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/doctor.ts

@@ -820,10 +820,6 @@ export async function doctorReportRemote(engine: BrainEngine): Promise<DoctorRep
   // v0.42.x (#1794, 4A): pool-budget nudge when GBRAIN_MAX_CONNECTIONS is set.
   checks.push(await checkPoolBudget(engine));
 
-  // #2552: warn when an explicit embed-concurrency override fans out against
-  // a local single-slot embedding endpoint (silent backfill starvation).
-  checks.push(await checkEmbedConcurrency());
-
   // v0.42.7 (#1696): link-extraction lag. Strictly SQL (single indexed COUNT),
   // safe on the thin-client/remote path — remote operators on checkout-less
   // Postgres brains are exactly who can't otherwise see the extraction backlog.
@@ -3819,61 +3815,6 @@ export function computePoolBudgetCheck(
   };
 }
 
-/**
- * #2552: warn when an explicit GBRAIN_EMBED_CONCURRENCY override fans out
- * against a local single-slot embedding endpoint (Ollama / llama-server /
- * localhost base URL). Requests serialize on the one loaded model, so N
- * parallel pages multiply latency xN and can exceed the fetch timeout with
- * no surfaced error — the backfill silently starves. (When the env var is
- * unset, embed auto-caps at LOCAL_EMBED_CONCURRENCY_CAP and this check
- * reports ok.) Pure; exported for tests.
- */
-export function computeEmbedConcurrencyCheck(
-  isLocalEndpoint: boolean,
-  envValue: string | undefined,
-  localCap: number,
-): Check {
-  const name = 'embed_concurrency';
-  if (!isLocalEndpoint) {
-    return { name, status: 'ok', message: 'Embedding endpoint is not a local inference server — cloud concurrency defaults apply.' };
-  }
-  const parsed = envValue ? parseInt(envValue, 10) : NaN;
-  if (envValue && Number.isFinite(parsed) && parsed > localCap) {
-    return {
-      name,
-      status: 'warn',
-      message:
-        `GBRAIN_EMBED_CONCURRENCY=${parsed} against a local embedding endpoint. ` +
-        `Local inference servers serialize requests, so ${parsed} parallel pages multiply ` +
-        `latency x${parsed} and can exceed the fetch timeout — the embed backfill stalls ` +
-        `with no error. Unset GBRAIN_EMBED_CONCURRENCY (auto-caps at ${localCap}) or set it <= ${localCap}.`,
-    };
-  }
-  return {
-    name,
-    status: 'ok',
-    message: `Local embedding endpoint detected; embed concurrency capped at ${envValue ? parsed : localCap}.`,
-  };
-}
-
-/** Thin gateway/env wrapper over `computeEmbedConcurrencyCheck`. */
-export async function checkEmbedConcurrency(): Promise<Check> {
-  try {
-    const { isLocalEmbeddingEndpoint, LOCAL_EMBED_CONCURRENCY_CAP } = await import('../core/ai/gateway.ts');
-    return computeEmbedConcurrencyCheck(
-      isLocalEmbeddingEndpoint(),
-      process.env.GBRAIN_EMBED_CONCURRENCY,
-      LOCAL_EMBED_CONCURRENCY_CAP,
-    );
-  } catch (err) {
-    return {
-      name: 'embed_concurrency',
-      status: 'ok',
-      message: `Skipped (${err instanceof Error ? err.message : String(err)})`,
-    };
-  }
-}
-
 /** Thin env/engine wrapper over `computePoolBudgetCheck`. */
 export async function checkPoolBudget(_engine: BrainEngine): Promise<Check> {
   try {

src/commands/embed.ts

@@ -1,6 +1,5 @@
 import type { BrainEngine } from '../core/engine.ts';
 import { embedBatch, currentEmbeddingSignature } from '../core/embedding.ts';
-import { isLocalEmbeddingEndpoint, LOCAL_EMBED_CONCURRENCY_CAP } from '../core/ai/gateway.ts';
 import type { ChunkInput } from '../core/types.ts';
 import { chunkText } from '../core/chunkers/recursive.ts';
 import { createProgress, type ProgressReporter } from '../core/progress.ts';
@@ -177,31 +176,6 @@ export class EmbeddingDimMismatchError extends Error {
   }
 }
 
