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>

docs/architecture/KEY_FILES.md

@@ -457,7 +457,6 @@ unresolvable+true|false, pre-v80 NULL/NULL rows survive).
 - `src/core/think/prompt.ts` extension — anti-bias rewrite. `withCalibration` option on `buildThinkSystemPrompt` adds anti-bias rules. `buildCalibrationBlock()` emits the `<calibration>` XML. `buildThinkUserMessage` has TWO shapes: default (question first), and with-calibration (retrieval → calibration → question) when opt-in. Wired into `runThink` via `opts.withCalibration` + `opts.calibrationHolder`.
 - `src/commands/calibration.ts` — CLI: `gbrain calibration` (read + print), `--regenerate`, `--undo-wave <ver>`, `ab-report`. MCP op `get_calibration_profile` (scope: read) backs the same data path. Source-scoped via `sourceScopeOpts(ctx)`.
 - `src/commands/serve-http.ts` extension — three admin routes: `/admin/api/calibration/profile`, `/admin/api/calibration/charts/:type` (image/svg+xml; type in {brier-trend, domain-bars, pattern-statements, abandoned-threads}), `/admin/api/calibration/pattern/:id` (drill-down).
-- `src/core/owner-holder.ts` — single source of truth for "the brain owner" holder string. `DEFAULT_OWNER_HOLDER = 'self'` (matches the consolidate facts→takes writer + `docs/takes-vs-facts.md`); `resolveOwnerHolder({override, configValue})` returns override > `emotional_weight.user_holder` config > `'self'`. Consumed by the calibration_profile cycle phase, `gbrain calibration` CLI, the `get_calibration_profile` op, `think`'s calibration block, `emotional-weight`'s `DEFAULT_USER_HOLDER`, and doctor's `calibration_freshness`. Pure; unit-tested in `test/owner-holder.test.ts`. Does NOT unify owner-identity fragmentation (`self`/`brain`/`people-<owner>`) — tracked separately.
 - `src/commands/takes.ts` extension — `gbrain takes revisit <slug>` opens $EDITOR on the source page with a `<!-- gbrain:revisit -->` cursor marker.
 - `src/commands/doctor.ts` extension — 4 checks: `abandoned_threads`, `calibration_freshness`, `grade_confidence_drift` (mitigation surface; math ships later), `voice_gate_health`.
 - `admin/src/pages/Calibration.tsx` — Calibration tab. Single-column layout. `<TrustedSVG>` wrapper handles `dangerouslySetInnerHTML` for the server-rendered SVG.

docs/takes-vs-facts.md

@@ -91,18 +91,3 @@ First full takes extraction run on a ~100K-page brain:
 4. **Self-reported ≠ verified.** "Reports 7 figures" → holder=person, weight=0.75, NOT world/1.0
 5. **No false precision.** Use 0.05 increments (0.35, 0.55, 0.75), not 0.74 or 0.82
 6. **"So what" test.** Skip Twitter handles, follower counts, obvious metadata
-
-## Owner-holder canonicalization
-
-"The brain owner" is, by convention, the holder string **`self`** — the value the
-dream `consolidate` phase stamps when it promotes the owner's hot facts into cold
-takes. Calibration, `think`, and the `doctor` calibration check resolve the owner
-holder through `resolveOwnerHolder` (`src/core/owner-holder.ts`): explicit override
-> `emotional_weight.user_holder` config > `self`.
-
-Known limitation (tracked in garrytan/gbrain#2465): the owner can also
-appear under `brain` (a take the owner asserts, via `propose_takes`) and
-`people/<owner>` (extraction that names the owner). The resolver selects the
-*default* canonical owner string for reads; it does not merge those other
-strings. Per-take attribution for other people (e.g. `people/george`) is
-unaffected and correct.

src/commands/calibration.ts

@@ -23,7 +23,6 @@ import { runPhaseCalibrationProfile } from '../core/cycle/calibration-profile.ts
 import { sourceScopeOpts, type OperationContext } from '../core/operations.ts';
 import type { GBrainConfig } from '../core/config.ts';
 import { GBrainError } from '../core/types.ts';
-import { resolveOwnerHolder } from '../core/owner-holder.ts';
 
