () => {
-    // Year-boundary edge pinned by the shared writer's own tests; here we
-    // pin the prefix wiring.
-    expect(computeParserProbeAuditFilename(new Date('2026-07-06T12:00:00Z'))).toBe(
-      'parser-probe-2026-W28.jsonl',
-    );
-  });
-
-  test('readRecent filters by window', () => {
-    const old = new Date(Date.now() - 10 * 86400000).toISOString();
-    logParserProbeEvent(makeEvent({ ts: old }));
-    expect(readRecentParserProbeEvents(7).length).toBe(0);
-  });
-});
-
-describe('parserProbeRanWithin — 24h rate-limit gate', () => {
-  const DAY_MS = 24 * 60 * 60 * 1000;
-
-  test('false when no runs are audited', () => {
-    expect(parserProbeRanWithin(DAY_MS)).toBe(false);
-  });
-
-  test('true when a run landed within the window', () => {
-    logParserProbeEvent(makeEvent({ ts: new Date(Date.now() - 60_000).toISOString() }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(true);
-  });
-
-  test('false when the last run is older than the window', () => {
-    logParserProbeEvent(makeEvent({ ts: new Date(Date.now() - 25 * 3600_000).toISOString() }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(false);
-  });
-
-  test('non-pass outcomes also hold the window (mirrors quality-probe semantics)', () => {
-    logParserProbeEvent(makeEvent({
-      outcome: 'no_embedding_key',
-      ts: new Date(Date.now() - 3600_000).toISOString(),
-    }));
-    expect(parserProbeRanWithin(DAY_MS)).toBe(true);
-  });
-});
@@ -31,15 +31,10 @@ describe('autopilot wiring: nightly quality probe', () => {
    expect(SOURCE).toContain(`runCrossModalBatchForProbe`);
  });

-  test('feature flag gate present: dual-plane read (DB row wins, file plane fallback)', () => {
+  test('feature flag gate present: cfg.autopilot.nightly_quality_probe.enabled', () => {
    // Per D10: the scheduler ONLY checks the feature flag. The 24h rate-limit
    // lives inside runNightlyQualityProbe itself (no scheduler-side precheck).
-    // The flag resolves through resolveProbeEnabled so `gbrain config set
-    // autopilot.nightly_quality_probe.enabled true` (the doctor hint, DB
-    // plane) and ~/.gbrain/config.json (file plane) BOTH work — a file-only
-    // read made the printed hint a silent no-op.
-    expect(SOURCE).toContain(`getConfig('autopilot.nightly_quality_probe.enabled')`);
-    expect(SOURCE).toMatch(/resolveProbeEnabled\(dbEnabled,\s*cfg\?\.autopilot\?\.nightly_quality_probe\?\.enabled\)/);
+    expect(SOURCE).toContain(`nightly_quality_probe?.enabled === true`);
  });

  test('NO scheduler-side rate-limit check (D10 simplification)', () => {
@@ -69,23 +64,12 @@ describe('autopilot wiring: nightly quality probe', () => {
    expect(SOURCE).toContain(`now:`);
  });

-  test('resolveRepoRoot prefers the gbrain package root (committed fixture home), not the brain repoPath', () => {
-    // The DI harness in nightly-quality-probe.test.ts passes process.cwd()
-    // (= the gbrain repo in CI), which papered over the wiring passing
-    // repoPath (= sync.repo_path, the user's BRAIN repo, where the fixture
-    // never exists). Pin the package-root resolution + existence check.
-    expect(SOURCE).toMatch(/fileURLToPath\(new URL\('\.\.\/\.\.', import\.meta\.url\)\)/);
-    expect(SOURCE).toContain(`'longmemeval-nightly.jsonl'`);
-    expect(SOURCE).toMatch(/fixtureAtPkgRoot \? pkgRoot : repoPath/);
-  });
-
  test('hasEmbeddingProvider reads from gateway.isAvailable("embedding") (codex round-2 #12 — in-process, not subprocess)', () => {
    expect(SOURCE).toContain(`isAvailable('embedding')`);
    expect(SOURCE).toContain(`gateway`);
  });

-  test('max_usd resolves dual-plane (default = 5 pinned by resolveProbeMaxUsd unit tests)', () => {
-    expect(SOURCE).toContain(`getConfig('autopilot.nightly_quality_probe.max_usd')`);
-    expect(SOURCE).toMatch(/resolveProbeMaxUsd\(dbMaxUsd,\s*cfg\?\.autopilot\?\.nightly_quality_probe\?\.max_usd\)/);
+  test('max_usd default = 5 when config unset (matches plan default per D10)', () => {
+    expect(SOURCE).toMatch(/max_usd\s*\?\?\s*5/);
  });
 });
@@ -1,77 +0,0 @@
-/**
- * Source-shape regression tests for the autopilot wiring of
- * `runConversationParserNightlyProbe` (step 4.6).
