fix(chunker): estimated-token hard cap — URL-dense/CJK-fallback chunks overflow strict embedding-server token limits

Takeover of #2847 (rebased onto current master). Fixes #2826.

- cjk.ts: estimateEmbeddingTokens() — conservative per-char-class token
  estimate (CJK 1.0, other 0.75, whitespace 0.1 per code unit).
- recursive.ts (MARKDOWN_CHUNKER_VERSION 3→4): countWords floored at
  ceil(nonWhitespaceChars/6); capByEstimatedTokens() final pass with
  ChunkOptions.maxTokens (default 1500).
- code.ts (CHUNKER_VERSION 4→5): capCodeChunks() applies the same cap to
  AST-path chunks that splitLargeNode can't subdivide.

Co-authored-by: paul-0320 <paul-0320@users.noreply.github.com>

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

src/commands/import.ts

@@ -170,14 +170,10 @@ 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, 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;
+  const { parseWorkers } = await import('../core/sync-concurrency.ts');
+  let workerCount: number;
   try {
-    workerCount = parseWorkers(workersArg ?? undefined);
+    workerCount = parseWorkers(workersArg ?? undefined) ?? 1;
   } catch (e) {
     console.error(e instanceof Error ? e.message : String(e));
     process.exit(1);
@@ -256,9 +252,8 @@ export async function runImport(
   }
   const files = resumeFilter(allFiles, dir, completed);
 
-  // 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);
+  // Determine actual worker count
+  const actualWorkers = workerCount > 1 ? workerCount : 1;
   if (actualWorkers > 1) {
     console.log(`Using ${actualWorkers} parallel workers`);
   }

src/core/ai/gateway.ts

@@ -1513,21 +1513,12 @@ 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 / 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,
-      )
+  // 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)
     : [truncated];
 
   const allEmbeddings: Float32Array[] = [];
@@ -1577,9 +1568,6 @@ 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.
  */
@@ -1587,17 +1575,15 @@ 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 || current.length >= maxCount)) {
+    if (current.length > 0 && currentTokens + estTokens > budgetTokens) {
       batches.push(current);
       current = [];
       currentTokens = 0;
@@ -1623,11 +1609,7 @@ 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) ||
-    // 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)
+    /max.*tokens.*per.*request/i.test(msg)
   );
 }
 

src/core/ai/recipes/dashscope.ts

@@ -31,10 +31,6 @@ 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

@@ -16,15 +16,6 @@ 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

@@ -58,8 +58,5 @@ export function getRecipe(id: string): Recipe | undefined {
 }
 
 export function listRecipes(): Recipe[] {
-  // 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()];
+  return [...ALL];
 }

src/core/ai/types.ts

@@ -46,16 +46,6 @@ 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/chunkers/code.ts

@@ -18,8 +18,9 @@
  * at runtime.
  */
 
-import { chunkText as recursiveChunk } from './recursive.ts';
+import { chunkText as recursiveChunk, capByEstimatedTokens, DEFAULT_MAX_EST_TOKENS } from './recursive.ts';
 import { buildQualifiedName } from './qualified-names.ts';
+import { estimateEmbeddingTokens } from '../cjk.ts';
 
 // Embed the tree-sitter runtime + per-language grammars as files.
 // `with { type: 'file' }` returns a path (string) at runtime. Bun bundles
@@ -111,7 +112,15 @@ import G_ZIG from '../../assets/wasm/grammars/tree-sitter-zig.wasm' with { type:
 // chunks get the new columns populated. Without this, the v28 backfill
 // gives every existing chunk a search_vector but subsequent Layer 5 AST
 // work would silently no-op.
-export const CHUNKER_VERSION = 4;
+//
+// v5: estimated-token hard cap on AST-path chunks (capCodeChunks). A node
+// splitLargeNode can't subdivide (giant single-statement function, huge
+// literal) previously shipped WHOLE regardless of size and could overflow
+// strict per-request embedding-token limits (local llama-server crashes
+// past ~2,050 tokens, measured). Mirrors the markdown
+// chunker's v4 cap; fallback-path chunks are already capped inside
+// recursiveChunk.
+export const CHUNKER_VERSION = 5;
 
