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/autopilot.ts

@@ -38,7 +38,6 @@ import { logSelfUpgrade } from '../core/audit/self-upgrade-audit.ts';
 import { detectInstallMethod } from './upgrade.ts';
 import { evaluateQuietHours } from '../core/minions/quiet-hours.ts';
 import { inspectLock } from '../core/db-lock.ts';
-import { registerCleanup } from '../core/process-cleanup.ts';
 
 /**
  * v0.37.7.0 #1162 — classify autopilot reconnect-loop errors.
@@ -434,37 +433,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
   let stopping = false;
   let childSupervisor: ChildWorkerSupervisor | null = null;
 
-  // #1872: graceful engine shutdown. On PGLite the cycle steps run INLINE in
-  // this process, so a hard `process.exit` mid-write (systemctl stop →
-  // SIGTERM) kills WASM Postgres with the WAL dirty and can corrupt the
-  // brain. Two exit paths must both close the engine:
-  //   - autopilot's own shutdown() below (owns SIGINT + internal stops like
-  //     max_crashes / cycle-failure-cap), and
-  //   - process-cleanup's SIGTERM handler (installed at cli.ts module load;
-  //     it runs the cleanup registry with a 3s deadline and then exits) —
-  //     which is why closeEngine is ALSO registered there.
-  // closeEngine aborts the in-flight inline cycle (runCycle checks the
-  // signal between phases and threads it into phase sub-work), gives it a
-  // short bounded window to wind down, then disconnects. PGLite's
-  // disconnect() drains the pending query and checkpoints before closing;
-  // a second call is a no-op (disconnect snapshots + nulls the handle), so
-  // both paths firing is safe.
-  const shutdownAbort = new AbortController();
-  let inflightInlineCycle: Promise<unknown> | null = null;
-  const closeEngine = async () => {
-    shutdownAbort.abort(new Error('autopilot shutdown'));
-    if (inflightInlineCycle) {
-      // ponytail: 2s cap keeps us inside process-cleanup's 3s deadline; a
-      // between-phase abort resolves instantly, a mid-phase one may not.
-      await Promise.race([
-        inflightInlineCycle.catch(() => { /* cycle errors already logged by the loop */ }),
-        new Promise((r) => setTimeout(r, 2_000)),
-      ]);
-    }
-    try { await engine.disconnect(); } catch { /* best-effort */ }
-  };
-  const deregisterEngineClose = registerCleanup('autopilot-engine-close', closeEngine);
-
   if (spawnManagedWorker) {
     const cliPath = resolveGbrainCliPath();
     // Cgroup-aware auto-sized RSS watchdog cap (issue #1678). The old flat
@@ -552,10 +520,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
         childSupervisor.killChild('SIGKILL');
       }
     }
-    // #1872: abort the in-flight inline cycle and close the engine BEFORE
-    // process.exit — a hard exit mid-write corrupts PGLite's WASM Postgres.
-    await closeEngine();
-    deregisterEngineClose();
     try { unlinkSync(lockPath); } catch { /* already gone */ }
     process.exit(0);
   };
@@ -1044,21 +1008,16 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
       // path's phase set). Now both converge on the same primitive.
       try {
         const { runCycle } = await import('../core/cycle.ts');
-        // #1872: track the promise so closeEngine can drain it on shutdown,
-        // and pass the abort signal so the cycle winds down between phases.
-        const cyclePromise = runCycle(engine, {
+        const report = await runCycle(engine, {
           brainDir: repoPath,
           // Autopilot daemon path: pulls by default (matches
           // pre-v0.17 autopilot behavior). CLI dream defaults false
           // for cron safety; that choice is scoped to dream only.
           pull: true,
-          signal: shutdownAbort.signal,
           yieldBetweenPhases: async () => {
             await new Promise(r => setImmediate(r));
           },
         });
-        inflightInlineCycle = cyclePromise;
-        const report = await cyclePromise.finally(() => { inflightInlineCycle = null; });
         // Only 'failed' (every attempted phase failed) trips the autopilot
         // circuit breaker. 'partial' means at least one phase warned or
         // failed while others ran — that's a soft signal, not a fatal

src/commands/dream.ts

@@ -26,7 +26,6 @@
 import type { BrainEngine } from '../core/engine.ts';
 import {
   runCycle,
-  resolveSourceForDir,
   ALL_PHASES,
   type CyclePhase,
   type CycleReport,
@@ -381,9 +380,9 @@ Options:
 
