fix(chunker): estimated-token hard cap — URL-dense/CJK-fallback chunks overflow strict embedding-server token limits

Takeover of #2847 (rebased onto current master). Fixes #2826.

- cjk.ts: estimateEmbeddingTokens() — conservative per-char-class token
  estimate (CJK 1.0, other 0.75, whitespace 0.1 per code unit).
- recursive.ts (MARKDOWN_CHUNKER_VERSION 3→4): countWords floored at
  ceil(nonWhitespaceChars/6); capByEstimatedTokens() final pass with
  ChunkOptions.maxTokens (default 1500).
- code.ts (CHUNKER_VERSION 4→5): capCodeChunks() applies the same cap to
  AST-path chunks that splitLargeNode can't subdivide.

Co-authored-by: paul-0320 <paul-0320@users.noreply.github.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/cli.ts

@@ -54,7 +54,7 @@ export function bigintToStringReplacer(_key: string, value: unknown): unknown {
 }
 
 // CLI-only commands that bypass the operation layer
-export const CLI_ONLY = new Set(['init', 'reinit-pglite', 'upgrade', 'post-upgrade', 'check-update', 'integrations', 'publish', 'check-backlinks', 'lint', 'report', 'import', 'export', 'files', 'embed', 'serve', 'call', 'config', 'doctor', 'migrate', 'eval', 'sync', 'extract', 'extract-conversation-facts', 'enrich', 'features', 'autopilot', 'graph-query', 'jobs', 'agent', 'apply-migrations', 'skillpack-check', 'skillpack', 'resolvers', 'integrity', 'repair-jsonb', 'orphans', 'sources', 'mounts', 'dream', 'check-resolvable', 'routing-eval', 'skillify', 'smoke-test', 'providers', 'storage', 'repos', 'code-def', 'code-refs', 'reindex', 'reindex-code', 'reindex-frontmatter', 'code-callers', 'code-callees', 'reconcile-links', 'frontmatter', 'auth', 'friction', 'claw-test', 'book-mirror', 'takes', 'think', 'salience', 'anomalies', 'calibration', 'transcripts', 'models', 'remote', 'recall', 'forget', 'edges-backfill', 'facts', 'cache', 'ze-switch', 'founder', 'brainstorm', 'lsd', 'schema', 'capture', 'onboard', 'conversation-parser', 'status', 'connect', 'skillopt', 'quarantine', 'self-upgrade', 'advisor', 'watch', 'reindex-search-vector']);
+export const CLI_ONLY = new Set(['init', 'reinit-pglite', 'upgrade', 'post-upgrade', 'check-update', 'integrations', 'publish', 'check-backlinks', 'lint', 'report', 'import', 'export', 'files', 'embed', 'serve', 'call', 'config', 'doctor', 'migrate', 'eval', 'sync', 'extract', 'extract-conversation-facts', 'enrich', 'features', 'autopilot', 'graph-query', 'jobs', 'agent', 'apply-migrations', 'skillpack-check', 'skillpack', 'resolvers', 'integrity', 'repair-jsonb', 'orphans', 'sources', 'mounts', 'dream', 'check-resolvable', 'routing-eval', 'skillify', 'smoke-test', 'providers', 'storage', 'repos', 'code-def', 'code-refs', 'reindex', 'reindex-code', 'reindex-frontmatter', 'code-callers', 'code-callees', 'reconcile-links', 'frontmatter', 'auth', 'friction', 'claw-test', 'book-mirror', 'takes', 'think', 'salience', 'anomalies', 'calibration', 'transcripts', 'models', 'remote', 'recall', 'forget', 'edges-backfill', 'cache', 'ze-switch', 'founder', 'brainstorm', 'lsd', 'schema', 'capture', 'onboard', 'conversation-parser', 'status', 'connect', 'skillopt', 'quarantine', 'self-upgrade', 'advisor', 'watch', 'reindex-search-vector']);
 // CLI-only commands whose handlers print their own --help text. These are
 // excluded from the generic short-circuit so detailed per-command and
 // per-subcommand usage stays reachable.
@@ -991,8 +991,6 @@ const THIN_CLIENT_REFUSED_COMMANDS = new Set([
   // hint pointing at the routable MCP tools; per-subcommand splits are
   // a v0.31.x follow-up TODO.
   'takes', 'sources',
-  // #1867: fence-backfill edits local .md fences + stamps the local DB.
-  'facts',
   // v0.32 thin-client routing audit (Codex round 2 findings #2, #4):
   // - `pages` purge-deleted is admin+localOnly (operations.ts:856-864)
   // - `files` list / file_url MCP ops are localOnly (operations.ts:1769-1879)
@@ -1028,7 +1026,6 @@ const THIN_CLIENT_REFUSE_HINTS: Record<string, string> = {
   migrate: "migrate runs on the host's local engine. Run on the host machine.",
   'apply-migrations': 'schema migrations run on the host. SSH and run there.',
   'repair-jsonb': 'repair-jsonb operates on the local DB only.',
-  facts: 'facts fence-backfill edits local entity-page fences. Run on the host machine.',
   integrity: 'integrity scans local files. Run on the host machine.',
   serve: 'serve starts a server. Run on the host, not the thin client.',
   dream: 'dream runs the autopilot cycle on the host. `gbrain remote ping` queues one. (Native `gbrain dream` thin-client routing planned for v0.31.2.)',
@@ -1858,13 +1855,6 @@ async function handleCliOnly(command: string, args: string[]) {
         await runEdgesBackfill(engine, args);
         break;
       }
-      case 'facts': {
-        // #1867 — re-runnable fence-backfill for row_num-NULL legacy fact
-        // rows (idempotent v0_32_2 phase B, exposed as an operator command).
-        const { runFactsCommand } = await import('./commands/facts.ts');
-        await runFactsCommand(engine, args);
-        break;
-      }
       case 'whoknows': {
         // v0.33 (Issue #?): expertise + relationship-proximity routing.
         // MCP op `find_experts` (read-scoped) backs the same code path; CLI
@@ -2345,7 +2335,6 @@ TOOLS
   check-backlinks <check|fix> [dir]  Find/fix missing back-links across brain
   lint <dir|file> [--fix]            Catch LLM artifacts, placeholder dates, bad frontmatter
   orphans [--json] [--count]         Find pages with no inbound wikilinks
-  facts fence-backfill [--dry-run]   Fence legacy fact rows (row_num NULL) onto entity pages
   salience [--days N] [--kind P]     v0.29: pages ranked by emotional + activity salience
   anomalies [--since D] [--sigma N]  v0.29: cohort-based statistical anomalies (tag, type)
   transcripts recent [--days N]      v0.29: recent raw .txt transcripts (local-only)