 // Lazy-loaded tree-sitter module (v0.22.x API: Parser is default export)
 let Parser: typeof import('web-tree-sitter') | null = null;
@@ -708,7 +717,7 @@ export async function chunkCodeTextFull(
     if (chunks.length === 0) {
       return { chunks: fallbackChunks(source, filePath, language, opts), edges: rawEdges };
     }
-    return { chunks: mergeSmallSiblings(chunks, chunkTarget), edges: rawEdges };
+    return { chunks: capCodeChunks(mergeSmallSiblings(chunks, chunkTarget)), edges: rawEdges };
   } catch {
     return { chunks: fallbackChunks(source, filePath, language, opts), edges: [] };
   } finally {
@@ -791,6 +800,33 @@ function mergeSmallSiblings(chunks: CodeChunk[], chunkTarget: number): CodeChunk
   return merged;
 }
 
+/**
+ * v5 final safety pass for AST-path chunks: split any chunk whose
+ * ESTIMATED embedding tokens (conservative per-char-class heuristic,
+ * cjk.ts) exceed DEFAULT_MAX_EST_TOKENS. Reaches chunks the AST logic
+ * can't subdivide — splitLargeNode returns [] for nodes with < 2 body
+ * children (giant single-statement functions, huge literals), which
+ * previously shipped whole at any size.
+ *
+ * Split pieces inherit the source chunk's metadata verbatim; start/end
+ * lines become approximate for pieces after the first. Acceptable —
+ * these chunks exist for embedding + retrieval, and the alternative was
+ * an embedding request the server rejects (or worse, crashes on).
+ */
+function capCodeChunks(chunks: CodeChunk[]): CodeChunk[] {
+  if (chunks.every((c) => estimateEmbeddingTokens(c.text) <= DEFAULT_MAX_EST_TOKENS)) {
+    return chunks;
+  }
+  const out: CodeChunk[] = [];
+  for (const c of chunks) {
+    const pieces = capByEstimatedTokens(c.text, DEFAULT_MAX_EST_TOKENS);
+    for (const piece of pieces) {
+      out.push({ ...c, text: piece, index: out.length, metadata: { ...c.metadata } });
+    }
+  }
+  return out;
+}
+
 function buildMergedChunk(group: CodeChunk[], index: number): CodeChunk {
   const first = group[0]!;
   const last = group[group.length - 1]!;

src/core/chunkers/recursive.ts

@@ -17,7 +17,13 @@
  * Lossless invariant: non-overlapping portions reassemble to original.
  */
 
-import { countCJKAwareWords, CJK_SENTENCE_DELIMITERS, CJK_CLAUSE_DELIMITERS } from '../cjk.ts';
+import {
+  countCJKAwareWords,
+  CJK_SENTENCE_DELIMITERS,
+  CJK_CLAUSE_DELIMITERS,
+  charEmbedTokenWeight,
+  estimateEmbeddingTokens,
+} from '../cjk.ts';
 
 /**
  * Markdown chunker version. Folded into the per-page chunker_version column
@@ -33,8 +39,20 @@ import { countCJKAwareWords, CJK_SENTENCE_DELIMITERS, CJK_CLAUSE_DELIMITERS } fr
  * re-embed (not re-chunk) so existing pages pick up the wrapper on the
  * post-upgrade reembed sweep. See
  * `src/core/contextual-retrieval-service.ts`.
+ *
+ * v4: estimated-token hard cap + whitespace-word undercount fix. The word
+ * pipeline counted a 150-char URL as ONE whitespace word, so URL/phone/
+ * email-dense docs (CJK density < 0.30 → whitespace fallback) produced
+ * 3-4K-char chunks that overflow strict per-request embedding-token
+ * limits (measured: local llama-server crashes past ~2,050 tokens; URL
+ * soup tokenizes at ~1.6 chars/token). Two changes:
+ *   1. countWords() floors the count at ceil(nonWhitespaceChars/6) so a
+ *      URL counts roughly per-character, not as one word.
+ *   2. capByEstimatedTokens() final pass guarantees every chunk fits
+ *      `maxTokens` (default 1500) under a conservative per-char-class
+ *      token estimate, regardless of how word counting misjudged it.
  */
-export const MARKDOWN_CHUNKER_VERSION = 3;
+export const MARKDOWN_CHUNKER_VERSION = 4;
 
 const DELIMITERS: string[][] = [
   ['\n\n'],                          // L0: paragraphs
@@ -48,8 +66,20 @@ export interface ChunkOptions {
   chunkSize?: number;    // target words per chunk (default 300)
   chunkOverlap?: number; // overlap words (default 50)
   maxChars?: number;     // hard cap on any chunk's char length (default 6000)
+  /**
+   * v4: hard cap on any chunk's ESTIMATED embedding tokens (default 1500).
+   * Estimate = conservative per-char-class weights (see cjk.ts
+   * estimateEmbeddingTokens) — deliberately high, so the real tokenizer
+   * count stays below this value. Default leaves headroom for the
+   * contextual-retrieval wrapper (≤ ~630 chars) under a ~2,050-token
+   * per-request embedding server limit.
+   */
+  maxTokens?: number;
 }
 