   --source <id>       Scope the cycle to one source so doctor's
                       cycle_freshness check sees a fresh stamp on
-                      completion. When omitted, gbrain derives the
-                      source from --dir / the configured checkout
-                      when it matches a source's local_path (#1869).
+                      completion. Without this, gbrain dream's
+                      timestamp never lands and federated brains
+                      see "stale cycle" forever.
   --source-id <id>    Alias for --source. Matches the v0.37.7.0+
                       naming used by import/extract/graph-query.
 
@@ -635,25 +634,6 @@ export async function runDream(engine: BrainEngine | null, args: string[]): Prom
     );
     process.exit(1);
   }
-
-  // #1869: a path-scoped run (--dir, or the configured sync.repo_path) whose
-  // directory matches a registered source's local_path IS that source's cycle
-  // — derive the source id so runCycle writes last_source_cycle_at /
-  // last_full_cycle_at on success and doctor's cycle_freshness check stops
-  // reading perpetually stale. Explicit --source still wins (resolved above).
-  // Fixed here at the command level, NOT in runCycle's stamp gate, so legacy
-  // global callers (autopilot-global-maintenance runs GLOBAL_PHASES with a
-  // brainDir and no sourceId) can't falsely stamp per-source freshness.
-  // A derived match on an archived source is skipped silently (falls back to
-  // legacy unscoped behavior) — stamping it would mask staleness on restore,
-  // mirroring the explicit --source archived guard above.
-  if (resolvedSourceId === undefined && engine !== null && brainDir !== null) {
-    const derived = await resolveSourceForDir(engine, brainDir);
-    if (derived !== undefined) {
-      const src = await fetchSource(engine, derived);
-      if (src?.archived !== true) resolvedSourceId = derived;
-    }
-  }
   // ─── issue #1678: bounded single-hold extract_atoms drain ──────────
   if (opts.drain) {
     if (engine === null) {

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/cycle.ts

@@ -855,16 +855,8 @@ interface SyncPhaseResult extends PhaseResult {
  * Resolve the source id for a brain directory by looking up the sources
  * table. Returns undefined when no registered source matches (falls back
  * to pre-v0.18 global config.sync.* keys).
- *
- * Exported for dream.ts (#1869): a `gbrain dream --dir <path>` run whose
- * path matches a registered source's local_path is a per-source cycle in
- * everything but name, so dream derives the source id up front and passes
- * it as opts.sourceId — landing the freshness stamp without changing
- * runCycle's stamp/lock semantics for legacy global callers (the
- * autopilot-global-maintenance handler runs GLOBAL_PHASES with a brainDir
- * and MUST NOT stamp per-source freshness; see rejected PR #2549).
  */
-export async function resolveSourceForDir(
+async function resolveSourceForDir(
   engine: BrainEngine,
   brainDir: string | null,
 ): Promise<string | undefined> {