 export interface CalibrationProfileRow {
   /** BIGSERIAL → string (postgres.js int8 wire shape; never Number() — int8
@@ -168,10 +167,7 @@ export async function runCalibration(
   config: GBrainConfig,
 ): Promise<void> {
   const { opts } = parseArgs(args);
-  const holder = resolveOwnerHolder({
-    override: opts.holder,
-    configValue: await engine.getConfig('emotional_weight.user_holder'),
-  });
+  const holder = opts.holder ?? 'garry';
   // Resolve --source / GBRAIN_SOURCE / .gbrain-source so the (now reachable, #2035)
   // calibration command targets the right source in a multi-source brain instead
   // of always reading `default`. No signal → 'default' (prior behavior).
@@ -257,15 +253,12 @@ export async function getCalibrationProfileOp(
   ctx: OperationContext,
   params: { holder?: string },
 ): Promise<CalibrationProfileRow | null> {
-  const holder = resolveOwnerHolder({
-    override: params.holder,
-    configValue: await ctx.engine.getConfig('emotional_weight.user_holder'),
-  });
+  const holder = params.holder ?? 'garry';
   if (typeof holder !== 'string' || holder.length === 0) {
     throw new GBrainError(
       'INVALID_HOLDER',
       'get_calibration_profile.holder must be a non-empty string',
-      'pass holder="<slug>" or omit to default to the owner holder (config emotional_weight.user_holder, else "self")',
+      'pass holder="<slug>" or omit to default to "garry"',
     );
   }
   const scope = sourceScopeOpts(ctx);

src/commands/doctor.ts

@@ -28,7 +28,6 @@ import type { DbUrlSource } from '../core/config.ts';
 import { gbrainPath, loadConfig } from '../core/config.ts';
 import { reflexEnabled } from '../core/context/reflex.ts';
 import { resolveSocketPath } from '../core/context/resolve-ipc.ts';
-import { resolveOwnerHolder } from '../core/owner-holder.ts';
 import { homedir } from 'os';
 import { dirname, isAbsolute, join, resolve as resolvePath } from 'path';
 import { fileURLToPath } from 'url';
@@ -1288,19 +1287,14 @@ export async function checkAbandonedThreads(engine: BrainEngine): Promise<Check>
 
 /**
  * calibration_freshness: warns when the active calibration profile is
- * older than 7 days (configurable). Default holder resolves via resolveOwnerHolder
- * (config emotional_weight.user_holder, else 'self'). Multi-source
+ * older than 7 days (configurable). Default holder 'garry'. Multi-source
  * brains see one row per source; this check uses the most recent across
  * all sources.
  */
 export async function checkCalibrationFreshness(engine: BrainEngine): Promise<Check> {
   try {
-    const ownerHolder = resolveOwnerHolder({
-      configValue: await engine.getConfig('emotional_weight.user_holder'),
-    });
     const rows = await engine.executeRaw<{ generated_at: Date | null }>(
-      `SELECT MAX(generated_at) AS generated_at FROM calibration_profiles WHERE holder = $1`,
-      [ownerHolder],
+      `SELECT MAX(generated_at) AS generated_at FROM calibration_profiles WHERE holder = 'garry'`,
     );
     const generated = rows[0]?.generated_at;
     if (!generated) {

src/commands/serve-http.ts

@@ -45,7 +45,6 @@ import {
   type IngestionContentType,
   type IngestionEvent,
 } from '../core/ingestion/types.ts';
-import { resolveOwnerHolder } from '../core/owner-holder.ts';
 
 /**
  * /health endpoint timeout. 3s rather than 5s: Fly.io's default
@@ -1191,7 +1190,7 @@ export async function runServeHttp(engine: BrainEngine, options: ServeHttpOption
   app.get('/admin/api/calibration/pattern/:id', requireAdmin, async (req: Request, res: Response) => {
     try {
       const { getLatestProfile } = await import('./calibration.ts');
-      const holder = resolveOwnerHolder({ override: (req.query.holder as string) || undefined, configValue: await engine.getConfig('emotional_weight.user_holder') });
+      const holder = (req.query.holder as string) || 'garry';
       const profile = await getLatestProfile(engine, { holder });
       if (!profile) {
         res.status(404).json({ error: 'no_profile' });
@@ -1241,7 +1240,7 @@ export async function runServeHttp(engine: BrainEngine, options: ServeHttpOption
   app.get('/admin/api/calibration/profile', requireAdmin, async (req: Request, res: Response) => {
     try {
       const { getLatestProfile } = await import('./calibration.ts');
-      const holder = resolveOwnerHolder({ override: (req.query.holder as string) || undefined, configValue: await engine.getConfig('emotional_weight.user_holder') });
+      const holder = (req.query.holder as string) || 'garry';
       const profile = await getLatestProfile(engine, { holder });
       res.json(profile);
     } catch (err) {
@@ -1258,7 +1257,7 @@ export async function runServeHttp(engine: BrainEngine, options: ServeHttpOption
         renderAbandonedThreadsCard,
         renderPatternStatementsCard,
       } = await import('../core/calibration/svg-renderer.ts');
-      const holder = resolveOwnerHolder({ override: (req.query.holder as string) || undefined, configValue: await engine.getConfig('emotional_weight.user_holder') });
+      const holder = (req.query.holder as string) || 'garry';
       const type = req.params.type;
       const profile = await getLatestProfile(engine, { holder });
 