+/** v4 default for ChunkOptions.maxTokens — see the field doc above. */
+export const DEFAULT_MAX_EST_TOKENS = 1500;
+
 export interface TextChunk {
   text: string;
   index: number;
@@ -73,6 +103,7 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
   const chunkSize = opts?.chunkSize || 300;
   const chunkOverlap = opts?.chunkOverlap || 50;
   const maxChars = opts?.maxChars || 6000;
+  const maxTokens = opts?.maxTokens || DEFAULT_MAX_EST_TOKENS;
 
   if (!text || text.trim().length === 0) return [];
 
@@ -89,8 +120,9 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
 
   const wordCount = countWords(stripped);
   if (wordCount <= chunkSize) {
-    // Single-chunk path: still apply the maxChars cap.
-    const capped = capByChars(stripped.trim(), maxChars);
+    // Single-chunk path: still apply the maxChars + maxTokens caps.
+    const capped = capByChars(stripped.trim(), maxChars)
+      .flatMap((t) => capByEstimatedTokens(t, maxTokens));
     return capped.map((t, i) => ({ text: t, index: i }));
   }
 
@@ -101,9 +133,14 @@ export function chunkText(text: string, opts?: ChunkOptions): TextChunk[] {
   // v0.32.7: hard char cap. Catches pathological CJK + whitespace-less text
   // that the word-level pipeline can't bound (a single Chinese paragraph can
   // exceed 8192 OpenAI embedding tokens at any word count).
+  // v4: estimated-token cap on top — the char cap alone passes token-dense
+  // content (URL soup at ~1.6 chars/token) that overflows strict embedding
+  // server limits.
   const capped: string[] = [];
   for (const chunk of withOverlap) {
-    capped.push(...capByChars(chunk.trim(), maxChars));
+    for (const piece of capByChars(chunk.trim(), maxChars)) {
+      capped.push(...capByEstimatedTokens(piece, maxTokens));
+    }
   }
   return capped.map((t, i) => ({ text: t, index: i }));
 }
@@ -132,6 +169,68 @@ function capByChars(text: string, maxChars: number): string[] {
   return out;
 }
 
+/**
+ * How far back (in chars) the token cap looks for a friendly cut point
+ * before falling back to a hard cut. 300 covers typical rollup/list line
+ * lengths so forced splits land at line starts, not mid-URL.
+ */
+const TOKEN_CAP_CUT_LOOKBACK = 300;
+
+/**
+ * v4: hard-cap a chunk's ESTIMATED embedding tokens. Final safety pass —
+ * runs after capByChars on every chunk, so no upstream miscounting
+ * (whitespace-word fallback, overlap inflation, char-cap survivors) can
+ * emit a chunk past `maxTokens`.
+ *
+ * Cut placement prefers, within the last TOKEN_CAP_CUT_LOOKBACK chars of
+ * the window: a newline, then any whitespace, then a hard cut. This keeps
+ * forced splits off mid-line/mid-URL positions for list-shaped content
+ * and inside code fences. No overlap is added (pieces stay lossless
+ * modulo the trims the char cap already applies).
+ *
+ * @internal exported for the code chunker (code.ts) and tests.
+ */
+export function capByEstimatedTokens(text: string, maxTokens: number): string[] {
+  if (text.length === 0) return [];
+  if (estimateEmbeddingTokens(text) <= maxTokens) return [text];
+
+  const out: string[] = [];
+  let start = 0;
+  while (start < text.length) {
+    // Greedily extend the window until the next char would break the cap.
+    // Always take at least one char so the loop makes forward progress.
+    let est = 0;
+    let end = start;
+    while (end < text.length) {
+      const w = charEmbedTokenWeight(text.charCodeAt(end));
+      if (est + w > maxTokens && end > start) break;
+      est += w;
+      end++;
+    }
+
+    if (end < text.length) {
+      const windowStart = Math.max(start + 1, end - TOKEN_CAP_CUT_LOOKBACK);
+      let cut = text.lastIndexOf('\n', end - 1);
+      if (cut < windowStart) {
+        cut = -1;
+        for (let i = end - 1; i >= windowStart; i--) {
+          const code = text.charCodeAt(i);
+          if (code === 0x20 || (code >= 0x09 && code <= 0x0d)) {
+            cut = i;
+            break;
+          }
+        }
+      }
+      if (cut >= windowStart) end = cut + 1;
+    }
+
+    const slice = text.slice(start, end).trim();
+    if (slice.length > 0) out.push(slice);
+    start = end;
+  }
+  return out;
+}
+
 function recursiveSplit(text: string, level: number, target: number): string[] {
   if (level >= DELIMITERS.length) {
     // Level 4: split on whitespace
@@ -317,7 +416,19 @@ function extractTrailingContext(text: string, targetWords: number): string {
  * Delegated to src/core/cjk.ts so the slugify whitelist, expansion
  * detection, and PGLite keyword fallback all agree on what "CJK enough"
  * means.
+ *
+ * v4: floored at ceil(nonWhitespaceChars/6). The whitespace fallback
+ * counts a 150-char URL as ONE word, so URL/phone/email-dense docs
+ * (whose ASCII mass pushes CJK density below the 0.30 threshold) were
+ * sized at a fraction of their real bulk and merged into 3-4K-char
+ * chunks. The floor makes long whitespace-less runs count roughly
+ * per-character while leaving normal Latin prose untouched (average
+ * English word ≈ 5 chars < 6, so the whitespace count still wins).
+ * Kept local to the chunker — search/expansion.ts keeps the original
+ * countCJKAwareWords semantics for its query-length check.
  */
 function countWords(text: string): number {
-  return countCJKAwareWords(text);
+  const cjkAware = countCJKAwareWords(text);
+  const nonWhitespace = text.replace(/\s/g, '').length;
+  return Math.max(cjkAware, Math.ceil(nonWhitespace / 6));
 }