test/autopilot-shutdown-engine-close.test.ts

@@ -1,63 +0,0 @@
-/**
- * #1872 — autopilot SIGTERM/SIGINT must close the engine before exit.
- *
- * On PGLite the cycle steps run INLINE in the autopilot process, so a hard
- * `process.exit` mid-write (systemctl stop → SIGTERM) kills WASM Postgres
- * with the WAL dirty and can corrupt the brain. Two exit paths must both
- * close the engine:
- *
- *   - autopilot's own shutdown() (owns SIGINT + internal stops like
- *     max_crashes / cycle-failure-cap), and
- *   - process-cleanup's SIGTERM handler (installed at cli.ts module load,
- *     which exits within its 3s cleanup deadline) — reached via the
- *     registered 'autopilot-engine-close' cleanup callback.
- *
- * Because the shutdown path is deep inside `runAutopilot()` (a long-running
- * daemon loop that ends in process.exit), a behavioral test would have to
- * spawn + signal a real daemon. Following the established precedent
- * (test/autopilot-supervisor-wiring.test.ts, test/autopilot-fanout-wiring.test.ts),
- * these static-shape regressions pin the load-bearing wiring instead.
- */
-import { describe, expect, it } from 'bun:test';
-import { readFileSync } from 'fs';
-import { join } from 'path';
-
-const AUTOPILOT_SRC = readFileSync(
-  join(import.meta.dir, '..', 'src', 'commands', 'autopilot.ts'),
-  'utf8',
-);
-
-describe('autopilot.ts graceful engine shutdown (#1872)', () => {
-  it('registers an engine-close callback in the process-cleanup registry (SIGTERM path)', () => {
-    // process-cleanup owns SIGTERM (installed at cli.ts:10) and hard-exits
-    // after its cleanup pass; without this registration the engine is never
-    // closed on `systemctl stop`.
-    expect(AUTOPILOT_SRC).toContain(
-      "import { registerCleanup } from '../core/process-cleanup.ts';",
-    );
-    expect(AUTOPILOT_SRC).toContain(
-      "registerCleanup('autopilot-engine-close', closeEngine)",
-    );
-  });
-
-  it('closeEngine aborts the in-flight inline cycle then disconnects the engine', () => {
-    // Abort first (runCycle checks the signal between phases and threads it
-    // into phase sub-work), bounded drain, then disconnect.
-    expect(AUTOPILOT_SRC).toMatch(
-      /const closeEngine = async \(\) => \{[\s\S]{0,900}shutdownAbort\.abort\([\s\S]{0,900}engine\.disconnect\(\)/,
-    );
-  });
-
-  it('the inline runCycle call carries the shutdown abort signal and is tracked as in-flight', () => {
-    // PGLite / --inline path: the cycle runs in-process, so shutdown must be
-    // able to (a) signal it to wind down and (b) await it before closing.
-    expect(AUTOPILOT_SRC).toMatch(/signal:\s*shutdownAbort\.signal/);
-    expect(AUTOPILOT_SRC).toMatch(/inflightInlineCycle\s*=\s*cyclePromise/);
-  });
-
-  it('shutdown() awaits closeEngine() before process.exit(0) (SIGINT + internal-stop path)', () => {
-    expect(AUTOPILOT_SRC).toMatch(
-      /await closeEngine\(\);[\s\S]{0,400}process\.exit\(0\)/,
-    );
-  });
-});

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/dream-dir-source-stamp.test.ts

