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/models.ts

@@ -50,10 +50,6 @@ const PER_TASK_KEYS: Array<{ key: string; tier: ModelTier; description: string }
   { key: 'models.eval.contradictions_judge', tier: 'utility',  description: 'Contradiction probe judge (v0.34 temporal-aware)' },
   { key: 'models.expansion',                tier: 'utility',   description: 'Query expansion for hybrid search' },
   { key: 'models.chat',                     tier: 'reasoning', description: 'Default `gateway.chat()` model' },
-  { key: 'models.propose_takes',            tier: 'reasoning', description: 'propose_takes claim extractor' },
-  { key: 'models.grade_takes',              tier: 'reasoning', description: 'grade_takes verdict judge' },
-  { key: 'models.calibration_profile',      tier: 'reasoning', description: 'Calibration profile generator' },
-  { key: 'models.brainstorm',               tier: 'reasoning', description: '`gbrain brainstorm` orchestrator' },
 ];
 
 interface ModelEntry {

src/core/brainstorm/judges.ts

@@ -61,10 +61,7 @@ export const MAX_OUTPUT_TOKENS_CEIL = 32_000;
  * (with a readable error) instead of the provider's opaque HTTP 400.
  */
 export const ANTHROPIC_OUTPUT_CAPS: Record<string, number> = {
-  'claude-fable-5': 64_000,
-  'claude-opus-4-8': 32_000,
   'claude-opus-4-7': 32_000,
-  'claude-sonnet-5': 64_000,
   'claude-sonnet-4-6': 64_000,
   'claude-haiku-4-5': 64_000,
   'claude-haiku-4-5-20251001': 64_000,

src/core/brainstorm/orchestrator.ts

@@ -32,7 +32,6 @@
  */
 
 import type { BrainEngine } from '../engine.ts';
-import { resolveModel } from '../model-config.ts';
 import { chat as defaultChat, embedQuery, type ChatResult, type ChatOpts } from '../ai/gateway.ts';
 import { hybridSearch, hybridSearchCached } from '../search/hybrid.ts';
 import { fetchFar, type CloseRef, type FarPage } from './domain-bank.ts';
@@ -539,14 +538,7 @@ async function _runBrainstormInner(
   const embedFn = opts.embedQueryFn ?? embedQuery;
 
   // ---- Phase 0: cost preview + TTY grace ----
-  // Tier-resolved (mirrors the cycle phases): honors models.brainstorm >
-  // models.default > models.tier.reasoning; the fallback keeps stock
-  // behavior identical (reasoning tier default IS claude-sonnet-4-6).
-  const modelStr = opts.modelOverride ?? await resolveModel(engine, {
-    configKey: 'models.brainstorm',
-    tier: 'reasoning',
-    fallback: 'anthropic:claude-sonnet-4-6',
-  });
+  const modelStr = opts.modelOverride ?? 'anthropic:claude-sonnet-4-6';
   const { aborted, estimate } = await previewCostAndWait({
     profile,
     model: modelStr,

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

@@ -26,8 +26,8 @@
  */
 
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
-import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
-import { resolveModel } from '../model-config.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';
 import { patternStatementTemplate, type PatternStatementSlots } from '../calibration/templates.ts';
 // v0.41 T10 — domain widening. The aggregator module resolves the active
@@ -229,16 +229,7 @@ class CalibrationProfilePhase extends BaseCyclePhase {
   ): Promise<{ summary: string; details: Record<string, unknown>; status?: PhaseStatus }> {
     const holder = opts.holder ?? 'garry';
     const promptVersion = opts.promptVersion ?? CALIBRATION_PROFILE_PROMPT_VERSION;
-    // Resolved once (see propose-takes.ts for the chain): models.calibration_profile
-    // > models.default > env > the gateway's chat model (itself resolved
-    // through models.chat + the reasoning tier). Provider-prefixed per #2451
-    // — a bare id would make gateway.chat() throw "missing a provider
-    // prefix". Drives the generator's chat call, the budget label, and the
-    // persisted model_id, so the three can never disagree.
-    const modelId = opts.model ?? await resolveModel(engine, {
-      configKey: 'models.calibration_profile',
-      fallback: getChatModel(),
-    });
+    const modelId = opts.model ?? TIER_DEFAULTS.reasoning;
     const gradeCompletion = opts.gradeCompletion ?? 1.0;
     const patternsGenerator = opts.patternsGenerator ?? defaultPatternsGenerator;
     const biasTagsGenerator = opts.biasTagsGenerator ?? defaultBiasTagsGenerator;
@@ -274,7 +265,6 @@ class CalibrationProfilePhase extends BaseCyclePhase {
         scorecard,
         holder,
         attempt,
-        modelHint: modelId,
         ...(feedback !== undefined ? { feedback } : {}),
       });
       return lines.join('\n');