src/commands/takes.ts

@@ -29,7 +29,6 @@ import {
 } from '../core/takes-fence.ts';
 import { withPageLock } from '../core/page-lock.ts';
 import { resolveSourceId } from '../core/source-resolver.ts';
-import { resolveOwnerHolder } from '../core/owner-holder.ts';
 
 // --- Helpers ---
 
@@ -365,7 +364,7 @@ async function cmdResolve(engine: BrainEngine, args: string[], sourceId?: string
   // --evidence is the v0.30.0 alias for --source on the resolve subcommand
   // (semantic clarity: "what evidence resolved this bet?").
   const source = flagValue(args, '--evidence') ?? flagValue(args, '--source');
-  const resolvedBy = flagValue(args, '--by') ?? resolveOwnerHolder({ configValue: await engine.getConfig('emotional_weight.user_holder') });
+  const resolvedBy = flagValue(args, '--by') ?? 'garry';
   const dirArg = flagValue(args, '--dir');
 
   const pageId = await getPageId(engine, slug, sourceId);

src/core/brainstorm/orchestrator.ts

@@ -68,7 +68,6 @@ import {
   type BrainstormCheckpoint,
   type CheckpointCross,
 } from './checkpoint.ts';
-import { resolveOwnerHolder } from '../owner-holder.ts';
 
 export { BudgetExhausted };
 
@@ -140,7 +139,7 @@ export interface BrainstormOptions {
   modelOverride?: string;
   /** Skip the cost-preview TTY grace window. Required for non-interactive callers. */
   skipCostPreview?: boolean;
-  /** When set, force the user holder for calibration profile lookup. Falls back to config (`emotional_weight.user_holder`) then `'self'`. */
+  /** When set, force the user holder for calibration profile lookup. Falls back to config (`emotional_weight.user_holder`) then `'garry'`. */
   holderOverride?: string;
   /** Source scope. */
   sourceId?: string;
@@ -624,7 +623,7 @@ async function _runBrainstormInner(
   }
 
   // ---- Phase 3: calibration context (cold-start fallback) ----
-  const holder = resolveOwnerHolder({ override: opts.holderOverride, configValue: config.emotional_weight?.user_holder });
+  const holder = opts.holderOverride ?? config.emotional_weight?.user_holder ?? 'garry';
   const calibContext = await loadCalibrationContext(engine, {
     holder,
     sourceId: opts.sourceId,

src/core/calibration/think-ab.ts

@@ -25,7 +25,7 @@ import type { BrainEngine } from '../engine.ts';
 
 export interface ABRunInput {
   question: string;
-  /** Holder context for calibration. Resolves via resolveOwnerHolder (config emotional_weight.user_holder, else 'self'). */
+  /** Holder context for calibration. Default 'garry'. */
   holder?: string;
   /** Engine for DB write. */
   engine: BrainEngine;

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/calibration-profile.ts

@@ -26,7 +26,6 @@
  */
 
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
-import { resolveOwnerHolder } from '../owner-holder.ts';
 import { chat as gatewayChat } from '../ai/gateway.ts';
 import { TIER_DEFAULTS } from '../model-config.ts';
 import { gateVoice, type VoiceGateGenerator, type VoiceGateJudge } from '../calibration/voice-gate.ts';
@@ -97,7 +96,7 @@ export type PatternStatementsGenerator = (input: {
 export type BiasTagsGenerator = (patterns: string[]) => Promise<string[]>;
 
