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/cli.ts

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

src/core/ai/gateway.ts

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

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/operations.ts

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

src/core/search/hybrid.ts

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

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

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

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

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

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

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

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

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

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

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

test/voyage-response-cap.test.ts

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