src/core/cycle/grade-takes.ts

@@ -36,9 +36,7 @@
 
 import { createHash } from 'node:crypto';
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
-import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
-import { resolveModel } from '../model-config.ts';
-import { splitProviderModelId } from '../model-id.ts';
+import { chat as gatewayChat } from '../ai/gateway.ts';
 import { GBrainError } from '../types.ts';
 import type { OperationContext } from '../operations.ts';
 import type { BrainEngine, Take, TakeResolution } from '../engine.ts';
@@ -397,24 +395,7 @@ class GradeTakesPhase extends BaseCyclePhase {
     const autoResolve = opts.autoResolve ?? false; // D17 default OFF
     const autoResolveThreshold = opts.autoResolveThreshold ?? 0.95; // D12 conservative
     const resolvedByLabel = opts.resolvedByLabel ?? 'gbrain:grade_takes';
-    // Resolve the judge model ONCE (see propose-takes.ts for the chain —
-    // same label-vs-actual split fixed here: the judge call rode the
-    // gateway's chat_model while the grade cache key, evidence signature,
-    // and budget label recorded a hardcoded 4.6).
-    // NOTE: changing the resolved judge model invalidates the grade cache
-    // (judge_model_id is part of its key) — a one-time, budget-capped
-    // re-grade wave that is CORRECT, since the actual judge did change.
-    const judgeModelFull = opts.model ?? await resolveModel(engine, {
-      configKey: 'models.grade_takes',
-      fallback: getChatModel(),
-    });
-    // Bare tail for cache keys / evidence signatures / stored ids — the
-    // grade cache has always been keyed on bare ids; the gateway default
-    // resolves provider-prefixed, and normalizing preserves cache continuity
-    // on stock installs (no spurious re-judge wave from a prefix change). A
-    // genuinely different configured judge still invalidates, which is
-    // correct. The FULL string drives the actual judge call.
-    const judgeModelId = splitProviderModelId(judgeModelFull).model || judgeModelFull;
+    const judgeModelId = opts.model ?? 'claude-sonnet-4-6';
 
     const useEnsemble = opts.useEnsemble ?? false;
     const ensembleThreshold = opts.ensembleThreshold ?? 0.85;
@@ -487,7 +468,7 @@ class GradeTakesPhase extends BaseCyclePhase {
       // Call the single-model judge. Errors on a single take log warning + continue.
       let verdict: JudgeVerdict;
       try {
-        verdict = await judge({ take, evidence, modelHint: judgeModelFull });
+        verdict = await judge({ take, evidence, modelHint: opts.model });
       } catch (err) {
         const msg = err instanceof Error ? err.message : String(err);
         result.warnings.push(`judge failed on take ${take.id}: ${msg}`);