- *
- * Same rationale as autopilot-nightly-probe-wiring.test.ts: the loop is
- * hard to drive end-to-end, so these pin the structural protections —
- * the dual-plane flag read, the D10 tokenmax mode-gate, the package-root
- * fixture resolution, the audit-flood guard, and the try/catch posture.
- *
- * The probe's own gate/scoring logic is pinned by the module's unit
- * tests; the audit trail by audit-parser-probe.serial.test.ts.
- */
-
-import { describe, test, expect } from 'bun:test';
-import { readFileSync } from 'node:fs';
-import { resolve } from 'node:path';
-
-const AUTOPILOT_SRC = resolve('src/commands/autopilot.ts');
-const SOURCE = readFileSync(AUTOPILOT_SRC, 'utf-8');
-
-describe('autopilot wiring: conversation-parser probe', () => {
-  test('invokes the phase module and the audit trail', () => {
-    expect(SOURCE).toContain(`runConversationParserNightlyProbe`);
-    expect(SOURCE).toContain(`conversation-parser/nightly-probe`);
-    expect(SOURCE).toContain(`logParserProbeEvent`);
-    expect(SOURCE).toContain(`audit-parser-probe`);
-  });
-
-  test('flag reads dual-plane: DB row (gbrain config set) wins, file plane fallback', () => {
-    expect(SOURCE).toContain(`getConfig('autopilot.conversation_parser_probe.enabled')`);
-    expect(SOURCE).toContain(`cfg?.autopilot?.conversation_parser_probe?.enabled === true`);
-  });
-
-  test('D10 mode-gate present: tokenmax brains run the probe by default', () => {
-    expect(SOURCE).toMatch(/parserEnabled \|\| searchMode === 'tokenmax'/);
-  });
-
-  test('fixtures resolve from the gbrain package root, NOT the brain repoPath', () => {
-    // The committed fixtures live in the gbrain source tree; resolving
-    // them against sync.repo_path would point into the user's brain repo.
-    expect(SOURCE).toMatch(/fileURLToPath\(new URL\('\.\.\/\.\.', import\.meta\.url\)\)/);
-    expect(SOURCE).toContain(`'conversation-formats', 'all.jsonl'`);
-    expect(SOURCE).toContain(`'conversation-formats', 'adversarial.jsonl'`);
-  });
-
-  test('missing fixtures skip quietly (no audit row, once-per-process stderr note)', () => {
-    // Compiled-binary installs carry no source tree; writing failure rows
-    // would flip doctor to WARN on every binary install.
-    expect(SOURCE).toContain(`parserProbeFixtureWarned`);
-  });
-
-  test('rate_limited outcomes are NOT audit-logged (flood guard)', () => {
-    expect(SOURCE).toMatch(/outcome !== 'rate_limited'\) logParserProbeEvent\(result\)/);
-  });
-
-  test('rate-limit gate delegates to the audit module, not inline event reads', () => {
-    expect(SOURCE).toContain(`parserProbeRanWithin(24 * 60 * 60 * 1000)`);
-  });
-
-  test('LLM-key gate reads gateway.isAvailable("chat") in-process', () => {
-    expect(SOURCE).toContain(`isAvailable('chat')`);
-  });
-
-  test('probe call wrapped in try/catch that does NOT bump consecutiveErrors', () => {
-    expect(SOURCE).toMatch(/catch[\s\S]*?autopilot\.parser_probe[\s\S]*?do NOT bump consecutiveErrors/);
-  });
-
-  test('DI shape: the exact 7 fields of the parser probe NightlyProbeDeps', () => {
-    expect(SOURCE).toContain(`isEnabled:`);
-    expect(SOURCE).toContain(`searchMode:`);
-    expect(SOURCE).toContain(`hasLlmKey:`);
-    expect(SOURCE).toContain(`resolveFixturePath:`);
-    expect(SOURCE).toContain(`resolveAdversarialPath:`);
-    expect(SOURCE).toContain(`shouldSkipForRateLimit:`);
-    expect(SOURCE).toContain(`now:`);
-  });
-});
@@ -1,47 +0,0 @@
-/**
- * Consistency guard: every cross-modal DEFAULT_SLOTS model must be listed
- * in its recipe's chat touchpoint. `openai:gpt-4o` drifted out of the
- * OpenAI recipe while remaining the slot-A default — the gateway then
- * rejected slot A ("not listed for OpenAI chat") on every install, and the
- * 3-slot judge panel could never reach its 2-model quorum without a Google
- * key, pinning every batch verdict at inconclusive (which the nightly
- * quality probe surfaces as a doctor WARN).