src/commands/facts.ts

@@ -1,48 +0,0 @@
-/**
- * gbrain facts — fact-store maintenance surface (#1867).
- *
- * `fence-backfill` re-runs the v0_32_2 fence-backfill phase on demand.
- * Remote `extract_facts` deposits that predate the fence-write backstop
- * (and any legacy DB-only insert) leave `row_num IS NULL` rows that the
- * cycle extract_facts guard refuses to reconcile past — previously the
- * only remedy was the one-shot v0_32_2 migration, which the ledger marks
- * complete and never re-runs. The phase is idempotent (only touches
- * `row_num IS NULL` rows), so exposing it as a command is safe to re-run
- * any time the backlog reappears.
- */
-import type { BrainEngine } from '../core/engine.ts';
-import { setCliExitVerdict } from '../core/cli-force-exit.ts';
-import { phaseBFenceFacts } from './migrations/v0_32_2.ts';
-
-function printHelp(): void {
-  process.stderr.write(
-    `Usage: gbrain facts fence-backfill [--dry-run]\n\n` +
-      `Fence-backfill: appends every legacy fact row (row_num IS NULL) to its\n` +
-      `entity page's \`## Facts\` fence and stamps row_num + source_markdown_slug\n` +
-      `back onto the DB row. Idempotent — re-runs only pick up rows still\n` +
-      `missing a fence assignment. Clears the backlog that makes the cycle's\n` +
-      `extract_facts phase skip fence→DB reconciliation.\n\n` +
-      `  --dry-run   report what would be fenced; no FS or DB writes\n`,
-  );
-}
-
-export async function runFactsCommand(engine: BrainEngine, args: string[]): Promise<void> {
-  const sub = args[0];
-  if (!sub || sub === '--help' || sub === '-h') {
-    printHelp();
-    return;
-  }
-  if (sub !== 'fence-backfill') {
-    process.stderr.write(`Unknown facts subcommand: ${sub}\n`);
-    printHelp();
-    setCliExitVerdict(1);
-    return;
-  }
-
-  const dryRun = args.includes('--dry-run');
-  const result = await phaseBFenceFacts(engine, { dryRun });
-  process.stderr.write(
-    `fence-backfill: ${result.status}${result.detail ? ` — ${result.detail}` : ''}\n`,
-  );
-  if (result.status === 'failed') setCliExitVerdict(1);
-}