src/core/cycle/propose-takes.ts

@@ -5,16 +5,11 @@
  * a tuned LLM extractor, writes the extracted gradeable claims to the
  * `take_proposals` queue. User accepts/rejects via `gbrain takes propose`.
  *
- * Idempotency contract (D17 schema spec; per-claim rows since migration v125):
- *   Every scan of a (source_id, page_slug, content_hash, prompt_version)
- *   tuple leaves at least one row — one per extracted claim, or a single
- *   status='empty' sentinel when extraction yields nothing — so an unchanged
- *   page never re-spends LLM tokens. Pre-v125 only proposal rows were
- *   written: a zero-claim page never entered the cache and was re-extracted
- *   on EVERY cycle (observed live: ~60 such pages × every cycle ≈ 1,400
- *   wasted extractor calls / ~$15 per day — ~90% of total autopilot spend).
- *   Bumping PROPOSE_TAKES_PROMPT_VERSION cleanly invalidates the cache so a
- *   tuned prompt re-runs proposals on every page.
+ * Idempotency contract (D17 schema spec):
+ *   The unique index on (source_id, page_slug, content_hash, prompt_version)
+ *   means an unchanged page never re-spends LLM tokens. Bumping
+ *   PROPOSE_TAKES_PROMPT_VERSION cleanly invalidates the cache so a tuned
+ *   prompt re-runs proposals on every page.
  *
  * F2 fence dedup:
  *   The phase reads the page's existing `<!-- gbrain:takes:begin -->` fence
@@ -45,7 +40,6 @@
 import { randomUUID, createHash } from 'node:crypto';
 import { BaseCyclePhase, type ScopedReadOpts, type BasePhaseOpts } from './base-phase.ts';
 import { chat as gatewayChat, getChatModel } from '../ai/gateway.ts';
-import { resolveModel } from '../model-config.ts';
 import { writeReceipt } from '../extract/receipt-writer.ts';
 import { upsertExtractRollup } from '../extract/rollup-writer.ts';
 import { GBrainError } from '../types.ts';
@@ -313,18 +307,6 @@ class ProposeTakesPhase extends BaseCyclePhase {
     const promptVersion = opts.promptVersion ?? PROPOSE_TAKES_PROMPT_VERSION;
     const pageLimit = opts.pageLimit ?? 100;
     const skipPagesWithFence = opts.skipPagesWithFence ?? false;
-    // Resolve the extractor model ONCE: models.propose_takes >
-    // models.default > GBRAIN_MODEL env > the gateway's chat model (which
-    // reconfigureGatewayWithEngine already resolved through models.chat +
-    // the reasoning tier). One resolved provider-prefixed string drives the
-    // actual chat call, the budget estimate, AND the stored model_id — so
-    // the recorded model can never disagree with the model that ran (#2451
-    // convention: stored ids are provider-prefixed, nested prefixes like
-    // openrouter:anthropic/... stay intact).
-    const extractorModelId = opts.model ?? await resolveModel(engine, {
-      configKey: 'models.propose_takes',
-      fallback: getChatModel(),
-    });
     const proposalRunId = `propose-${new Date().toISOString().slice(0, 19).replace(/[-:T]/g, '')}-${randomUUID().slice(0, 8)}`;
 
     const result: ProposeTakesResult = {
@@ -348,6 +330,8 @@ class ProposeTakesPhase extends BaseCyclePhase {
       opts.reporter.start('propose_takes.pages' as never, pages.length);
     }
 
+    const modelId = opts.model ?? getChatModel();
+
     for (const page of pages) {
       result.pages_scanned += 1;
       this.tick(opts);
@@ -377,7 +361,7 @@ class ProposeTakesPhase extends BaseCyclePhase {
 
       // Budget pre-check before the LLM call. Estimate: ~1500 input tokens + 500 output.
       const budget = this.checkBudget({
-        modelId: extractorModelId,
+        modelId,
         estimatedInputTokens: 1500,
         maxOutputTokens: 500,
       });
@@ -396,7 +380,7 @@ class ProposeTakesPhase extends BaseCyclePhase {
           pagePath: page.slug,
           pageBody: body,
           existingTakes,
-          modelHint: extractorModelId,
+          modelHint: opts.model,
         });
       } catch (err) {
         const msg = err instanceof Error ? err.message : String(err);
@@ -404,42 +388,16 @@ class ProposeTakesPhase extends BaseCyclePhase {
         continue;
       }
 
-      // Zero-claim scans MUST still enter the idempotency cache. Pre-v125
-      // only proposal rows were written, so a page whose extraction yielded
-      // no gradeable claims never got a row for its (page, content_hash) —
-      // the cache check above missed on every subsequent cycle and the LLM
-      // call was re-spent on the same unchanged page, forever. The sentinel
-      // row (status='empty', empty claim_text) is invisible to the review
-      // queue (pending_idx is partial on status='pending'); it exists only
-      // so the cache check hits.
-      if (proposals.length === 0) {
-        await engine.executeRaw(
-          `INSERT INTO take_proposals
-             (source_id, page_slug, content_hash, prompt_version, proposal_run_id,
-              status, claim_text, kind, holder, weight, domain, dedup_against_fence_rows, model_id)
-           VALUES ($1, $2, $3, $4, $5, 'empty', '', 'none', 'brain', 0, NULL, NULL, $6)
-           ON CONFLICT (source_id, page_slug, content_hash, prompt_version, md5(claim_text)) DO NOTHING`,
-          [sourceId, page.slug, ch, promptVersion, proposalRunId, extractorModelId],
-        );
-        continue;
-      }
-
-      // Write proposals to take_proposals, one row per claim. The v125
-      // idempotency index includes md5(claim_text), so a same-page
-      // multi-claim run keeps EVERY claim — the pre-v125 four-column unique
-      // index made claims 2..N conflict with claim 1 and ON CONFLICT DO
-      // NOTHING silently dropped them (the review queue only ever saw the
-      // first claim of each page version). RETURNING id keeps
-      // proposals_inserted honest: it counts rows that actually landed,
-      // not insert attempts.
+      // Write proposals to take_proposals. Each row is a separate INSERT
+      // because the composite idempotency key is on the per-page tuple — a
+      // bulk UPSERT would collapse a same-page-multi-claim run into one row.
       for (const p of proposals) {
-        const landed = await engine.executeRaw<{ id: number }>(
+        await engine.executeRaw(
           `INSERT INTO take_proposals
              (source_id, page_slug, content_hash, prompt_version, proposal_run_id,
               claim_text, kind, holder, weight, domain, dedup_against_fence_rows, model_id)
            VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12)
-           ON CONFLICT (source_id, page_slug, content_hash, prompt_version, md5(claim_text)) DO NOTHING
-           RETURNING id`,
+           ON CONFLICT (source_id, page_slug, content_hash, prompt_version) DO NOTHING`,
           [
             sourceId,
             page.slug,
@@ -452,10 +410,10 @@ class ProposeTakesPhase extends BaseCyclePhase {
             p.weight,
             p.domain ?? null,
             JSON.stringify(existingTakes),
-            extractorModelId,
+            modelId,
           ],
         );
-        if (landed.length > 0) result.proposals_inserted += 1;
+        result.proposals_inserted += 1;
       }
     }
 