 export interface CalibrationProfileOpts extends BasePhaseOpts {
-  /** Holder to generate the profile for. Default resolves via resolveOwnerHolder (config emotional_weight.user_holder, else 'self'). */
+  /** Holder to generate the profile for. Default 'garry'. */
   holder?: string;
   /** Inject the patterns generator (tests). */
   patternsGenerator?: PatternStatementsGenerator;
@@ -228,10 +227,7 @@ class CalibrationProfilePhase extends BaseCyclePhase {
     _ctx: OperationContext,
     opts: CalibrationProfileOpts,
   ): Promise<{ summary: string; details: Record<string, unknown>; status?: PhaseStatus }> {
-    const holder = resolveOwnerHolder({
-      override: opts.holder,
-      configValue: await engine.getConfig('emotional_weight.user_holder'),
-    });
+    const holder = opts.holder ?? 'garry';
     const promptVersion = opts.promptVersion ?? CALIBRATION_PROFILE_PROMPT_VERSION;
     const modelId = opts.model ?? TIER_DEFAULTS.reasoning;
     const gradeCompletion = opts.gradeCompletion ?? 1.0;

src/core/cycle/emotional-weight.ts

@@ -14,8 +14,6 @@
  * See `loadHighEmotionTags` for the resolution path.
  */
 
-import { DEFAULT_OWNER_HOLDER } from '../owner-holder.ts';
-
 /**
  * Default high-emotion tag seed list. Pages with any tag in this set get the
  * tag-emotion boost in the formula below. Override via config key
@@ -45,12 +43,11 @@ export const HIGH_EMOTION_TAGS: ReadonlySet<string> = new Set([
 ]);
 
 /**
- * Holder name treated as "the user" for the user-as-holder ratio. Configurable
- * via the `emotional_weight.user_holder` config key; defaults to the canonical
- * owner holder ('self', DEFAULT_OWNER_HOLDER) so it matches the consolidate
- * facts→takes writer instead of a hardcoded name.
+ * Holder name treated as "the user" for the Garry-as-holder ratio. Configurable
+ * via the `emotional_weight.user_holder` config key (defaults to 'garry' to
+ * match the v0.28 schema's takes table convention).
  */
-export const DEFAULT_USER_HOLDER = DEFAULT_OWNER_HOLDER;
+export const DEFAULT_USER_HOLDER = 'garry';
 
 export interface EmotionalWeightTake {
   holder: string;

src/core/operations.ts

@@ -3289,7 +3289,7 @@ const get_calibration_profile: Operation = {
     holder: {
       type: 'string',
       description:
-        "Holder slug, e.g. 'self' or 'people/charlie-example'. Defaults to config emotional_weight.user_holder, else 'self', when omitted.",
+        "Holder slug, e.g. 'garry' or 'people/charlie-example'. Defaults to 'garry' when omitted.",
     },
   },
   handler: async (ctx, p) => {

src/core/owner-holder.ts

@@ -1,24 +0,0 @@
-/**
- * Canonical holder string for "the brain owner," resolved in ONE place so the
- * calibration / think / doctor / emotional-weight defaults stop disagreeing.
- *
- * The default matches the consolidate facts→takes writer
- * (src/core/cycle/phases/consolidate.ts: holder:'self') and docs/takes-vs-facts.md.
- * Do NOT introduce a fourth literal — three already exist historically
- * ('garry', 'system', 'self'); this is the source of truth.
- *
- * NORMALIZATION NOTE: the brain owner may also appear under other holder
- * strings — 'brain' (propose_takes when the author asserts a claim) and
- * people/<owner> (extraction that names the owner). This resolver only selects
- * the *default* canonical owner string for reads; it does NOT merge those other
- * strings. Unifying them is owner-identity entity-resolution, tracked separately
- * (see garrytan/gbrain#2465). Until then, historical owner takes
- * under 'brain'/people-<owner> are not folded into the default profile.
- */
-export const DEFAULT_OWNER_HOLDER = 'self';
-
-export function resolveOwnerHolder(
-  opts: { override?: string | null; configValue?: string | null },
-): string {
-  return opts.override ?? opts.configValue ?? DEFAULT_OWNER_HOLDER;
-}