src/core/cjk.ts

@@ -65,3 +65,64 @@ export function countCJKAwareWords(s: string): number {
 export function escapeLikePattern(s: string): string {
   return s.replace(/\\/g, '\\\\').replace(/%/g, '\\%').replace(/_/g, '\\_');
 }
+
+/**
+ * Conservative per-char-class embedding-token weights (markdown chunker v4).
+ *
+ * Why this exists: the chunker's "word" counting drastically UNDER-counts
+ * whitespace-less ASCII runs (a 150-char URL = 1 whitespace word), so
+ * word-based size targets can emit chunks that overflow an embedding
+ * server's per-request token limit. Measured on a local Qwen3-embedding
+ * llama-server stack:
+ *   - URL/phone/email-dense text tokenizes at ~1.6 chars/token
+ *   - base64-ish / minified blobs approach ~1.3 chars/token (worst case)
+ *   - Korean prose tokenizes NO WORSE than 1 char/token in practice
+ *
+ * Weights are deliberately HIGH (tokens are overestimated) so any cap
+ * based on this estimate is safe against real tokenizers:
+ *   - CJK char        → 1.0 token  (real CJK prose is cheaper)
+ *   - other non-space → 0.75 token (≈1.33 chars/token, covers base64)
+ *   - whitespace      → 0.1 token  (mostly folds into neighbor tokens)
+ */
+export const EMBED_TOKEN_WEIGHT_CJK = 1.0;
+export const EMBED_TOKEN_WEIGHT_OTHER = 0.75;
+export const EMBED_TOKEN_WEIGHT_WS = 0.1;
+
+/** BMP CJK check by UTF-16 code unit — same ranges as CJK_SLUG_CHARS. */
+export function isCJKCodeUnit(code: number): boolean {
+  return (
+    (code >= 0x4e00 && code <= 0x9fff) || // Han
+    (code >= 0x3040 && code <= 0x309f) || // Hiragana
+    (code >= 0x30a0 && code <= 0x30ff) || // Katakana
+    (code >= 0xac00 && code <= 0xd7af)    // Hangul Syllables
+  );
+}
+
+/**
+ * Per-code-unit token weight. Unrecognized whitespace (exotic Unicode
+ * spaces) intentionally falls into OTHER — that only overestimates.
+ */
+export function charEmbedTokenWeight(code: number): number {
+  if (isCJKCodeUnit(code)) return EMBED_TOKEN_WEIGHT_CJK;
+  if (
+    code === 0x20 || (code >= 0x09 && code <= 0x0d) ||
+    code === 0xa0 || code === 0x3000
+  ) {
+    return EMBED_TOKEN_WEIGHT_WS;
+  }
+  return EMBED_TOKEN_WEIGHT_OTHER;
+}
+
+/**
+ * Tokenizer-free embedding-token estimate (conservative overestimate).
+ * See weight docs above. Astral chars count as 2 OTHER code units —
+ * another overestimate, which is the safe direction.
+ */
+export function estimateEmbeddingTokens(s: string): number {
+  if (s.length === 0) return 0;
+  let est = 0;
+  for (let i = 0; i < s.length; i++) {
+    est += charEmbedTokenWeight(s.charCodeAt(i));
+  }
+  return Math.ceil(est);
+}