src/core/cycle/synthesize.ts

@@ -58,11 +58,8 @@ const SUMMARY_SLUG_RE = /^[a-z0-9][a-z0-9\-]*(\/[a-z0-9][a-z0-9\-]*)*$/;
  * resolver returns for known Anthropic aliases.
  */
 const MODEL_CONTEXT_TOKENS: Record<string, number> = {
-  'claude-fable-5': 1_000_000,
-  'claude-opus-4-8': 1_000_000,
   'claude-opus-4-7': 1_000_000,
   'claude-opus-4-6': 1_000_000,
-  'claude-sonnet-5': 1_000_000,
   'claude-sonnet-4-6': 200_000,
   'claude-sonnet-4-5': 200_000,
   'claude-haiku-4-5-20251001': 200_000,

src/core/migrate.ts

@@ -5671,41 +5671,6 @@ export const MIGRATIONS: Migration[] = [
 `);
     },
   },
-  {
-    version: 125,
-    name: 'take_proposals_empty_scan_sentinels_and_per_claim_rows',
-    // v0.42.x — kill the propose_takes rescan loop + stop dropping claims.
-    //
-    // Two defects, one schema touch:
-    //   1. Zero-claim scans never entered the idempotency cache (only
-    //      proposal rows were written), so pages whose extraction yielded
-    //      nothing were re-extracted on EVERY cycle. Observed live: ~60
-    //      such pages per run ≈ 1,400 wasted extractor calls / ~$15 per
-    //      day — ~90% of total autopilot LLM spend. Fix: the phase now
-    //      writes a status='empty' sentinel row per zero-claim scan; the
-    //      status CHECK gains the 'empty' value. Sentinels are excluded
-    //      from the partial pending index, so the review queue never sees
-    //      them.
-    //   2. The 4-column unique index collapsed a same-page multi-claim run
-    //      to its FIRST claim: rows 2..N conflicted and ON CONFLICT DO
-    //      NOTHING silently dropped them (verified live: exactly one row
-    //      per (page, hash) across 3 days of runs). Fix: the idempotency
-    //      index gains md5(claim_text) — per-claim rows, while the
-    //      4-column prefix still serves the per-scan cache lookup.
-    //
-    // Existing data is index-safe by construction: the old index guaranteed
-    // at most one row per 4-tuple, so the widened index has no duplicates
-    // to trip on. The DROP+ADD CONSTRAINT pair is idempotent as a unit.
-    idempotent: true,
-    sql: `
-      ALTER TABLE take_proposals DROP CONSTRAINT IF EXISTS take_proposals_status_check;
-      ALTER TABLE take_proposals ADD CONSTRAINT take_proposals_status_check
-        CHECK (status IN ('pending','accepted','rejected','superseded','empty'));
-      DROP INDEX IF EXISTS take_proposals_idempotency_idx;
-      CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-        ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
-    `,
-  },
 ];
 
 export const LATEST_VERSION = MIGRATIONS.length > 0

src/core/pglite-schema.ts

@@ -762,7 +762,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
   proposal_run_id             TEXT         NOT NULL,
   status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
   claim_text                  TEXT         NOT NULL,
   kind                        TEXT         NOT NULL,
   holder                      TEXT         NOT NULL,
@@ -777,7 +777,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
   ON take_proposals (source_id, status, proposed_at DESC)
   WHERE status = 'pending';

src/core/schema-embedded.ts

@@ -1274,14 +1274,8 @@ CREATE INDEX IF NOT EXISTS calibration_profiles_published_idx
   WHERE published = true;
 