src/commands/migrations/v0_32_2.ts

@@ -39,7 +39,6 @@ import type { BrainEngine } from '../../core/engine.ts';
 import { loadConfig, toEngineConfig } from '../../core/config.ts';
 import { createEngine } from '../../core/engine-factory.ts';
 import { upsertFactRow, parseFactsFence } from '../../core/facts-fence.ts';
-import { resolvePageFilePath } from '../../core/markdown.ts';
 
 let testEngineOverride: BrainEngine | null = null;
 export function __setTestEngineOverride(engine: BrainEngine | null): void {
@@ -149,16 +148,9 @@ function isLocalPathDirty(localPath: string): boolean {
   }
 }
 
-/**
- * Exported (not just via `__testing`) because `gbrain facts fence-backfill`
- * (#1867) re-runs this phase on demand: remote `extract_facts` deposits that
- * predate the fence-write backstop leave row_num-NULL rows the cycle guard
- * refuses to reconcile past. The phase is idempotent (only touches
- * `row_num IS NULL` rows), so re-running is always safe.
- */
-export async function phaseBFenceFacts(
+async function phaseBFenceFacts(
   engine: BrainEngine | null,
-  opts: Pick<OrchestratorOpts, 'dryRun'>,
+  opts: OrchestratorOpts,
 ): Promise<OrchestratorPhaseResult> {
   if (opts.dryRun) {
     // Dry-run: report what WOULD happen without touching FS or DB.
@@ -246,11 +238,7 @@ export async function phaseBFenceFacts(
     for (const [key, group] of groups) {
       const [sourceId, entitySlug] = key.split('\0');
       const localPath = localPathById.get(sourceId)!;
-      // resolvePageFilePath, NOT a bare join — non-default sources fence
-      // into `<local_path>/.sources/<id>/<slug>.md`, the same path the
-      // fence-write backstop and put_page write-through compute. A bare
-      // join here diverges fence and DB for non-default sources (#2044).
-      const filePath = resolvePageFilePath(localPath, entitySlug, sourceId);
+      const filePath = join(localPath, `${entitySlug}.md`);
       const tmpPath = `${filePath}.tmp`;
 
       try {
@@ -281,21 +269,6 @@ export async function phaseBFenceFacts(
         const existingFence = parseFactsFence(body);
         const existingKeySet = new Set(existingFence.facts.map(f => `${f.claim}\0${f.source ?? ''}`));
 