src/core/embedding.ts

@@ -79,34 +79,15 @@ 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
- * owns progress-callback granularity and (#1818) bounded parallel dispatch
- * of the sub-batches — the embed-stale.ts worker-pool pattern, scoped down.
+ * is purely about progress-callback granularity.
  */
 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 = {},
@@ -122,44 +103,13 @@ export async function embedBatch(
   if (texts.length <= BATCH_SIZE && !options.onBatchComplete) {
     return gatewayEmbed(texts, gwOpts);
   }
-  // #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[] }> = [];
+  const results: Float32Array[] = [];
   for (let i = 0; i < texts.length; i += BATCH_SIZE) {
-    slices.push({ start: i, texts: texts.slice(i, 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);
   }
-  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;
 }
 

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

@@ -39,8 +39,6 @@ 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,
@@ -95,14 +93,6 @@ 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)', () => {
@@ -159,27 +149,6 @@ 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)', () => {
@@ -210,12 +179,6 @@ 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);
@@ -424,92 +387,26 @@ 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).
-      configureCapless();
+      configureGoogle();
       expect(warnings.length).toBe(firstCallCount);
     } finally {
       console.warn = original;
-      RECIPES.delete(caplessRecipe.id);
     }
 
     // The warning text should match the documented contract.
@@ -518,12 +415,11 @@ describe('startup warning for recipes missing max_batch_tokens', () => {
     );
     expect(contractMatch.length).toBe(1);
 
-    // 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.
+    // Voyage declares max_batch_tokens → suppressed. OpenAI is the
+    // canonical fast-path recipe → also suppressed by id. Both must be
+    // absent from the warnings.
     expect(warnings.find(w => w.includes('"voyage"'))).toBeUndefined();
     expect(warnings.find(w => w.includes('"openai"'))).toBeUndefined();
-    expect(warnings.find(w => w.includes('"google"'))).toBeUndefined();
-    expect(warnings.find(w => w.includes('"capless-test"'))).toBeDefined();
+    expect(warnings.find(w => w.includes('"google"'))).toBeDefined();
   });
 });

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

@@ -52,7 +52,16 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
     }
   });
 
-  test('configureGateway does NOT warn for google now that it declares batch caps (#970)', () => {
+  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);
+
     warnSpy.mockClear();
     resetGateway();
     configureGateway({
@@ -60,20 +69,11 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
       embedding_dimensions: 768,
       env: { GOOGLE_GENERATIVE_AI_API_KEY: 'fake' },
     });
-    const messages = warnSpy.mock.calls.map(c => String(c[0] ?? ''));
+    messages = warnSpy.mock.calls.map(c => String(c[0] ?? ''));
     expect(
       messages.some(m => m.includes('"google"') && m.includes('without max_batch_tokens')),
-      '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);
+      'google should warn when configured because it has fixed-cap models',
+    ).toBe(true);
   });
 
   test('every recipe with empty models[] declares user_provided_models OR has openai-fast-path', () => {

test/ai/recipe-dashscope.test.ts

@@ -55,11 +55,6 @@ 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/chunker-version-gate.test.ts

@@ -15,19 +15,22 @@ import { describe, test, expect } from 'bun:test';
 import { CHUNKER_VERSION } from '../src/core/chunkers/code.ts';
 
 describe('Layer 12 — CHUNKER_VERSION constant', () => {
-  test('bumped to 4 for Cathedral II', () => {
+  test('bumped to 5 for the estimated-token hard cap', () => {
     // v3: v0.19.0 Chonkie parity (tokenizer + small-sibling merge).
     // v4: v0.20.0 Cathedral II (qualified names + parent scope + doc_comment
     //     + fence extraction + chunk-grain FTS). Folded into content_hash
     //     so any bump forces clean re-chunks on next sync.
-    expect(CHUNKER_VERSION).toBe(4);
+    // v5: estimated-token hard cap on AST-path chunks (capCodeChunks) so
+    //     un-subdividable giant nodes can't overflow strict embedding
+    //     server token limits.
+    expect(CHUNKER_VERSION).toBe(5);
   });
 