 -- take_proposals: propose_takes phase queue. Idempotency cache via the
--- composite unique index (source_id, page_slug, content_hash, prompt_version,
--- md5(claim_text)) — the 4-column prefix is the per-scan cache key (mirrors
--- v0.23 dream_verdicts); md5(claim_text) makes rows per-claim so multi-claim
--- pages keep every claim (v125). status='empty' rows are zero-claim scan
--- sentinels: they hold the cache slot for a page version whose extraction
--- yielded nothing — the phase never re-spends the LLM call on that page
--- version. Excluded from the partial pending index. proposal_run_id supports
--- --rollback by run.
+-- composite unique index (source_id, page_slug, content_hash, prompt_version)
+-- mirrors v0.23 dream_verdicts. proposal_run_id supports --rollback by run.
 CREATE TABLE IF NOT EXISTS take_proposals (
   id                          BIGSERIAL PRIMARY KEY,
   source_id                   TEXT         NOT NULL REFERENCES sources(id) ON DELETE CASCADE,
@@ -1292,7 +1286,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
   proposal_run_id             TEXT         NOT NULL,
   status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
   claim_text                  TEXT         NOT NULL,
   kind                        TEXT         NOT NULL,
   holder                      TEXT         NOT NULL,
@@ -1307,7 +1301,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
   ON take_proposals (source_id, status, proposed_at DESC)
   WHERE status = 'pending';

src/schema.sql

@@ -1270,14 +1270,8 @@ CREATE INDEX IF NOT EXISTS calibration_profiles_published_idx
   WHERE published = true;
 
 -- take_proposals: propose_takes phase queue. Idempotency cache via the
--- composite unique index (source_id, page_slug, content_hash, prompt_version,
--- md5(claim_text)) — the 4-column prefix is the per-scan cache key (mirrors
--- v0.23 dream_verdicts); md5(claim_text) makes rows per-claim so multi-claim
--- pages keep every claim (v125). status='empty' rows are zero-claim scan
--- sentinels: they hold the cache slot for a page version whose extraction
--- yielded nothing — the phase never re-spends the LLM call on that page
--- version. Excluded from the partial pending index. proposal_run_id supports
--- --rollback by run.
+-- composite unique index (source_id, page_slug, content_hash, prompt_version)
+-- mirrors v0.23 dream_verdicts. proposal_run_id supports --rollback by run.
 CREATE TABLE IF NOT EXISTS take_proposals (
   id                          BIGSERIAL PRIMARY KEY,
   source_id                   TEXT         NOT NULL REFERENCES sources(id) ON DELETE CASCADE,
@@ -1288,7 +1282,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   proposed_at                 TIMESTAMPTZ  NOT NULL DEFAULT now(),
   proposal_run_id             TEXT         NOT NULL,
   status                      TEXT         NOT NULL DEFAULT 'pending'
-                                           CHECK (status IN ('pending','accepted','rejected','superseded','empty')),
+                                           CHECK (status IN ('pending','accepted','rejected','superseded')),
   claim_text                  TEXT         NOT NULL,
   kind                        TEXT         NOT NULL,
   holder                      TEXT         NOT NULL,
@@ -1303,7 +1297,7 @@ CREATE TABLE IF NOT EXISTS take_proposals (
   predicted_brier_bucket_n    INTEGER
 );
 CREATE UNIQUE INDEX IF NOT EXISTS take_proposals_idempotency_idx
-  ON take_proposals (source_id, page_slug, content_hash, prompt_version, md5(claim_text));
+  ON take_proposals (source_id, page_slug, content_hash, prompt_version);
 CREATE INDEX IF NOT EXISTS take_proposals_pending_idx
   ON take_proposals (source_id, status, proposed_at DESC)
   WHERE status = 'pending';

test/calibration-profile.test.ts

@@ -38,7 +38,6 @@ function buildMockEngine(opts: { scorecard: TakesScorecard }): {
 } {
   const captured: CapturedSql[] = [];
   const engine = {
-    async getConfig() { return null; },
     kind: 'pglite',
     async getScorecard() {
       return opts.scorecard;