-        // Seed appended row_nums from MAX(fence max, DB max) — same #2044
-        // divergence guard as writeFactsToFence. When the fence at the
-        // resolved path is missing/behind but the DB already holds stamped
-        // rows for this slug (legacy wrong-path fence writes), fence-max+1
-        // collides with idx_facts_fence_key on the post-rename UPDATE,
-        // failing the page and leaving fence and DB disagreeing.
-        const dbMaxRows = await engine.executeRaw<{ max: number | string | null }>(
-          `SELECT MAX(row_num) AS max FROM facts
-            WHERE source_id = $1 AND source_markdown_slug = $2`,
-          [sourceId, entitySlug],
-        );
-        const dbMaxRowNum = Number(dbMaxRows[0]?.max ?? 0) || 0;
-        const fenceMaxRowNum = existingFence.facts.reduce((m, f) => Math.max(m, f.rowNum), 0);
-        let nextRowNum = Math.max(fenceMaxRowNum, dbMaxRowNum) + 1;
-
         const assignments: Array<{ id: string; row_num: number }> = [];
         for (const row of group) {
           const key = `${row.fact}\0${row.source ?? ''}`;
@@ -318,7 +291,6 @@ export async function phaseBFenceFacts(
                 .toISOString().slice(0, 10)
             : undefined;
           const { body: updated, rowNum } = upsertFactRow(body, {
-            rowNum:     nextRowNum++,
             claim:      row.fact,
             kind:       row.kind,
             confidence: row.confidence,
@@ -409,7 +381,7 @@ async function phaseCVerify(
     for (const g of groups) {
       const localPath = localPathById.get(g.source_id);
       if (!localPath) continue;
-      const filePath = resolvePageFilePath(localPath, g.source_markdown_slug, g.source_id);
+      const filePath = join(localPath, `${g.source_markdown_slug}.md`);
       if (!existsSync(filePath)) {
         mismatches.push(`${g.source_markdown_slug} (file missing)`);
         continue;

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/extract-facts.ts

@@ -176,11 +176,9 @@ export async function runExtractFacts(
   if (legacyCount > 0) {
     result.guardTriggered = true;
     result.warnings.push(
-      `extract_facts: ${legacyCount} legacy fact rows pending fence backfill ` +
-      `(row_num IS NULL — v0.31 rows or remote extract_facts deposits that ` +
-      `predate the fence backstop). Run \`gbrain facts fence-backfill\` ` +
-      `(idempotent, re-runnable) before this phase can safely reconcile ` +
-      `fence → DB.`,
+      `extract_facts: ${legacyCount} legacy v0.31 fact rows pending fence backfill. ` +
+      `Run \`gbrain apply-migrations --yes\` to complete v0_32_2 before this phase ` +
+      `can safely reconcile fence → DB.`,
     );
     return result;
   }

src/core/engine.ts

@@ -1739,14 +1739,13 @@ export interface BrainEngine {
    * single-row supersede flow because fence reconciliation is the canonical
    * source-of-truth direction, not the consolidator path.
    *
-   * Insertion runs in a single transaction. A collision on the v51
-   * partial UNIQUE index `(source_id, source_markdown_slug, row_num)`
-   * skips ONLY that row (ON CONFLICT DO NOTHING, #2044) — the rest of
-   * the batch still commits, so a redundant deposit against an
-   * already-indexed fence row is idempotent instead of a hard failure.
+   * Insertion is atomic per call: all rows commit in a single transaction
+   * or none commit (the transaction rolls back on any constraint
+   * violation, e.g. the v51 partial UNIQUE index on
+   * `(source_id, source_markdown_slug, row_num)`).
    *
-   * Returns the inserted ids in input-order (colliding rows omitted) so
-   * callers can correlate fence-row → DB-id without a separate lookup.
+   * Returns the inserted ids in input-order so callers can correlate
+   * fence-row → DB-id without a separate lookup.
    */
   insertFacts(
     rows: Array<NewFact & { row_num: number; source_markdown_slug: string }>,