-/**
- * #2552: resolve the bulk-embed worker count. Env override or the
- * cloud-tuned default of 20 — but when the operator did NOT set
- * GBRAIN_EMBED_CONCURRENCY and the embedding endpoint is a local inference
- * server (Ollama / llama-server / localhost base URL), cap at
- * LOCAL_EMBED_CONCURRENCY_CAP: 20 parallel pages against a single-slot
- * server serialize on the one loaded model, multiply latency x20 past the
- * fetch timeout, and starve the backfill with no surfaced error. An
- * explicit env value always wins (`gbrain doctor` warns instead).
- * Pacing only ever LOWERS concurrency (Codex P2).
- */
-export function resolveEmbedConcurrency(paceMaxConcurrency?: number): number {
-  const envSet = !!process.env.GBRAIN_EMBED_CONCURRENCY;
-  const base = parseInt(process.env.GBRAIN_EMBED_CONCURRENCY || '20', 10);
-  let resolved = base;
-  if (!envSet && isLocalEmbeddingEndpoint() && base > LOCAL_EMBED_CONCURRENCY_CAP) {
-    resolved = LOCAL_EMBED_CONCURRENCY_CAP;
-    serr(
-      `[embed] local embedding endpoint detected — capping concurrency at ` +
-      `${LOCAL_EMBED_CONCURRENCY_CAP} (set GBRAIN_EMBED_CONCURRENCY to override)`,
-    );
-  }
-  return paceMaxConcurrency ? Math.min(resolved, paceMaxConcurrency) : resolved;
-}
-
 /**
  * Pre-flight check: read the actual schema column dim and compare to the
  * gateway's resolved dim. Throws `EmbeddingDimMismatchError` on mismatch
@@ -703,8 +677,10 @@ async function embedAll(
   // Paced runs lower this to the resolved cap (the real lever vs pooler-slot
   // starvation); unpaced keeps the env/default 20. Codex P2: only ever LOWER —
   // never raise above an operator's existing env cap.
-  // #2552: local endpoints auto-cap — see resolveEmbedConcurrency.
-  const CONCURRENCY = resolveEmbedConcurrency(staleOpts?.paceMaxConcurrency);
+  const BASE_CONCURRENCY = parseInt(process.env.GBRAIN_EMBED_CONCURRENCY || '20', 10);
+  const CONCURRENCY = staleOpts?.paceMaxConcurrency
+    ? Math.min(BASE_CONCURRENCY, staleOpts.paceMaxConcurrency)
+    : BASE_CONCURRENCY;
 
   async function embedOnePage(page: typeof pages[number]) {
     // #1737: bail before doing any work for this page if the run was aborted.
@@ -879,8 +855,10 @@ async function embedAllStale(
   // Paced runs lower concurrency to the resolved cap (E-1: worker count IS the
   // lever on this single pool, no separate permit). Codex P2: pacing only ever
   // LOWERS concurrency — never raise above an operator's existing env cap.
-  // #2552: local endpoints auto-cap — see resolveEmbedConcurrency.
-  const CONCURRENCY = resolveEmbedConcurrency(staleOpts?.paceMaxConcurrency);
+  const BASE_CONCURRENCY = parseInt(process.env.GBRAIN_EMBED_CONCURRENCY || '20', 10);
+  const CONCURRENCY = staleOpts?.paceMaxConcurrency
+    ? Math.min(BASE_CONCURRENCY, staleOpts.paceMaxConcurrency)
+    : BASE_CONCURRENCY;
   const pacer = staleOpts?.pacer ?? createNoopPacer();
 
   // D3 + D3a + D8: wall-clock budget. 30 min default; env override.

src/core/ai/gateway.ts

@@ -683,33 +683,6 @@ export function getEmbeddingDimensions(): number {
   return requireConfig().embedding_dimensions ?? DEFAULT_EMBEDDING_DIMENSIONS;
 }
 