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/grade-takes-ensemble.test.ts

@@ -47,7 +47,6 @@ function buildMockEngine(opts: { takes: Take[] }): {
   const captured: CapturedSql[] = [];
   const resolves: CapturedResolve[] = [];
   const engine = {
-    async getConfig() { return null; },
     kind: 'pglite',
     async listTakes() {
       return opts.takes;

test/grade-takes.test.ts

@@ -46,14 +46,12 @@ interface CapturedResolve {
 function buildMockEngine(opts: {
   takes: Take[];
   cachedGrades?: Set<string>; // composite-key strings already in take_grade_cache
-  config?: Record<string, string>; // engine.getConfig plane (models.grade_takes etc.)
 }): { engine: BrainEngine; captured: CapturedSql[]; resolves: CapturedResolve[] } {
   const captured: CapturedSql[] = [];
   const resolves: CapturedResolve[] = [];
   const cached = opts.cachedGrades ?? new Set<string>();
 
   const engine = {
-    async getConfig(key: string) { return opts.config?.[key] ?? null; },
     kind: 'pglite',
     async listTakes() {
       return opts.takes;
@@ -226,28 +224,6 @@ describe('runPhaseGradeTakes — phase integration', () => {
     expect(resolves).toHaveLength(0); // no canonical mutation
   });
 
-  test('models.grade_takes config drives the judge call; cache key stays bare-tailed', async () => {
-    // Pre-fix the judge call rode the gateway's chat_model while the cache
-    // key / budget label recorded a hardcoded 'claude-sonnet-4-6'. The phase
-    // now resolves models.grade_takes; the judge gets the FULL string and
-    // the cache row keys on the bare tail (continuity with historical rows).
-    const takes = [buildTake({ id: 1, sinceDate: '2023-01-01' })];
-    const { engine, captured } = buildMockEngine({
-      takes,
-      config: { 'models.grade_takes': 'anthropic:claude-sonnet-5' },
-    });
-    const hints: Array<string | undefined> = [];
-    const judge: JudgeFn = async ({ modelHint }) => {
-      hints.push(modelHint);
-      return { verdict: 'correct', confidence: 0.9, reasoning: 'held' };
-    };
-    const result = await runPhaseGradeTakes(buildCtx(engine), { judge });
-    expect(result.status).toBe('ok');
-    expect(hints).toEqual(['anthropic:claude-sonnet-5']); // actual call gets the FULL string
-    const inserts = captured.filter(c => c.sql.includes('INSERT INTO take_grade_cache'));
-    expect(inserts[0]!.params[2]).toBe('claude-sonnet-5'); // judge_model_id is the bare tail
-  });
-
   test('D17: auto-resolve OFF by default — even high-confidence verdict does NOT mutate takes', async () => {
     const takes = [buildTake({ id: 1, sinceDate: '2023-01-01' })];
     const { engine, resolves } = buildMockEngine({ takes });