- */
-import { describe, expect, test } from 'bun:test';
-
-import { DEFAULT_SLOTS } from '../src/core/cross-modal-eval/runner.ts';
-import { getRecipe } from '../src/core/ai/recipes/index.ts';
-import { splitProviderModelId } from '../src/core/model-id.ts';
-import { canonicalLookup } from '../src/core/model-pricing.ts';
-
-describe('cross-modal DEFAULT_SLOTS ↔ recipe consistency', () => {
-  test('every default slot model is listed in its recipe chat touchpoint', () => {
-    for (const slot of DEFAULT_SLOTS) {
-      const { provider, model } = splitProviderModelId(slot.model);
-      expect(provider).not.toBeNull();
-      const recipe = getRecipe(provider!);
-      expect(recipe, `slot ${slot.id}: unknown recipe "${provider}"`).toBeDefined();
-      const chatModels = recipe!.touchpoints.chat?.models ?? [];
-      expect(
-        chatModels,
-        `slot ${slot.id}: "${model}" not listed for ${provider} chat — the judge slot can never run`,
-      ).toContain(model);
-    }
-  });
-
-  test('every default slot model has a canonical pricing entry', () => {
-    // Without one, estimateCost silently drops the slot from the
-    // --max-usd pre-flight and est_cost_usd audit rows (~1/3 under-count).
-    for (const slot of DEFAULT_SLOTS) {
-      expect(
-        canonicalLookup(slot.model),
-        `slot ${slot.id}: "${slot.model}" missing from CANONICAL_PRICING`,
-      ).toBeDefined();
-    }
-  });
-
-  test('slots span three distinct providers (uncorrelated blind spots)', () => {
-    const providers = new Set(DEFAULT_SLOTS.map(s => splitProviderModelId(s.model).provider));
-    expect(providers.size).toBe(3);
-  });
-});
@@ -1,51 +0,0 @@
-/**
- * Tests for computeConversationParserProbeHealthCheck — the pure function
- * behind doctor's conversation_parser_probe_health check, which replaced
- * the v0.41.13.0 hardcoded "Skipped" stub when the autopilot wiring
- * landed. Mirrors the branch coverage style of the quality-probe check.
- */
-import { describe, expect, test } from 'bun:test';
-
-import { computeConversationParserProbeHealthCheck } from '../src/commands/doctor.ts';
-
-const ev = (outcome: string, reason?: string, ts = new Date().toISOString()) => ({
-  outcome,
-  ts,
-  ...(reason !== undefined ? { reason } : {}),
-});
-
-describe('computeConversationParserProbeHealthCheck', () => {
-  test('disabled + no events → ok with paste-ready enable hint', () => {
-    const check = computeConversationParserProbeHealthCheck(false, []);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('gbrain config set autopilot.conversation_parser_probe.enabled true');
-  });
-
-  test('enabled + no events yet → ok, next run by autopilot', () => {
-    const check = computeConversationParserProbeHealthCheck(true, []);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('no probe events');
-  });
-
-  test('disabled flag but events exist (tokenmax mode-gate ran it) → events win over the hint', () => {
-    const check = computeConversationParserProbeHealthCheck(false, [ev('pass')]);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('all pass');
-  });
-
-  test('any non-pass outcome in the window → warn, latest surfaced with reason', () => {
-    const check = computeConversationParserProbeHealthCheck(true, [
-      ev('pass'),
-      ev('adversarial_false_positive', '1 adversarial fixture(s) parsed to non-empty'),
-    ]);
-    expect(check.status).toBe('warn');
-    expect(check.message).toContain('adversarial_false_positive');
-    expect(check.message).toContain('parsed to non-empty');
-  });
-
-  test('all pass → ok with run count', () => {
-    const check = computeConversationParserProbeHealthCheck(true, [ev('pass'), ev('pass')]);
-    expect(check.status).toBe('ok');
-    expect(check.message).toContain('2 probe run(s)');
-  });
-});
@@ -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);
+  });
+});
@@ -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);
@@ -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);
+  });
+});
@@ -1,74 +0,0 @@
-// Regression test for the nightly-quality-probe config-plane split-brain.
-//
-// The doctor check prints a paste-ready enable hint — `gbrain config set
-// autopilot.nightly_quality_probe.enabled true` — which writes the DB config
-// plane. But both the autopilot gate and the doctor check used to read ONLY
-// the file plane (~/.gbrain/config.json via loadConfig), so following the
-// hint was a silent no-op: the probe never ran and doctor kept reporting
-// "disabled (opt-in)".