-/**
- * #2552: cap for parallel bulk-embed workers against a local inference
- * server. A single-slot Ollama/llama-server serializes requests, so the
- * cloud-tuned 20-worker fan-out multiplies latency x20 and blows past the
- * fetch timeout with no surfaced error (the backfill silently starves).
- */
-export const LOCAL_EMBED_CONCURRENCY_CAP = 2;
-
-/**
- * #2552: true when the configured embedding model routes to a local
- * inference server — the `ollama` / `llama-server` recipes, or any recipe
- * whose base URL was explicitly pointed at localhost. Bulk callers use this
- * to pick CPU-safe concurrency defaults; `gbrain doctor` uses it to warn
- * about an explicit cloud-sized override. Fail-open: unconfigured or
- * unresolvable gateway → false (cloud behavior, the historical default).
- */
-export function isLocalEmbeddingEndpoint(): boolean {
-  try {
-    const { recipe } = resolveRecipe(getEmbeddingModel());
-    if (recipe.id === 'ollama' || recipe.id === 'llama-server') return true;
-    const base = requireConfig().base_urls?.[recipe.id] ?? '';
-    return /\/\/(localhost|127\.0\.0\.1|\[::1\])(:|\/|$)/i.test(base);
-  } catch {
-    return false;
-  }
-}
-
 /**
  * v0.28.11: returns the configured multimodal embedding model when set,
  * or undefined if the brain falls back to `embedding_model` for multimodal

src/core/ai/recipes/ollama.ts

@@ -29,17 +29,9 @@ export const ollama: Recipe = {
       trust_custom_dims: true, // #2271: local models carry varied native dims
       cost_per_1m_tokens_usd: 0,
       price_last_verified: '2026-04-20',
-      // #2552: Ollama's true batch capacity depends on the locally loaded
-      // model + OLLAMA_NUM_PARALLEL, but the previous `no_batch_cap: true`
-      // meant a whole page went out in ONE request — on a CPU-only box that
-      // multiplies latency past the fetch timeout and the backfill starves
-      // with no surfaced error. Ollama doesn't return a recognizable
-      // token-limit error either, so the recursive-halving safety net never
-      // fires; a conservative static pre-split cap is the only guard.
-      // 4096 tokens x 2 chars/token ~= 8K chars per request (code-dense
-      // pages run ~2 chars/token, not the tiktoken-ish 4).
-      max_batch_tokens: 4096,
-      chars_per_token: 2,
+      // Ollama's batch capacity depends on the locally loaded model + the
+      // OLLAMA_NUM_PARALLEL config; no static cap to declare. v0.32 (#779).
+      no_batch_cap: true,
     },
   },
   setup_hint: 'Install Ollama from https://ollama.ai, then `ollama pull nomic-embed-text` and `ollama serve`.',

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/doctor-categories.ts

@@ -141,7 +141,6 @@ export const OPS_CHECK_NAMES: ReadonlySet<string> = new Set([
   'pgbouncer_prepare',
   'pgvector',
   'pool_budget',
-  'embed_concurrency',
   'progressive_batch_audit_health',
   'queue_health',
   'reranker_health',

src/core/pglite-engine.ts

@@ -56,7 +56,7 @@ import { GBrainError, PAGE_SORT_SQL, ENRICH_ORDER_SQL } from './types.ts';
 import { finalizeLastSeen } from './chronicle/last-seen.ts';
 import { computeAnomaliesFromBuckets } from './cycle/anomaly.ts';
 import { resolveBoostMap, resolveHardExcludes } from './search/source-boost.ts';
-import { buildSourceFactorCase, buildHardExcludeClause, buildVisibilityClause, buildRecencyComponentSql, buildBestPerPagePoolCte, buildOrFallbackWebsearchQuery, buildWebsearchQueryExpr } from './search/sql-ranking.ts';
+import { buildSourceFactorCase, buildHardExcludeClause, buildVisibilityClause, buildRecencyComponentSql, buildBestPerPagePoolCte, buildOrFallbackWebsearchQuery } from './search/sql-ranking.ts';
 import {
   normalizeEngineColumn,
   buildVectorCastFragment,
@@ -1591,8 +1591,6 @@ export class PGLiteEngine implements BrainEngine {
     }
 