src/core/facts/fence-write.ts

@@ -218,27 +218,11 @@ export async function writeFactsToFence(
       }
 
       // 2. Upsert each fact onto the fence in input order. row_num
-      //    monotonically increases, append-only, seeded from the MAX of
-      //    the fence and the DB index (#2044): when fence and DB have
-      //    diverged (e.g. legacy writes that stamped DB rows against a
-      //    fence at a path this code no longer reads), fence-max+1 can
-      //    collide with an existing DB row_num, tripping
-      //    idx_facts_fence_key and rolling back the whole batch.
-      const dbMaxRows = await engine.executeRaw<{ max: number | string | null }>(
-        `SELECT MAX(row_num) AS max FROM facts
-          WHERE source_id = $1 AND source_markdown_slug = $2`,
-        [target.sourceId, target.slug],
-      );
-      const dbMaxRowNum = Number(dbMaxRows[0]?.max ?? 0) || 0;
-      const fenceMaxRowNum = parseFactsFence(body).facts
-        .reduce((m, f) => Math.max(m, f.rowNum), 0);
-      let nextRowNum = Math.max(fenceMaxRowNum, dbMaxRowNum) + 1;
-
+      //    monotonically increases (max-existing + 1 per call, append-only).
       const assignedRowNums: number[] = [];
       for (const f of facts) {
         const validFromStr = (f.validFrom ?? new Date()).toISOString().slice(0, 10);
         const { body: updated, rowNum } = upsertFactRow(body, {
-          rowNum:      nextRowNum++,
           claim:       f.fact,
           kind:        (f.kind ?? 'fact') as 'fact' | 'event' | 'preference' | 'commitment' | 'belief',
           confidence:  f.confidence ?? 1.0,

src/core/pglite-engine.ts

@@ -4102,13 +4102,11 @@ export class PGLiteEngine implements BrainEngine {
   ): Promise<{ inserted: number; ids: number[] }> {
     if (rows.length === 0) return { inserted: 0, ids: [] };
 
-    // Single transaction; per-row INSERTs (not multi-row VALUES) keep the
-    // embedding-vs-no-embedding branching readable; batch sizes are small
-    // (5-30 rows per page in practice) so the loop overhead is negligible
-    // vs the embedding compute cost. #2044: ON CONFLICT DO NOTHING on the
-    // v51 partial UNIQUE index makes a residual fence/DB row_num collision
-    // skip that row instead of rolling back the whole batch (parity with
-    // postgres-engine.ts).
+    // Single transaction so the v51 partial UNIQUE index can roll back the
+    // whole batch on constraint violation. Per-row INSERTs (not multi-row
+    // VALUES) keep the embedding-vs-no-embedding branching readable; batch
+    // sizes are small (5-30 rows per page in practice) so the loop overhead
+    // is negligible vs the embedding compute cost.
     const ids = await this.db.transaction(async (tx) => {
       const out: number[] = [];
       for (const input of rows) {
@@ -4151,11 +4149,7 @@ export class PGLiteEngine implements BrainEngine {
                  $14, $15,
                  $16, $17, $18, $19,
                  $20
-               )
-               ON CONFLICT (source_id, source_markdown_slug, row_num)
-                 WHERE row_num IS NOT NULL
-                 DO NOTHING
-               RETURNING id`
+               ) RETURNING id`
             : `INSERT INTO facts (
                  source_id, entity_slug, fact, kind, visibility, notability, context,
                  valid_from, valid_until, source, source_session, confidence,
@@ -4169,16 +4163,12 @@ export class PGLiteEngine implements BrainEngine {
                  $15, $16,
                  $17, $18, $19, $20,
                  $21
-               )
-               ON CONFLICT (source_id, source_markdown_slug, row_num)
-                 WHERE row_num IS NOT NULL
-                 DO NOTHING
-               RETURNING id`,
+               ) RETURNING id`,
           embedStr === null
             ? [ctx.source_id, entitySlug, input.fact, kind, visibility, notability, context, validFrom, validUntil, input.source, sourceSession, confidence, embeddedAt, input.row_num, input.source_markdown_slug, claimMetric, claimValue, claimUnit, claimPeriod, eventType]
             : [ctx.source_id, entitySlug, input.fact, kind, visibility, notability, context, validFrom, validUntil, input.source, sourceSession, confidence, embedStr, embeddedAt, input.row_num, input.source_markdown_slug, claimMetric, claimValue, claimUnit, claimPeriod, eventType],
         );
-        if (ins.rows[0]) out.push(ins.rows[0].id);
+        out.push(ins.rows[0].id);
       }
       return out;
     });