src/core/think/index.ts

@@ -23,7 +23,6 @@ import { runGather, renderPagesBlock, takesHitToTakeForPrompt } from './gather.t
 import { renderTakesBlock } from './sanitize.ts';
 import { buildThinkSystemPrompt, buildThinkUserMessage } from './prompt.ts';
 import { resolveCitations, type ParsedCitation } from './cite-render.ts';
-import { resolveOwnerHolder } from '../owner-holder.ts';
 import { resolveModel } from '../model-config.ts';
 import { chat as gatewayChat, probeChatModel, type ChatResult } from '../ai/gateway.ts';
 import { AIConfigError } from '../ai/errors.ts';
@@ -77,8 +76,8 @@ export interface RunThinkOpts {
    */
   withCalibration?: boolean;
   /**
-   * Holder to retrieve the calibration profile for. Resolves via resolveOwnerHolder
-   * (config emotional_weight.user_holder, else 'self'). Only consulted when withCalibration=true.
+   * Holder to retrieve the calibration profile for. Default 'garry'. Only
+   * consulted when withCalibration=true.
    */
   calibrationHolder?: string;
   /**
@@ -309,10 +308,7 @@ export async function runThink(
     try {
       const { getLatestProfile } = await import('../../commands/calibration.ts');
       const profile = await getLatestProfile(engine, {
-        holder: resolveOwnerHolder({
-          override: opts.calibrationHolder,
-          configValue: await engine.getConfig('emotional_weight.user_holder'),
-        }),
+        holder: opts.calibrationHolder ?? 'garry',
       });
       if (profile) {
         calibrationBlockOpts = {

test/calibration-cli.test.ts

@@ -27,12 +27,6 @@ function buildMockEngine(opts: { rows: CalibrationProfileRow[] }): {
   const capturedParams: unknown[][] = [];
   const engine = {
     kind: 'pglite',
-    // #2464: getCalibrationProfileOp resolves the owner holder via
-    // resolveOwnerHolder(config emotional_weight.user_holder, else 'self'), so the
-    // mock must implement getConfig. null = key unset → resolver falls back to 'self'.
-    async getConfig(): Promise<string | null> {
-      return null;
-    },
     async executeRaw<T>(sql: string, params?: unknown[]): Promise<T[]> {
       capturedSql.push(sql);
       capturedParams.push(params ?? []);
@@ -217,17 +211,17 @@ describe('formatProfileText', () => {
 // ─── getCalibrationProfileOp ────────────────────────────────────────
 
 describe('getCalibrationProfileOp (MCP)', () => {
-  test('defaults holder to "self" when omitted (config emotional_weight.user_holder unset)', async () => {
-    const { engine } = buildMockEngine({ rows: [buildProfile({ holder: 'self' })] });
+  test('defaults holder to "garry" when omitted', async () => {
+    const { engine } = buildMockEngine({ rows: [buildProfile({ holder: 'garry' })] });
     const ctx = buildCtx(engine);
     const result = await getCalibrationProfileOp(ctx, {});
-    expect(result?.holder).toBe('self');
+    expect(result?.holder).toBe('garry');
   });
 
   test('routes through sourceScopeOpts: scalar source-bound client gets source-scoped result', async () => {
     const rows = [
-      buildProfile({ holder: 'self', source_id: 'default' }),
-      buildProfile({ holder: 'self', source_id: 'tenant-b' }),
+      buildProfile({ holder: 'garry', source_id: 'default' }),
+      buildProfile({ holder: 'garry', source_id: 'tenant-b' }),
     ];
     const { engine } = buildMockEngine({ rows });
     const ctx = buildCtx(engine, { sourceId: 'tenant-b' });
@@ -237,8 +231,8 @@ describe('getCalibrationProfileOp (MCP)', () => {
 
   test('federated read scope sees the union of allowed sources', async () => {
     const rows = [
-      buildProfile({ holder: 'self', source_id: 'tenant-a' }),
-      buildProfile({ holder: 'self', source_id: 'tenant-z' }),
+      buildProfile({ holder: 'garry', source_id: 'tenant-a' }),
+      buildProfile({ holder: 'garry', source_id: 'tenant-z' }),
     ];
     const { engine } = buildMockEngine({ rows });
     const ctx = buildCtx(engine, { allowedSources: ['tenant-a', 'tenant-b'] });