     // v0.20.0 Cathedral II Layer 10 C1/C2: language + symbol-kind filters.
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query, innerLimit, limit, offset];
     let extraFilter = '';
     if (opts?.language) {
@@ -1632,7 +1630,6 @@ export class PGLiteEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
     const keywordSql =
       `WITH ranked AS (
@@ -1640,14 +1637,14 @@ export class PGLiteEngine implements BrainEngine {
            p.slug, p.id as page_id, p.title, p.type, p.source_id,
            p.effective_date, p.effective_date_source,
            cc.id as chunk_id, cc.chunk_index, cc.chunk_text, cc.chunk_source,
-           ts_rank(cc.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score,
+           ts_rank(cc.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score,
            CASE WHEN p.updated_at < (
              SELECT MAX(te.created_at) FROM timeline_entries te WHERE te.page_id = p.id
            ) THEN true ELSE false END AS stale
          FROM content_chunks cc
          JOIN pages p ON p.id = cc.page_id
          JOIN sources s ON s.id = p.source_id
-         WHERE cc.search_vector @@ ${ftsQueryExpr} ${detailFilter}${extraFilter} ${hardExcludeClause} ${visibilityClause}
+         WHERE cc.search_vector @@ websearch_to_tsquery('${ftsLang}', $1) ${detailFilter}${extraFilter} ${hardExcludeClause} ${visibilityClause}
            -- v0.27.1: hide image rows from default text-keyword search so
            -- OCR text doesn't drown text-page hits. Image-similarity queries
            -- run a separate vector path on embedding_image.
@@ -1715,10 +1712,7 @@ export class PGLiteEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query, limit, offset];
     let extraFilter = '';
     if (opts?.type) {
@@ -1766,7 +1760,7 @@ export class PGLiteEngine implements BrainEngine {
          COALESCE(rep.chunk_index, 0) as chunk_index,
          COALESCE(rep.chunk_text, '') as chunk_text,
          COALESCE(rep.chunk_source, 'compiled_truth') as chunk_source,
-         ts_rank_cd(p.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score,
+         ts_rank_cd(p.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score,
          CASE WHEN p.updated_at < (
            SELECT MAX(te.created_at) FROM timeline_entries te WHERE te.page_id = p.id
          ) THEN true ELSE false END AS stale
@@ -1781,7 +1775,7 @@ export class PGLiteEngine implements BrainEngine {
          ORDER BY (cc.chunk_source = 'compiled_truth') DESC, cc.chunk_index ASC
          LIMIT 1
        ) rep ON true
-       WHERE p.search_vector @@ ${ftsQueryExpr}
+       WHERE p.search_vector @@ websearch_to_tsquery('${ftsLang}', $1)
          ${extraFilter} ${hardExcludeClause} ${visibilityClause}
        ORDER BY score DESC, p.id ASC
        LIMIT $2 OFFSET $3`;
@@ -1968,8 +1962,6 @@ export class PGLiteEngine implements BrainEngine {
       });
     }
 
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query, limit, offset];
     let extraFilter = '';
     if (opts?.language) {
@@ -2004,21 +1996,20 @@ export class PGLiteEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
     const { rows } = await this.db.query(
       `SELECT
          p.slug, p.id as page_id, p.title, p.type, p.source_id,
          p.effective_date, p.effective_date_source,
          cc.id as chunk_id, cc.chunk_index, cc.chunk_text, cc.chunk_source,
-         ts_rank(cc.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score,
+         ts_rank(cc.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score,
          CASE WHEN p.updated_at < (
            SELECT MAX(te.created_at) FROM timeline_entries te WHERE te.page_id = p.id
          ) THEN true ELSE false END AS stale
        FROM content_chunks cc
        JOIN pages p ON p.id = cc.page_id
        JOIN sources s ON s.id = p.source_id
-       WHERE cc.search_vector @@ ${ftsQueryExpr} ${detailFilter}${extraFilter} ${hardExcludeClause} ${visibilityClause}
+       WHERE cc.search_vector @@ websearch_to_tsquery('${ftsLang}', $1) ${detailFilter}${extraFilter} ${hardExcludeClause} ${visibilityClause}
        ORDER BY score DESC
        LIMIT $2 OFFSET $3`,
       params