src/core/postgres-engine.ts

@@ -4302,12 +4302,10 @@ export class PostgresEngine implements BrainEngine {
     // ONCE per process so the cast matches the actual column type
     // (halfvec vs vector). The probe is cached after first call.
     const castSuffix = await this.resolveFactsEmbeddingCast();
-    // Single transaction; per-row INSERTs (not multi-row VALUES) keep the
-    // embedding-vs-no-embedding branching readable; batch sizes are small
-    // (5-30 rows per page in practice). #2044: ON CONFLICT DO NOTHING on
-    // the v51 partial UNIQUE index makes a residual fence/DB row_num
-    // collision skip that row instead of rolling back the whole batch —
-    // the fence stays system-of-record and reconciliation catches up.
+    // Single transaction so the v51 partial UNIQUE index can roll back
+    // the whole batch on constraint violation. Per-row INSERTs (not
+    // multi-row VALUES) keep the embedding-vs-no-embedding branching
+    // readable; batch sizes are small (5-30 rows per page in practice).
     // No supersede flow in this path — fence reconciliation is the
     // canonical source-of-truth direction, not the consolidator path.
     const ids = await sql.begin(async (tx) => {
@@ -4348,13 +4346,9 @@ export class PostgresEngine implements BrainEngine {
             ${input.row_num}, ${input.source_markdown_slug},
             ${claimMetric}, ${claimValue}, ${claimUnit}, ${claimPeriod},
             ${eventType}
-          )
-          ON CONFLICT (source_id, source_markdown_slug, row_num)
-            WHERE row_num IS NOT NULL
-            DO NOTHING
-          RETURNING id
+          ) RETURNING id
         `;
-        if (ins[0]) out.push(Number(ins[0].id));
+        out.push(Number(ins[0].id));
       }
       return out;
     });

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/extract-facts-phase.test.ts

@@ -304,9 +304,7 @@ describe('runExtractFacts — empty-fence guard (Codex R2-#7)', () => {
     expect(r.legacyRowsPending).toBe(1);
     expect(r.factsInserted).toBe(0);
     expect(r.factsDeleted).toBe(0);
-    // #1867: the remedy hint points at the re-runnable backfill command,
-    // not the one-shot v0_32_2 migration (which the ledger never re-runs).
-    expect(r.warnings.some(w => w.includes('gbrain facts fence-backfill'))).toBe(true);
+    expect(r.warnings.some(w => w.includes('apply-migrations'))).toBe(true);
 
     // Legacy row was NOT touched.
     // eslint-disable-next-line @typescript-eslint/no-explicit-any