   test('is stable across imports (not recomputed at call time)', async () => {
     const a = (await import('../src/core/chunkers/code.ts')).CHUNKER_VERSION;
     const b = (await import('../src/core/chunkers/code.ts')).CHUNKER_VERSION;
     expect(a).toBe(b);
-    expect(a).toBe(4);
+    expect(a).toBe(5);
   });
 });
 

test/chunkers/code.test.ts

@@ -10,8 +10,8 @@ import { describe, test, expect } from 'bun:test';
 import { chunkCodeText, detectCodeLanguage, CHUNKER_VERSION } from '../../src/core/chunkers/code.ts';
 
 describe('CHUNKER_VERSION', () => {
-  test('v0.20.0 Cathedral II Layer 12 bumped to 4', () => {
-    expect(CHUNKER_VERSION).toBe(4);
+  test('v5: estimated-token hard cap on AST-path chunks', () => {
+    expect(CHUNKER_VERSION).toBe(5);
   });
 });
 

test/chunkers/recursive.test.ts

@@ -135,13 +135,14 @@ describe('Recursive Text Chunker', () => {
 });
 
 describe('CJK chunking (v0.32.7)', () => {
-  test('MARKDOWN_CHUNKER_VERSION is 3', async () => {
+  test('MARKDOWN_CHUNKER_VERSION is 4', async () => {
     // v0.40.3.0: bumped 2→3 to signal the post-upgrade reembed sweep that
-    // contextual retrieval wrapping is now applied at embed time. Chunk
-    // boundaries themselves are unchanged; the bump forces re-embed for
-    // pages where chunker_version < 3.
+    // contextual retrieval wrapping is now applied at embed time.
+    // v4: estimated-token hard cap + whitespace-word undercount floor
+    // (URL-dense docs produced chunks past strict embedding server token
+    // limits). Boundary change → forces re-chunk for chunker_version < 4.
     const mod = await import('../../src/core/chunkers/recursive.ts');
-    expect(mod.MARKDOWN_CHUNKER_VERSION).toBe(3);
+    expect(mod.MARKDOWN_CHUNKER_VERSION).toBe(4);
   });
 