src/core/postgres-engine.ts

@@ -64,7 +64,7 @@ import { ConnectionManager } from './connection-manager.ts';
 import { logConnectionEvent } from './connection-audit.ts';
 import { validateSlug, contentHash, rowToPage, rowToStalePage, rowToChunk, rowToSearchResult, parseEmbedding, tryParseEmbedding, takeRowToTake, takeHitRowToHit, isUndefinedTableError, warnOncePerProcess } from './utils.ts';
 import { resolveBoostMap, resolveHardExcludes } from './search/source-boost.ts';
-import { buildSourceFactorCase, buildHardExcludeClause, buildVisibilityClause, buildRecencyComponentSql, buildBestPerPagePoolCte, buildOrFallbackWebsearchQuery, buildWebsearchQueryExpr } from './search/sql-ranking.ts';
+import { buildSourceFactorCase, buildHardExcludeClause, buildVisibilityClause, buildRecencyComponentSql, buildBestPerPagePoolCte, buildOrFallbackWebsearchQuery } from './search/sql-ranking.ts';
 import { DEFAULT_EMBEDDING_MODEL, DEFAULT_EMBEDDING_DIMENSIONS } from './ai/defaults.ts';
 import { DELETE_BATCH_SIZE } from './engine-constants.ts';
 
@@ -1691,8 +1691,6 @@ export class PostgresEngine implements BrainEngine {
     const hardExcludePrefixes = resolveHardExcludes(opts?.exclude_slug_prefixes, opts?.include_slug_prefixes);
     const hardExcludeClause = buildHardExcludeClause('p.slug', hardExcludePrefixes);
 
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query];
     let typeClause = '';
     if (type) {
@@ -1763,7 +1761,6 @@ export class PostgresEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
     const rawQuery = `
       WITH ranked_chunks AS (
@@ -1771,11 +1768,11 @@ export class PostgresEngine implements BrainEngine {
           p.slug, p.id as page_id, p.title, p.type, p.source_id,
           p.effective_date, p.effective_date_source,
           cc.id as chunk_id, cc.chunk_index, cc.chunk_text, cc.chunk_source,
-          ts_rank(cc.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score
+          ts_rank(cc.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score
         FROM content_chunks cc
         JOIN pages p ON p.id = cc.page_id
         JOIN sources s ON s.id = p.source_id
-        WHERE cc.search_vector @@ ${ftsQueryExpr}
+        WHERE cc.search_vector @@ websearch_to_tsquery('${ftsLang}', $1)
           ${typeClause}
           ${typesClause}
           ${excludeSlugsClause}
@@ -1866,10 +1863,7 @@ export class PostgresEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query];
     let typeClause = '';
     if (opts?.type) {
@@ -1929,7 +1923,7 @@ export class PostgresEngine implements BrainEngine {
         COALESCE(rep.chunk_index, 0) as chunk_index,
         COALESCE(rep.chunk_text, '') as chunk_text,
         COALESCE(rep.chunk_source, 'compiled_truth') as chunk_source,
-        ts_rank_cd(p.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score,
+        ts_rank_cd(p.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score,
         false AS stale
       FROM pages p
       JOIN sources s ON s.id = p.source_id
@@ -1942,7 +1936,7 @@ export class PostgresEngine implements BrainEngine {
         ORDER BY (cc.chunk_source = 'compiled_truth') DESC, cc.chunk_index ASC
         LIMIT 1
       ) rep ON true
-      WHERE p.search_vector @@ ${ftsQueryExpr}
+      WHERE p.search_vector @@ websearch_to_tsquery('${ftsLang}', $1)
         ${typeClause}
         ${typesClause}
         ${excludeSlugsClause}
@@ -2006,8 +2000,6 @@ export class PostgresEngine implements BrainEngine {
     const hardExcludePrefixes = resolveHardExcludes(opts?.exclude_slug_prefixes, opts?.include_slug_prefixes);
     const hardExcludeClause = buildHardExcludeClause('p.slug', hardExcludePrefixes);
 
-    // #2380: slash-bearing queries match both the split-word and literal
-    // slash forms — see buildWebsearchQueryExpr in ./search/sql-ranking.ts.
     const params: unknown[] = [query];
     let typeClause = '';
     if (type) {
@@ -2068,19 +2060,18 @@ export class PostgresEngine implements BrainEngine {
     // FTS config name (e.g. 'english', 'pt_br'). Validated by getFtsLanguage()
     // — safe to interpolate into raw SQL.
     const ftsLang = getFtsLanguage();
-    const ftsQueryExpr = buildWebsearchQueryExpr(ftsLang, '$1', query);
 