test/facts-fence-backfill-command.test.ts

@@ -1,133 +0,0 @@
-/**
- * #1867 — `gbrain facts fence-backfill` command tests.
- *
- * The command re-runs the (idempotent) v0_32_2 phase B on demand so
- * row_num-NULL backlogs — remote extract_facts deposits that predate
- * the fence-write backstop — can be cleared without re-running the
- * one-shot migration. Real PGLite + real tempdir filesystem.
- */
-
-import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:test';
-import { mkdtempSync, rmSync, existsSync, readFileSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { runFactsCommand } from '../src/commands/facts.ts';
-import { phaseBFenceFacts } from '../src/commands/migrations/v0_32_2.ts';
-
-let engine: PGLiteEngine;
-let brainDir: string;
-
-beforeAll(async () => {
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-});
-
-afterAll(async () => {
-  await engine.disconnect();
-  try {
-    if (brainDir) rmSync(brainDir, { recursive: true, force: true });
-  } catch { /* best-effort */ }
-});
-
-beforeEach(async () => {
-  brainDir = mkdtempSync(join(tmpdir(), 'facts-backfill-cmd-test-'));
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query('DELETE FROM facts');
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query(
-    `UPDATE sources SET local_path = $1 WHERE id = 'default'`,
-    [brainDir],
-  );
-});
-
-async function seedLegacyFact(fact: string): Promise<void> {
-  // The row_num-NULL shape a remote extract_facts deposit leaves behind
-  // when it lands via the legacy DB-only path.
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  await (engine as any).db.query(
-    `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                        valid_from, source, confidence)
-     VALUES ('default', 'people/alice', $1, 'fact', 'private', 'medium',
-             now(), 'mcp:extract_facts', 1.0)`,
-    [fact],
-  );
-}
-
-describe('gbrain facts fence-backfill', () => {
-  test('fences row_num-NULL rows and stamps the DB', async () => {
-    await seedLegacyFact('Deposited remotely');
-
-    await runFactsCommand(engine, ['fence-backfill']);
-
-    // The fence exists on disk with the claim.
-    const filePath = join(brainDir, 'people/alice.md');
-    expect(existsSync(filePath)).toBe(true);
-    expect(readFileSync(filePath, 'utf-8')).toContain('Deposited remotely');
-
-    // The backlog is cleared: no row_num-NULL rows remain.
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num, source_markdown_slug FROM facts',
-    );
-    expect(rows.rows).toHaveLength(1);
-    expect(rows.rows[0].row_num).toBe(1);
-    expect(rows.rows[0].source_markdown_slug).toBe('people/alice');
-  });
-
-  test('re-run is a no-op (idempotent)', async () => {
-    await seedLegacyFact('Deposited remotely');
-    await runFactsCommand(engine, ['fence-backfill']);
-    await runFactsCommand(engine, ['fence-backfill']);
-
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query('SELECT id FROM facts');
-    expect(rows.rows).toHaveLength(1);
-    const body = readFileSync(join(brainDir, 'people/alice.md'), 'utf-8');
-    expect(body.match(/Deposited remotely/g)).toHaveLength(1);
-  });
-
-  test('--dry-run reports without writing', async () => {
-    await seedLegacyFact('Deposited remotely');
-    await runFactsCommand(engine, ['fence-backfill', '--dry-run']);
-
-    expect(existsSync(join(brainDir, 'people/alice.md'))).toBe(false);
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num FROM facts',
-    );
-    expect(rows.rows[0].row_num).toBeNull();
-  });
-
-  test('diverged page: appends past the DB row_num max instead of colliding (#2044 class)', async () => {
-    // The #2044 divergence shape the backfill must survive: the DB already
-    // holds stamped rows 1..3 for the slug (legacy wrong-path fence write),
-    // but the fence at the resolved path is missing. Fence-max+1 (= 1) would
-    // collide with the stamped rows on the post-rename UPDATE, failing the
-    // page and leaving the renamed fence disagreeing with the DB.
-    for (let n = 1; n <= 3; n++) {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      await (engine as any).db.query(
-        `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                            valid_from, source, confidence, row_num, source_markdown_slug)
-         VALUES ('default', 'people/alice', $1, 'fact', 'private', 'medium',
-                 now(), 'mcp:extract_facts', 1.0, $2, 'people/alice')`,
-        [`stamped ${n}`, n],
-      );
-    }
-    await seedLegacyFact('Deposited remotely');
-
-    const result = await phaseBFenceFacts(engine, { dryRun: false });
-    expect(result.status).toBe('complete');
-
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      `SELECT row_num FROM facts WHERE fact = 'Deposited remotely'`,
-    );
-    expect(rows.rows[0].row_num).toBe(4);
-    const body = readFileSync(join(brainDir, 'people/alice.md'), 'utf-8');
-    expect(body).toContain('| 4 | Deposited remotely |');
-  });
-});

test/fence-write.test.ts

@@ -292,49 +292,6 @@ describe('lookupSourceLocalPath', () => {
   });
 });
 