   test('long pure-Chinese paragraph splits into multiple chunks', () => {

test/chunkers/token-cap.test.ts

@@ -0,0 +1,195 @@
+/**
+ * Markdown chunker v4 / code chunker v5 — estimated-token hard cap
+ * regression tests.
+ *
+ * Reproduces a field failure: a local llama-server embedding backend
+ * (`-ub 2048`) crashes deterministically (trace/BPT trap → EOF at the
+ * client) when a single chunk exceeds ~2,050 real tokens. Two content
+ * shapes triggered it:
+ *
+ *   1. Korean docs carrying one long source URL per line.
+ *      The URLs' ASCII mass pushes CJK density below 0.30, flipping
+ *      countCJKAwareWords to whitespace counting, where a 150-char URL
+ *      counts as ONE word → chunks ballooned to 3-4K chars ≈ 2,000+
+ *      real tokens (URL soup tokenizes at ~1.6 chars/token).
+ *
+ *   2. Large JSON code blocks (~7K chars) that the word pipeline
+ *      undercounts the same way (few whitespace tokens).
+ *
+ * The fix: every emitted chunk must satisfy
+ *   estimateEmbeddingTokens(chunk) <= maxTokens (default 1500)
+ * where the estimate deliberately OVERSTATES real tokenizer counts.
+ */
+
+import { describe, test, expect } from 'bun:test';
+import { chunkText, capByEstimatedTokens, DEFAULT_MAX_EST_TOKENS } from '../../src/core/chunkers/recursive.ts';
+import { chunkCodeText } from '../../src/core/chunkers/code.ts';
+import { estimateEmbeddingTokens } from '../../src/core/cjk.ts';
+
+/** Synthesize the failing shape: Korean rollup lines each ending in a long Notion URL. */
+function urlDenseKoreanRollup(lines: number): string {
+  const out: string[] = ['# 링크가 줄마다 붙는 한국어 예시 문서', ''];
+  for (let i = 0; i < lines; i++) {
+    const hex32 = (i * 2654435761 >>> 0).toString(16).padStart(8, '0').repeat(4);
+    out.push(
+      `- **항목 ${i}**: 이 줄은 청커 동작 검증을 위한 의미 없는 한국어 예시 문장입니다 · 전화 000-0000-${String(1000 + i)} · ` +
+      `이메일 user${i}@example.com · 링크: https://docs.example.com/pages/${hex32}?v=abcdef0123456789&ref=sample`,
+    );
+  }
+  return out.join('\n');
+}
+
+/** Synthesize a large pretty-printed JSON block with CJK values. */
+function bigJsonBlock(targetChars: number): string {
+  const entries: string[] = [];
+  let i = 0;
+  let len = 0;
+  while (len < targetChars) {
+    const row =
+      `  "item_${i}": { "name": "예시-${i}", "url": "https://example.com/api/v2/items/${i}?token=abc${i}def", "qty": ${i % 100}, "memo": "한국어 값이 섞인 예시 데이터" }`;
+    entries.push(row);
+    len += row.length;
+    i++;
+  }
+  return `{\n${entries.join(',\n')}\n}`;
+}
+
+describe('v4 estimated-token cap — URL-dense Korean doc (field-failure shape)', () => {
+  test('every chunk stays under the estimated-token cap', () => {
+    const md = urlDenseKoreanRollup(60);
+    const chunks = chunkText(md);
+    expect(chunks.length).toBeGreaterThan(0);
+    for (const c of chunks) {
+      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
+    }
+  });
+
+  test('no chunk reaches the measured 3K-char danger zone for URL soup', () => {
+    const md = urlDenseKoreanRollup(60);
+    const chunks = chunkText(md);
+    // 1500 est tokens at the OTHER weight (0.75/char) bounds chunks to
+    // ~2,000 chars for pure ASCII — well under the ~3,300 chars where
+    // URL-dense content crosses ~2,050 real tokens (1.6 chars/token).
+    for (const c of chunks) {
+      expect(c.text.length).toBeLessThanOrEqual(2600);
+    }
+  });
+
+  test('content is preserved (no lines dropped by the cap)', () => {
+    const md = urlDenseKoreanRollup(60);
+    const chunks = chunkText(md);
+    const joined = chunks.map((c) => c.text).join('\n');
+    // Spot-check first / middle / last rollup lines survive chunking.
+    for (const marker of ['항목 0', '항목 30', '항목 59']) {
+      expect(joined).toContain(marker);
+    }
+  });
+});
+
+describe('v4 estimated-token cap — large JSON blocks', () => {
+  test('7K-char pretty JSON through the prose path stays under the cap', () => {
+    const md = `설정 파일 원문 보존:\n\n\`\`\`\n${bigJsonBlock(7000)}\n\`\`\`\n`;
+    const chunks = chunkText(md);
+    expect(chunks.length).toBeGreaterThan(1);
+    for (const c of chunks) {
+      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
+    }
+  });
+
+  test('7K-char minified JSON (single whitespace-less token) stays under the cap', () => {
+    const minified = bigJsonBlock(7000).replace(/\n\s*/g, '');
+    const chunks = chunkText(minified);
+    expect(chunks.length).toBeGreaterThan(1);
+    for (const c of chunks) {
+      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS);
+    }
+  });
+
+  test('json fence via the code chunker stays under the cap (+header slack)', async () => {
+    const chunks = await chunkCodeText(bigJsonBlock(7000), 'fence.json');
+    expect(chunks.length).toBeGreaterThan(0);
+    for (const c of chunks) {
+      // buildChunk prepends a short "[JSON] fence.json:…" header AFTER the
+      // body-level cap; allow ~60 est tokens of header slack. Real-token
+      // safety margin (2,050 − overestimated 1,500) absorbs this easily.
+      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(DEFAULT_MAX_EST_TOKENS + 60);
+    }
+  });
+});
+
+describe('v4 word-count floor — behavior preserved for normal content', () => {
+  test('Latin prose chunking is unchanged by the floor (avg word < 6 chars)', () => {
+    const prose = Array.from({ length: 120 }, (_, i) =>
+      `This is sentence number ${i} and it talks about ordinary things in plain words.`,
+    ).join(' ');
+    const chunks = chunkText(prose);
+    // Historical behavior: ~1,560 whitespace words → multiple ~300-word chunks.
+    expect(chunks.length).toBeGreaterThan(3);
+    for (const c of chunks) {
+      const words = c.text.split(/\s+/).length;
+      expect(words).toBeLessThanOrEqual(300 * 1.5 + 50); // merge cap + overlap
+    }
+  });
+
+  test('Korean prose (CJK-dense, no URLs) never triggers the token cap', () => {
+    const prose = Array.from({ length: 80 }, (_, i) =>
+      `이 문장은 순수 한국어 산문의 청킹 동작을 확인하기 위한 ${i}번째 예시 문장입니다.`,
+    ).join(' ');
+    const chunks = chunkText(prose);
+    expect(chunks.length).toBeGreaterThan(1);
+    for (const c of chunks) {
+      // CJK-dense chunks are char-counted (≈450 max) — nowhere near 1500.
+      expect(estimateEmbeddingTokens(c.text)).toBeLessThanOrEqual(700);
+    }
+  });
+});
+
+describe('capByEstimatedTokens unit behavior', () => {
+  test('returns input unchanged when under the cap', () => {
+    expect(capByEstimatedTokens('short text', 1500)).toEqual(['short text']);
+    expect(capByEstimatedTokens('', 1500)).toEqual([]);
+  });
+
+  test('prefers newline cut points within the lookback window', () => {
+    const line = 'x'.repeat(100);
+    const text = Array.from({ length: 40 }, () => line).join('\n');
+    const pieces = capByEstimatedTokens(text, 1000);
+    expect(pieces.length).toBeGreaterThan(1);
+    for (const p of pieces) {
+      // Every piece should be whole lines (multiples of the 100-char line).
+      for (const l of p.split('\n')) {
+        expect(l).toBe(line);
+      }
+    }
+  });
+
+  test('makes forward progress on whitespace-less input (hard cut)', () => {
+    const blob = 'a'.repeat(10_000);
+    const pieces = capByEstimatedTokens(blob, 1000);
+    expect(pieces.length).toBeGreaterThan(1);
+    expect(pieces.join('')).toBe(blob);
+    for (const p of pieces) {
+      expect(estimateEmbeddingTokens(p)).toBeLessThanOrEqual(1000);
+    }
+  });
+});
+
+describe('estimateEmbeddingTokens — weight sanity', () => {
+  test('overestimates URL-dense ASCII (0.75/char ≥ measured ~0.63/char)', () => {
+    const url = 'https://docs.example.com/pages/a1b2c3d4e5f6a1b2c3d4e5f6a1b2c3d4?v=abc&ref=sample';
+    const est = estimateEmbeddingTokens(url);
+    expect(est).toBeGreaterThanOrEqual(Math.floor(url.length * 0.7));
+  });
+
+  test('counts CJK at 1 token/char', () => {
+    expect(estimateEmbeddingTokens('가나다라마')).toBe(5);
+  });
+
+  test('whitespace is nearly free', () => {
+    expect(estimateEmbeddingTokens('   \n\t  ')).toBeLessThanOrEqual(1);
+  });
+
+  test('empty string is 0', () => {
+    expect(estimateEmbeddingTokens('')).toBe(0);
+  });
+});