test/calibration-profile.test.ts

@@ -32,7 +32,7 @@ interface CapturedSql {
   params: unknown[];
 }
 
-function buildMockEngine(opts: { scorecard: TakesScorecard; userHolder?: string | null }): {
+function buildMockEngine(opts: { scorecard: TakesScorecard }): {
   engine: BrainEngine;
   captured: CapturedSql[];
 } {
@@ -42,10 +42,6 @@ function buildMockEngine(opts: { scorecard: TakesScorecard; userHolder?: string
     async getScorecard() {
       return opts.scorecard;
     },
-    async getConfig(key: string): Promise<string | null> {
-      if (key === 'emotional_weight.user_holder') return opts.userHolder ?? null;
-      return null;
-    },
     async executeRaw<T>(sql: string, params?: unknown[]): Promise<T[]> {
       captured.push({ sql, params: params ?? [] });
       return [];
@@ -238,7 +234,7 @@ describe('runPhaseCalibrationProfile — phase integration', () => {
     // grade_completion, domain_scorecards_json, patterns[], voice_passed, voice_attempts,
     // bias_tags[], model_id
     expect(insert!.params[0]).toBe('default'); // source_id
-    expect(insert!.params[1]).toBe('self'); // holder (resolved via resolveOwnerHolder, no override)
+    expect(insert!.params[1]).toBe('garry'); // holder
     expect(insert!.params[2]).toBe(12); // total_resolved
     expect(insert!.params[9]).toBe(true); // voice_gate_passed
     expect(insert!.params[10]).toBe(1); // voice_gate_attempts
@@ -334,24 +330,4 @@ describe('runPhaseCalibrationProfile — phase integration', () => {
     const insert = captured.find(c => c.sql.includes('INSERT INTO calibration_profiles'));
     expect(insert!.params[0]).toBe('tenant-b');
   });
-
-  test('cold-brain summary uses resolved owner holder self when user_holder unset', async () => {
-    const { engine } = buildMockEngine({
-      scorecard: { total_bets: 0, resolved: 0, correct: 0, incorrect: 0, partial: 0,
-        accuracy: null, brier: null, partial_rate: null, unresolvable_count: 0, unresolvable_rate: null },
-    });
-    const result = await runPhaseCalibrationProfile(buildCtx(engine), {});
-    expect(result.summary).toContain('holder=self');
-    expect(result.summary).not.toContain('holder=garry');
-  });
-
-  test('configured user_holder overrides the default in the cold-brain summary', async () => {
-    const { engine } = buildMockEngine({
-      scorecard: { total_bets: 0, resolved: 0, correct: 0, incorrect: 0, partial: 0,
-        accuracy: null, brier: null, partial_rate: null, unresolvable_count: 0, unresolvable_rate: null },
-      userHolder: 'people/charlie-example',
-    });
-    const result = await runPhaseCalibrationProfile(buildCtx(engine), {});
-    expect(result.summary).toContain('holder=people/charlie-example');
-  });
 });

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/doctor-calibration-checks.test.ts

@@ -32,12 +32,6 @@ function buildMockEngine(opts: {
 }): BrainEngine {
   return {
     kind: 'pglite',
-    // #2464: checkCalibrationFreshness resolves the owner holder via
-    // resolveOwnerHolder(config emotional_weight.user_holder, else 'self'), so the
-    // mock must implement getConfig. null = key unset → resolver falls back to 'self'.
-    async getConfig(): Promise<string | null> {
-      return null;
-    },
     async executeRaw<T>(sql: string): Promise<T[]> {
       if (opts.throwOn && opts.throwOn.test(sql)) {
         throw new Error('mock engine error: ' + sql.slice(0, 50));

test/emotional-weight.test.ts

@@ -112,7 +112,3 @@ describe('computeEmotionalWeight', () => {
     expect(HIGH_EMOTION_TAGS.has('mental-health')).toBe(true);
   });
 });
-
-test('DEFAULT_USER_HOLDER is the canonical owner holder self', () => {
-  expect(DEFAULT_USER_HOLDER).toBe('self');
-});

test/owner-holder.test.ts

@@ -1,27 +0,0 @@
-import { describe, test, expect } from 'bun:test';
-import { resolveOwnerHolder, DEFAULT_OWNER_HOLDER } from '../src/core/owner-holder.ts';
-
-describe('owner-holder', () => {
-  test('DEFAULT_OWNER_HOLDER is self', () => {
-    expect(DEFAULT_OWNER_HOLDER).toBe('self');
-  });
-
-  test('defaults to self when nothing provided', () => {
-    expect(resolveOwnerHolder({})).toBe('self');
-  });
-
-  test('null/undefined config falls back to self', () => {
-    expect(resolveOwnerHolder({ configValue: null })).toBe('self');
-    expect(resolveOwnerHolder({ configValue: undefined })).toBe('self');
-  });
-
-  test('uses config value when set and no override', () => {
-    expect(resolveOwnerHolder({ configValue: 'people/charlie-example' }))
-      .toBe('people/charlie-example');
-  });
-
-  test('override beats config and default', () => {
-    expect(resolveOwnerHolder({ override: 'world', configValue: 'people/charlie-example' }))
-      .toBe('world');
-  });
-});