test/propose-takes-rescan.test.ts

@@ -1,160 +0,0 @@
-/**
- * propose_takes rescan-loop + dropped-claims regression tests (migration v125).
- *
- * Two live-observed defects, both fixed by the v125 schema + phase change:
- *
- *   1. RESCAN LOOP — a page whose extraction yielded zero claims never
- *      entered the idempotency cache (only proposal rows were written), so
- *      every cycle re-spent the extractor call on the same unchanged page.
- *      Live impact: ~60 such pages × every cycle ≈ 1,400 wasted LLM calls
- *      (~$15) per day — ~90% of total autopilot spend. Fix: status='empty'
- *      sentinel row per zero-claim scan.
- *
- *   2. DROPPED CLAIMS — the 4-column unique index collapsed a same-page
- *      multi-claim run to its first claim (rows 2..N conflicted, ON CONFLICT
- *      DO NOTHING dropped them silently; verified live: exactly 1 row per
- *      (page, hash) over 3 days). Fix: idempotency index gains
- *      md5(claim_text).
- *
- * Hermetic: PGLite engine + injected extractor; no gateway, no LLM.
- */
-import { describe, test, expect, beforeAll, afterAll } from 'bun:test';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { runPhaseProposeTakes, type ProposeTakesExtractor, type ProposedTake } from '../src/core/cycle/propose-takes.ts';
-import type { OperationContext } from '../src/core/operations.ts';
-
-let engine: PGLiteEngine;
-
-function ctx(): OperationContext {
-  return {
-    engine,
-    remote: false,
-    config: {} as OperationContext['config'],
-    logger: { info() {}, warn() {}, error() {}, debug() {} } as unknown as OperationContext['logger'],
-  } as unknown as OperationContext;
-}
-
-/** Extractor stub that counts invocations per page slug. */
-function countingExtractor(
-  claimsBySlug: Record<string, ProposedTake[]>,
-): { extractor: ProposeTakesExtractor; calls: string[] } {
-  const calls: string[] = [];
-  const extractor: ProposeTakesExtractor = async ({ pagePath }) => {
-    calls.push(pagePath);
-    return claimsBySlug[pagePath] ?? [];
-  };
-  return { extractor, calls };
-}
-
-beforeAll(async () => {
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-  await engine.putPage('notes/zero-claims', {
-    type: 'note',
-    title: 'pure narrative',
-    compiled_truth: 'A quiet walk in the park. Nothing opinionated happened at all today.',
-  });
-  await engine.putPage('notes/three-claims', {
-    type: 'note',
-    title: 'opinionated',
-    compiled_truth: 'I bet acme-example wins the market. widget-co will struggle. fund-a is overexposed.',
-  });
-});
-
-afterAll(async () => {
-  await engine.disconnect();
-});
-
-describe('rescan loop — zero-claim scans enter the cache', () => {
-  test('second run cache-hits: extractor is NOT called again on unchanged pages', async () => {
-    const claims = {
-      'notes/three-claims': [
-        { claim_text: 'acme-example wins the market', kind: 'bet' as const, holder: 'brain', weight: 0.7 },
-        { claim_text: 'widget-co will struggle', kind: 'take' as const, holder: 'brain', weight: 0.6 },
-        { claim_text: 'fund-a is overexposed', kind: 'take' as const, holder: 'brain', weight: 0.55 },
-      ],
-    };
-
-    const first = countingExtractor(claims);
-    const r1 = await runPhaseProposeTakes(ctx(), { extractor: first.extractor });
-    expect(r1.status).toBe('ok');
-    // Both pages extracted on the first pass.
-    expect(first.calls).toContain('notes/zero-claims');
-    expect(first.calls).toContain('notes/three-claims');
-
-    const second = countingExtractor(claims);
-    const r2 = await runPhaseProposeTakes(ctx(), { extractor: second.extractor });
-    expect(r2.status).toBe('ok');
-    // THE regression: pre-fix the zero-claim page missed the cache every
-    // run and was re-extracted here. (Run 1's receipt page legitimately
-    // appears once — it's a new page — and its zero-claim scan now caches
-    // too; pre-fix, receipts re-scanned forever as well.)
-    expect(second.calls).not.toContain('notes/zero-claims');
-    expect(second.calls).not.toContain('notes/three-claims');
-
-    // Run 2 inserted nothing, so no new receipt page exists: run 3 must be
-    // fully quiescent — zero extractor calls, zero cache misses.
-    const third = countingExtractor(claims);
-    const r3 = await runPhaseProposeTakes(ctx(), { extractor: third.extractor });
-    expect(r3.status).toBe('ok');
-    expect(third.calls).toEqual([]);
-    expect((r3.details as Record<string, unknown>).cache_misses).toBe(0);
-  });
-
-  test('zero-claim scan wrote an "empty" sentinel invisible to the pending queue', async () => {
-    const sentinel = await engine.executeRaw<{ status: string; claim_text: string }>(
-      `SELECT status, claim_text FROM take_proposals WHERE page_slug = 'notes/zero-claims'`,
-      [],
-    );
-    expect(sentinel.length).toBe(1);
-    expect(sentinel[0].status).toBe('empty');
-    expect(sentinel[0].claim_text).toBe('');
-    const pending = await engine.executeRaw<{ n: number }>(
-      `SELECT COUNT(*)::int AS n FROM take_proposals WHERE page_slug = 'notes/zero-claims' AND status = 'pending'`,
-      [],
-    );
-    expect(Number(pending[0].n)).toBe(0);
-  });
-});
-
-describe('dropped claims — per-claim rows survive the idempotency index', () => {
-  test('a 3-claim page stores 3 rows and reports an honest inserted count', async () => {
-    const rows = await engine.executeRaw<{ claim_text: string }>(
-      `SELECT claim_text FROM take_proposals WHERE page_slug = 'notes/three-claims' AND status = 'pending' ORDER BY id`,
-      [],
-    );
-    // Pre-fix the 4-column unique index kept only the FIRST claim.
-    expect(rows.length).toBe(3);
-    expect(rows.map(r => r.claim_text)).toEqual([
-      'acme-example wins the market',
-      'widget-co will struggle',
-      'fund-a is overexposed',
-    ]);
-  });
-
-  test('content change re-extracts and stores the new version separately', async () => {
-    await engine.putPage('notes/zero-claims', {
-      type: 'note',
-      title: 'pure narrative',
-      compiled_truth: 'Updated: I now believe acme-example is undervalued and will re-rate within a year.',
-    });
-    const claims = {
-      'notes/zero-claims': [
-        { claim_text: 'acme-example is undervalued', kind: 'take' as const, holder: 'brain', weight: 0.6 },
-      ],
-    };
-    const run = countingExtractor(claims);
-    const r = await runPhaseProposeTakes(ctx(), { extractor: run.extractor });
-    expect(r.status).toBe('ok');
-    // Changed page re-extracts; the unchanged 3-claim page stays cached.
-    expect(run.calls).toEqual(['notes/zero-claims']);
-    expect((r.details as Record<string, unknown>).proposals_inserted).toBe(1);
-    const all = await engine.executeRaw<{ status: string }>(
-      `SELECT status FROM take_proposals WHERE page_slug = 'notes/zero-claims' ORDER BY id`,
-      [],
-    );
-    // Old hash's sentinel + new hash's pending claim coexist.
-    expect(all.map(r2 => r2.status).sort()).toEqual(['empty', 'pending']);
-  });
-});