test/embed-batch-concurrency.test.ts

@@ -1,161 +0,0 @@
-/**
- * #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 { afterEach, beforeEach } from 'bun:test';
+import { 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. ✓
-function applyLegacyIfEmpty() {
+beforeEach(() => {
   try {
     // Only re-apply if the gateway was reset (or never configured).
     // Tests that explicitly configured a different model in their
@@ -62,28 +62,4 @@ function applyLegacyIfEmpty() {
   } catch {
     applyLegacy();
   }
-}
-
-beforeEach(applyLegacyIfEmpty);
-
-// PR #3130 shard-order fix: beforeEach alone leaves ONE window open — a file
-// whose LAST afterEach calls resetGateway() poisons the NEXT file's
-// beforeAll, which runs BEFORE any beforeEach fires. A beforeAll there that
-// does engine.initSchema() then sizes the embedding column from the gateway
-// DEFAULTS (zembed-1/1280d) instead of the pinned legacy 1536, and every
-// 1536-d Float32Array fixture in that file dies with
-// "expected 1280 dimensions, not 1536". Which file pair collides is a
-// function of shard composition, so adding/removing ANY test file can
-// surface it (that is exactly how it bit shard 9).
-//
-// Preload hooks are registered before any file-local hooks, and bun runs
-// after-hooks inside-out (file-local afterEach first, then this one), so
-// this repairs the empty slot immediately after the poisoning reset —
-// before the next file's beforeAll can observe it.
-//
-// Known remaining window: a file whose afterAll() resets the gateway (no
-// hook runs between its afterAll and the next file's beforeAll). Files
-// that reset in afterAll and can precede a schema-creating file should
-// re-apply their own config, or the victim file should configureGateway()
-// explicitly in its beforeAll.
-afterEach(applyLegacyIfEmpty);
+});

test/import-default-workers.test.ts

@@ -1,69 +0,0 @@
-/**
- * #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);
-  });
-});