     const rawQuery = `
       SELECT
         p.slug, p.id as page_id, p.title, p.type, p.source_id,
         p.effective_date, p.effective_date_source,
         cc.id as chunk_id, cc.chunk_index, cc.chunk_text, cc.chunk_source,
-        ts_rank(cc.search_vector, ${ftsQueryExpr}) * ${sourceFactorCase} AS score,
+        ts_rank(cc.search_vector, websearch_to_tsquery('${ftsLang}', $1)) * ${sourceFactorCase} AS score,
         false AS stale
       FROM content_chunks cc
       JOIN pages p ON p.id = cc.page_id
       JOIN sources s ON s.id = p.source_id
-      WHERE cc.search_vector @@ ${ftsQueryExpr}
+      WHERE cc.search_vector @@ websearch_to_tsquery('${ftsLang}', $1)
         ${typeClause}
         ${typesClause}
         ${excludeSlugsClause}

src/core/search/sql-ranking.ts

@@ -251,28 +251,6 @@ export function buildOrFallbackWebsearchQuery(query: string): string | null {
   return tokens.join(' OR ');
 }
 
-/**
- * #2380: FTS query expression for slash-bearing queries. Postgres' default
- * text-search parser classifies `foo/bar` as a single `file`-alias lexeme —
- * on BOTH the query side and the index side. So a raw `foo/bar` query only
- * matched documents carrying the identical joined lexeme (literal paths),
- * and a slash-split query only matches documents whose text had the words
- * separated. Neither form alone covers both document shapes; OR the two
- * parses so a slash query matches prose ("foo and bar", stemmed, AND
- * semantics) AND literal slash forms ("src/core/x.ts") alike.
- *
- * Slash-free queries return the plain single-parse expression — byte-
- * identical SQL and identical ts_rank to the historical behavior.
- *
- * `ftsLang` is validated by getFtsLanguage() (safe to interpolate);
- * `param` is a `$N` placeholder, never user text.
- */
-export function buildWebsearchQueryExpr(ftsLang: string, param: string, query: string): string {
-  const plain = `websearch_to_tsquery('${ftsLang}', ${param})`;
-  if (!query.includes('/')) return plain;
-  return `(websearch_to_tsquery('${ftsLang}', translate(${param}, '/', ' ')) || ${plain})`;
-}
-
 // ============================================================
 // v0.29.1 — Recency component SQL builder
 // ============================================================

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

@@ -28,8 +28,8 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
     resetGateway();
   });
 
-  test('LiteLLM and llama-server declare no_batch_cap: true', () => {
-    for (const id of ['litellm', 'llama-server']) {
+  test('Ollama, LiteLLM, llama-server all declare no_batch_cap: true', () => {
+    for (const id of ['ollama', 'litellm', 'llama-server']) {
       const r = getRecipe(id);
       expect(r, `${id} not registered`).toBeDefined();
       expect(
@@ -39,18 +39,6 @@ describe('v0.32 #779: no_batch_cap suppresses the missing-max_batch_tokens warni
     }
   });
 
-  test('#2552: Ollama declares a conservative static batch cap, not no_batch_cap', () => {
-    // A CPU-only Ollama box wedges when a whole page ships in one request;
-    // Ollama never returns a token-limit error so the recursive-halving
-    // safety net can't fire. The pre-split cap is the only guard.
-    const r = getRecipe('ollama');
-    expect(r).toBeDefined();
-    const e = r!.touchpoints.embedding!;
-    expect(e.no_batch_cap).toBeUndefined();
-    expect(e.max_batch_tokens).toBe(4096);
-    expect(e.chars_per_token).toBe(2);
-  });
-
   test('configureGateway does NOT warn for ollama/litellm/llama-server', () => {
     warnSpy.mockClear();
     resetGateway();

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-local-endpoint.serial.test.ts