-//
-// resolveProbeEnabled / resolveProbeMaxUsd pin the dual-plane rule (same
-// precedent as `mcp.publish_skills` in serve-http.ts): DB row wins when
-// present, file plane is the fallback.
-import { describe, expect, test } from 'bun:test';
-
-import {
-  resolveProbeEnabled,
-  resolveProbeMaxUsd,
-} from '../src/core/cycle/nightly-quality-probe.ts';
-
-describe('resolveProbeEnabled — dual-plane flag resolution', () => {
-  test('DB plane "true" enables regardless of file plane (the doctor hint path)', () => {
-    expect(resolveProbeEnabled('true', undefined)).toBe(true);
-    expect(resolveProbeEnabled('true', false)).toBe(true);
-  });
-
-  test('explicit DB "false" wins over file-plane true (config set off sticks)', () => {
-    expect(resolveProbeEnabled('false', true)).toBe(false);
-  });
-
-  test('file plane is the fallback when no DB row exists', () => {
-    expect(resolveProbeEnabled(null, true)).toBe(true);
-    expect(resolveProbeEnabled(undefined, true)).toBe(true);
-    expect(resolveProbeEnabled(null, undefined)).toBe(false);
-    expect(resolveProbeEnabled(null, false)).toBe(false);
-  });
-
-  test('file plane stays strict boolean — string "true" in config.json does not enable', () => {
-    // Matches the pre-fix autopilot gate (`=== true`); the doctor check used
-    // Boolean(...) and could disagree with autopilot on a string value.
-    // Both call sites now share this helper, so they can no longer diverge.
-    expect(resolveProbeEnabled(null, 'true')).toBe(false);
-    expect(resolveProbeEnabled(null, 1)).toBe(false);
-  });
-
-  test('non-"true" DB strings are off (mcp.publish_skills semantics)', () => {
-    expect(resolveProbeEnabled('1', true)).toBe(false);
-    expect(resolveProbeEnabled('yes', true)).toBe(false);
-    expect(resolveProbeEnabled('', true)).toBe(false);
-  });
-});
-
-describe('resolveProbeMaxUsd — dual-plane cost cap resolution', () => {
-  test('DB plane wins when parseable', () => {
-    expect(resolveProbeMaxUsd('2.5', 10)).toBe(2.5);
-    expect(resolveProbeMaxUsd('0', 10)).toBe(0);
-  });
-
-  test('malformed or negative DB value falls through to file plane', () => {
-    expect(resolveProbeMaxUsd('banana', 3)).toBe(3);
-    expect(resolveProbeMaxUsd('-1', 3)).toBe(3);
-  });
-
-  test('file plane used when no DB row; default when both absent/invalid', () => {
-    expect(resolveProbeMaxUsd(null, 7)).toBe(7);
-    expect(resolveProbeMaxUsd(null, '4')).toBe(4);
-    expect(resolveProbeMaxUsd(null, undefined)).toBe(5);
-    expect(resolveProbeMaxUsd(null, 'banana')).toBe(5);
-    expect(resolveProbeMaxUsd(undefined, -2)).toBe(5);
-  });
-
-  test('explicit fallback override is honored', () => {
-    expect(resolveProbeMaxUsd(null, undefined, 12)).toBe(12);
-  });
-});
@@ -132,20 +132,17 @@ describe('runNightlyQualityProbe (DI stub harness)', () => {
    });
  });

-  test('enabled + recent run within 24h → outcome: rate_limited, NO audit row', async () => {
+  test('enabled + recent run within 24h → outcome: rate_limited', async () => {
    // Pre-seed a recent audit event by running the probe once first.
    await withEnv({ GBRAIN_AUDIT_DIR: auditTmp }, async () => {
      // First run succeeds.
      await runNightlyQualityProbe(makeDeps());
-      // Second run, same hour → rate_limited. A skip is a non-event: the
-      // autopilot loop invokes the probe every cycle (~5-10 min), so
-      // logging each skip would flood the audit file and flip doctor's
-      // any-non-pass-is-bad filter to a permanent WARN.
+      // Second run, same hour → rate_limited.
      const r2 = await runNightlyQualityProbe(makeDeps());
      expect(r2.outcome).toBe('rate_limited');
      const events = await readEvents();
-      expect(events.length).toBe(1);
-      expect(events[0].outcome).toBe('pass');
+      expect(events.length).toBe(2);
+      expect(events[1].outcome).toBe('rate_limited');
    });
  });