@@ -1,99 +0,0 @@
-/**
- * #1869 — `gbrain dream --dir <path>` stamps cycle freshness when the path
- * matches a registered source's local_path.
- *
- * Pre-fix, only `--source <id>` runs wrote last_source_cycle_at /
- * last_full_cycle_at (runCycle's stamp gate reads opts.sourceId, and dream
- * never derived one from --dir), so a path-scoped brain showed doctor's
- * cycle_freshness as perpetually stale.
- *
- * The fix lives in dream.ts (derive the source id from the resolved brain
- * dir via resolveSourceForDir), NOT in runCycle's stamp gate — a runCycle-
- * wide change would make the autopilot-global-maintenance handler (global
- * phases, brainDir set, no sourceId) falsely stamp per-source freshness
- * (the #2194 poisoning class; see rejected PR #2549).
- *
- * Same real-PGLite/no-mocks discipline as test/dream.test.ts; same
- * GBRAIN_HOME isolation as test/cycle-last-full-cycle-at.test.ts (the
- * cycle's PGLite file lock lives under ~/.gbrain).
- */
-import { describe, test, expect, beforeAll, afterAll, beforeEach, afterEach } from 'bun:test';
-import { mkdtempSync, rmSync } from 'fs';
-import { join } from 'path';
-import { tmpdir } from 'os';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { resetPgliteState } from './helpers/reset-pglite.ts';
-import { runDream } from '../src/commands/dream.ts';
-import { withEnv } from './helpers/with-env.ts';
-
-let engine: PGLiteEngine;
-let brainDir: string;
-let gbrainHome: string;
-
-beforeAll(async () => {
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-}, 60_000);
-
-afterAll(async () => {
-  await engine.disconnect();
-});
-
-beforeEach(async () => {
-  await resetPgliteState(engine);
-  brainDir = mkdtempSync(join(tmpdir(), 'gbrain-dream-stamp-'));
-  gbrainHome = mkdtempSync(join(tmpdir(), 'gbrain-dream-stamp-home-'));
-}, 60_000);
-
-afterEach(() => {
-  rmSync(brainDir, { recursive: true, force: true });
-  rmSync(gbrainHome, { recursive: true, force: true });
-});
-
-async function seedSource(id: string, archived = false): Promise<void> {
-  await engine.executeRaw(
-    `INSERT INTO sources (id, name, local_path, config, archived, created_at)
-     VALUES ($1, $2, $3, '{}'::jsonb, $4, NOW())`,
-    [id, id, brainDir, archived],
-  );
-}
-
-async function readLastFullCycleAt(sourceId: string): Promise<string | null> {
-  const rows = await engine.executeRaw<{ config: Record<string, unknown> | null }>(
-    `SELECT config FROM sources WHERE id = $1`,
-    [sourceId],
-  );
-  const raw = rows[0]?.config?.last_full_cycle_at;
-  return typeof raw === 'string' ? raw : null;
-}
-
-describe('gbrain dream --dir <path> freshness stamp (#1869)', () => {
-  test('--dir matching a source local_path stamps last_full_cycle_at', async () => {
-    await withEnv({ GBRAIN_HOME: gbrainHome }, async () => {
-      await seedSource('path-scoped');
-      expect(await readLastFullCycleAt('path-scoped')).toBeNull();
-
-      const report = await runDream(engine, ['--dir', brainDir, '--phase', 'lint', '--json']);
-      expect(report).toBeTruthy();
-      if (report) expect(['ok', 'clean']).toContain(report.status);
-
-      // Pre-fix this stays null forever: dream never passed a sourceId, so
-      // runCycle's stamp gate skipped the write.
-      expect(await readLastFullCycleAt('path-scoped')).not.toBeNull();
-    });
-  }, 60_000);
-
-  test('--dir matching an ARCHIVED source does not stamp it', async () => {
-    await withEnv({ GBRAIN_HOME: gbrainHome }, async () => {
-      await seedSource('mothballed', true);
-
-      const report = await runDream(engine, ['--dir', brainDir, '--phase', 'lint', '--json']);
-      expect(report).toBeTruthy();
-
-      // Stamping an archived source would mask data staleness when it is
-      // later restored (mirrors the explicit --source archived guard).
-      expect(await readLastFullCycleAt('mothballed')).toBeNull();
-    });
-  }, 60_000);
-});

test/dream.test.ts

@@ -562,22 +562,12 @@ describe('runDream — --source / --source-id (v0.41.13)', () => {
 
   // ─── Back-compat: bare `gbrain dream` does NOT write per-source stamp ─
 
-  test('gbrain dream (no --source) stamps only the source whose local_path matches --dir (#1869)', async () => {
-    // Pre-#1869 this asserted NO source was ever stamped without an explicit
-    // --source — which is exactly the bug: a path-scoped `gbrain dream --dir`
-    // run never landed a freshness stamp and doctor's cycle_freshness stayed
-    // stale forever. New truth: the source whose local_path matches the
-    // resolved brain dir is derived and stamped; unrelated sources stay
-    // untouched (cross-source isolation).
-    await seedSource('alpha'); // local_path = repo → derived + stamped
-    await engine.executeRaw(
-      `INSERT INTO sources (id, name, local_path, config, archived, created_at)
-       VALUES ($1, $2, $3, '{}'::jsonb, false, NOW())`,
-      ['beta', 'beta', '/somewhere/else'],
-    );
+  test('gbrain dream (no --source) leaves all sources untouched (back-compat regression)', async () => {
+    await seedSource('alpha');
+    await seedSource('beta');
     const report = await runDream(engine, ['--dir', repo, '--phase', 'lint', '--json']);
     expect(report).toBeTruthy();
-    expect(await readLastFullCycleAt('alpha')).not.toBeNull();
+    expect(await readLastFullCycleAt('alpha')).toBeNull();
     expect(await readLastFullCycleAt('beta')).toBeNull();
   }, 60_000);