@@ -1,118 +0,0 @@
-/**
- * #2552: cloud-tuned embedding defaults silently wedge CPU-only local
- * endpoints (Ollama). Three-part fix under test:
- *
- *  1. `isLocalEmbeddingEndpoint()` — gateway helper detecting local
- *     inference servers (ollama / llama-server recipes, localhost base URL).
- *  2. `resolveEmbedConcurrency()` — embed auto-caps the 20-worker fan-out
- *     at LOCAL_EMBED_CONCURRENCY_CAP for local endpoints unless the
- *     operator set GBRAIN_EMBED_CONCURRENCY explicitly.
- *  3. `computeEmbedConcurrencyCheck()` — doctor warns when an explicit env
- *     override fans out against a local endpoint.
- *
- * Serial: mutates process.env and the module-global gateway config.
- */
-
-import { afterAll, afterEach, describe, expect, test } from 'bun:test';
-import {
-  configureGateway,
-  resetGateway,
-  isLocalEmbeddingEndpoint,
-  LOCAL_EMBED_CONCURRENCY_CAP,
-} from '../src/core/ai/gateway.ts';
-import { resolveEmbedConcurrency } from '../src/commands/embed.ts';
-import { computeEmbedConcurrencyCheck } from '../src/commands/doctor.ts';
-
-const SAVED_ENV = process.env.GBRAIN_EMBED_CONCURRENCY;
-
-afterEach(() => {
-  resetGateway();
-  if (SAVED_ENV === undefined) delete process.env.GBRAIN_EMBED_CONCURRENCY;
-  else process.env.GBRAIN_EMBED_CONCURRENCY = SAVED_ENV;
-});
-
-afterAll(() => {
-  resetGateway();
-});
-
-describe('#2552 isLocalEmbeddingEndpoint', () => {
-  test('false when the gateway is not configured (fail-open to cloud behavior)', () => {
-    resetGateway();
-    expect(isLocalEmbeddingEndpoint()).toBe(false);
-  });
-
-  test('true for the ollama recipe', () => {
-    configureGateway({ embedding_model: 'ollama:nomic-embed-text', env: {} });
-    expect(isLocalEmbeddingEndpoint()).toBe(true);
-  });
-
-  test('true for the llama-server recipe', () => {
-    configureGateway({ embedding_model: 'llama-server:my-gguf', env: {} });
-    expect(isLocalEmbeddingEndpoint()).toBe(true);
-  });
-
-  test('false for a cloud recipe', () => {
-    configureGateway({
-      embedding_model: 'openai:text-embedding-3-small',
-      env: { OPENAI_API_KEY: 'fake' },
-    });
-    expect(isLocalEmbeddingEndpoint()).toBe(false);
-  });
-
-  test('true when a cloud recipe base URL is explicitly pointed at localhost', () => {
-    configureGateway({
-      embedding_model: 'openai:text-embedding-3-small',
-      env: { OPENAI_API_KEY: 'fake' },
-      base_urls: { openai: 'http://localhost:8080/v1' },
-    });
-    expect(isLocalEmbeddingEndpoint()).toBe(true);
-  });
-});
-
-describe('#2552 resolveEmbedConcurrency', () => {
-  test('caps at LOCAL_EMBED_CONCURRENCY_CAP for a local endpoint when env is unset', () => {
-    delete process.env.GBRAIN_EMBED_CONCURRENCY;
-    configureGateway({ embedding_model: 'ollama:nomic-embed-text', env: {} });
-    expect(resolveEmbedConcurrency()).toBe(LOCAL_EMBED_CONCURRENCY_CAP);
-  });
-
-  test('explicit env override always wins, even against a local endpoint', () => {
-    process.env.GBRAIN_EMBED_CONCURRENCY = '10';
-    configureGateway({ embedding_model: 'ollama:nomic-embed-text', env: {} });
-    expect(resolveEmbedConcurrency()).toBe(10);
-  });
-
-  test('cloud endpoints keep the historical default of 20', () => {
-    delete process.env.GBRAIN_EMBED_CONCURRENCY;
-    configureGateway({ env: { OPENAI_API_KEY: 'fake' } });
-    expect(resolveEmbedConcurrency()).toBe(20);
-  });
-
-  test('pacing only ever lowers concurrency', () => {
-    delete process.env.GBRAIN_EMBED_CONCURRENCY;
-    configureGateway({ embedding_model: 'ollama:nomic-embed-text', env: {} });
-    expect(resolveEmbedConcurrency(1)).toBe(1);
-    expect(resolveEmbedConcurrency(16)).toBe(LOCAL_EMBED_CONCURRENCY_CAP);
-  });
-});
-
-describe('#2552 computeEmbedConcurrencyCheck (doctor)', () => {
-  test('ok for non-local endpoints', () => {
-    expect(computeEmbedConcurrencyCheck(false, '20', 2).status).toBe('ok');
-  });
-
-  test('warn when an explicit override exceeds the local cap', () => {
-    const check = computeEmbedConcurrencyCheck(true, '20', 2);
-    expect(check.status).toBe('warn');
-    expect(check.message).toContain('GBRAIN_EMBED_CONCURRENCY=20');
-  });
-
-  test('ok when env is unset against a local endpoint (auto-cap applies)', () => {
-    expect(computeEmbedConcurrencyCheck(true, undefined, 2).status).toBe('ok');
-  });
-
-  test('ok when the override is at or under the cap', () => {
-    expect(computeEmbedConcurrencyCheck(true, '2', 2).status).toBe('ok');
-    expect(computeEmbedConcurrencyCheck(true, '1', 2).status).toBe('ok');
-  });
-});