test/propose-takes.test.ts

@@ -41,16 +41,12 @@ interface CapturedSql {
 function buildMockEngine(opts: {
   pages: Page[];
   existingProposals?: Set<string>; // composite-key strings already in take_proposals
-  config?: Record<string, string>; // engine.getConfig plane (models.tier.* etc.)
 }): { engine: BrainEngine; captured: CapturedSql[] } {
   const captured: CapturedSql[] = [];
   const existing = opts.existingProposals ?? new Set<string>();
 
   const engine = {
     kind: 'pglite',
-    async getConfig(key: string) {
-      return opts.config?.[key] ?? null;
-    },
     async listPages() {
       return opts.pages;
     },
@@ -63,12 +59,7 @@ function buildMockEngine(opts: {
         if (existing.has(key)) return [{ id: 1 } as unknown as T];
         return [];
       }
-      // INSERT ... RETURNING id — emulate a successful insert so the
-      // honest proposals_inserted counter (counts RETURNING rows, not
-      // attempts) sees the row land. Conflicted inserts would return [].
-      if (sql.includes('INSERT INTO take_proposals') && sql.includes('RETURNING id')) {
-        return [{ id: 1 } as unknown as T];
-      }
+      // INSERT — return nothing
       return [];
     },
   } as unknown as BrainEngine;
@@ -276,29 +267,6 @@ describe('runPhaseProposeTakes — phase integration', () => {
     expect(inserts[0]!.params[9]).toBe('market'); // domain
   });
 
-  test('extractor model resolves through models.propose_takes config', async () => {
-    // Pre-fix the phase's only model knob was the gateway chat model — there
-    // was no per-phase config key. The phase now resolves once via
-    // resolveModel(models.propose_takes > models.default > env > gateway
-    // chat model); the extractor hint and the stored model_id must both
-    // reflect the configured override (full provider-prefixed string, #2451).
-    const pages = [buildPage({ slug: 'wiki/concepts/tier-routing', body: 'Tier-routed models will win.' })];
-    const { engine, captured } = buildMockEngine({
-      pages,
-      config: { 'models.propose_takes': 'anthropic:claude-sonnet-5' },
-    });
-    const seen: Array<string | undefined> = [];
-    const extractor: ProposeTakesExtractor = async ({ modelHint }) => {
-      seen.push(modelHint);
-      return [{ claim_text: 'tier-routed models win', kind: 'bet', holder: 'brain', weight: 0.7 }];
-    };
-    const result = await runPhaseProposeTakes(buildCtx(engine), { extractor });
-    expect(result.status).toBe('ok');
-    expect(seen).toEqual(['anthropic:claude-sonnet-5']); // chat call gets the FULL string
-    const inserts = captured.filter(c => c.sql.includes('INSERT INTO take_proposals'));
-    expect(inserts[0]!.params[11]).toBe('anthropic:claude-sonnet-5'); // stored model_id matches the call
-  });
-
   test('cache hit: page already in take_proposals is skipped', async () => {
     const body = 'A page that was already processed.';
     const pages = [buildPage({ slug: 'wiki/old-page', body })];