test/pglite-engine.test.ts

@@ -216,67 +216,6 @@ describe('PGLiteEngine: Search', () => {
     expect(results.length).toBe(0);
   });
 
-  // Regression (#2380): queries containing `/` used to bypass FTS AND
-  // semantics. Postgres' default text-search parser classifies `foo/bar` as
-  // a `file`-alias token mapped to the `simple` dictionary, so it became a
-  // single un-stemmed lexeme `'foo/bar'` that never matches indexed text —
-  // the primary FTS pass returned 0 and the OR fallback took over, matching
-  // pages that contain EITHER term. searchKeyword/searchTitles now normalize
-  // `/` to whitespace before websearch_to_tsquery parses, so the primary
-  // AND pass matches directly.
-  test('searchKeyword: slash query matches with AND semantics, not OR fallback', async () => {
-    // Decoy shares only ONE of the two query terms ('enterprise').
-    await engine.putPage('concepts/enterprise-pricing', {
-      type: 'concept', title: 'Widget Pricing',
-      compiled_truth: 'Enterprise pricing for widgets.',
-    });
-    await engine.upsertChunks('concepts/enterprise-pricing', [
-      { chunk_index: 0, chunk_text: 'Enterprise pricing for widgets', chunk_source: 'compiled_truth' },
-    ]);
-
-    // Both terms co-occur only in the novamind chunk. Pre-fix this returned
-    // BOTH pages (primary pass zero-hit → OR fallback); post-fix the primary
-    // AND pass returns exactly the co-occurrence page.
-    const results = await engine.searchKeyword('NovaMind/enterprise');
-    expect(results.length).toBe(1);
-    expect(results[0].slug).toBe('companies/novamind');
-  });
-
-  test('searchTitles: slash query matches with AND semantics, not OR fallback', async () => {
-    await engine.putPage('companies/novamind-enterprise', {
-      type: 'company', title: 'NovaMind Enterprise Platform',
-      compiled_truth: 'Placeholder body.',
-    });
-    await engine.putPage('guides/enterprise-sales', {
-      type: 'concept', title: 'Enterprise Sales Guide',
-      compiled_truth: 'Placeholder body.',
-    });
-
-    // Pre-fix: `NovaMind/Enterprise` parsed as one file-alias lexeme → the
-    // primary title pass returned 0 and the OR fallback matched BOTH titles.
-    const results = await engine.searchTitles('NovaMind/Enterprise');
-    expect(results.length).toBe(1);
-    expect(results[0].slug).toBe('companies/novamind-enterprise');
-  });
-
-  test('searchKeyword: slash query still matches the literal slash form (file paths)', async () => {
-    // The INDEX side also emits the joined file-alias lexeme for literal
-    // `foo/bar` text, so a query normalized to split words alone would go
-    // blind to documents containing the literal slash form (paths, URLs).
-    // buildWebsearchQueryExpr ORs both parses; this pins the raw arm.
-    await engine.putPage('runbooks/widget-deploy', {
-      type: 'concept', title: 'Widget Deploy Runbook',
-      compiled_truth: 'Runbook for the acme/widget deployment pipeline.',
-    });
-    await engine.upsertChunks('runbooks/widget-deploy', [
-      { chunk_index: 0, chunk_text: 'Runbook for the acme/widget deployment pipeline', chunk_source: 'compiled_truth' },
-    ]);
-
-    const results = await engine.searchKeyword('acme/widget');
-    expect(results.length).toBe(1);
-    expect(results[0].slug).toBe('runbooks/widget-deploy');
-  });
-
   test('tsvector trigger populates search_vector on insert', async () => {
     // Verify the PL/pgSQL trigger fires and content_chunks.search_vector is
     // populated from chunk_text. v0.20.0 Cathedral II Layer 3 moved FTS from