-describe('writeFactsToFence — fence/DB divergence (#2044)', () => {
-  test('seeds row_num past the DB max when the fence lags the DB', async () => {
-    // Simulate the divergence class from #2044: DB rows were stamped with
-    // row_nums against a fence written at a path this code no longer reads
-    // (e.g. the pre-"Local patch 2026-06-11" wrong-path writes). The page
-    // on disk has NO fence, but the DB already holds row_num 1..3 for the
-    // slug. Pre-fix, the next deposit re-assigned row_num=1 from fence
-    // text alone and the whole insertFacts batch failed on
-    // idx_facts_fence_key.
-    for (let n = 1; n <= 3; n++) {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      await (engine as any).db.query(
-        `INSERT INTO facts (source_id, entity_slug, fact, kind, visibility, notability,
-                            valid_from, source, confidence, row_num, source_markdown_slug)
-         VALUES ('default', 'people/dana', $1, 'fact', 'private', 'medium',
-                 now(), 'mcp:extract_facts', 1.0, $2, 'people/dana')`,
-        [`old claim ${n}`, n],
-      );
-    }
-
-    const result = await writeFactsToFence(
-      engine,
-      { sourceId: 'default', localPath: brainDir, slug: 'people/dana' },
-      [baseInput({ fact: 'second deposit' })],
-    );
-
-    expect(result.fenceWriteFailed).toBeUndefined();
-    expect(result.inserted).toBe(1);
-
-    // The new row landed PAST the DB max, not at fence-max+1 (= 1).
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    const rows = await (engine as any).db.query(
-      'SELECT row_num FROM facts WHERE id = $1',
-      [result.ids[0]],
-    );
-    expect(rows.rows[0].row_num).toBe(4);
-
-    // And the on-disk fence carries the same row_num — fence and DB agree.
-    const body = readFileSync(join(brainDir, 'people/dana.md'), 'utf-8');
-    expect(body).toContain('| 4 | second deposit |');
-  });
-});
-
 // Cleanup any leftover tempdirs after the whole suite.
 afterAll(() => {
   // No-op: each test cleaned up via the beforeEach; this is a safety net.

test/insert-facts-batch.test.ts

@@ -5,7 +5,7 @@
  *   - Batch insert N rows persists row_num + source_markdown_slug
  *   - Empty batch is a no-op
  *   - Returns ids in input-order
- *   - v51 partial UNIQUE collision skips only the colliding row (#2044)
+ *   - v51 partial UNIQUE index rolls back the whole batch on a collision
  *   - deleteFactsForPage scopes by (source_id, source_markdown_slug);
  *     never touches other pages or pre-v51 NULL-source_markdown_slug rows
  *   - deleteFactsForPage on an empty page returns deleted:0 (idempotent)
@@ -135,29 +135,30 @@ describe('engine.insertFacts — batch insert', () => {
     });
   });
 
-  test('v51 partial UNIQUE collision skips ONLY the colliding row (#2044)', async () => {
+  test('v51 partial UNIQUE index rolls back the whole batch on collision', async () => {
     // Seed row #1 first.
     await engine.insertFacts([fixtureFact(1, { fact: 'seeded' })], { source_id: 'default' });
 
-    // Batch-insert rows that include a colliding row_num=1. Pre-#2044 this
-    // threw and rolled back the whole batch, making a second remote
-    // extract_facts deposit to an already-fenced page a hard failure. Now
-    // ON CONFLICT DO NOTHING skips the colliding row and keeps the rest.
-    const result = await engine.insertFacts(
-      [
-        fixtureFact(2, { fact: 'second' }),
-        fixtureFact(1, { fact: 'collides' }),  // row_num=1 on same (source_id, source_markdown_slug)
-        fixtureFact(3, { fact: 'third' }),
-      ],
-      { source_id: 'default' },
-    );
-    expect(result.inserted).toBe(2);
-    expect(result.ids).toHaveLength(2);
+    // Now try to batch-insert rows that include a colliding row_num=1.
+    let threw = false;
+    try {
+      await engine.insertFacts(
+        [
+          fixtureFact(2, { fact: 'second' }),
+          fixtureFact(1, { fact: 'collides' }),  // row_num=1 on same (source_id, source_markdown_slug)
+          fixtureFact(3, { fact: 'third' }),
+        ],
+        { source_id: 'default' },
+      );
+    } catch {
+      threw = true;
+    }
+    expect(threw).toBe(true);
 
-    // The seeded row survives untouched; the colliding claim is skipped.
+    // Verify the transaction rolled back — only the seeded row should remain.
     // eslint-disable-next-line @typescript-eslint/no-explicit-any
     const rows = await (engine as any).db.query('SELECT fact FROM facts ORDER BY id');
-    expect(rows.rows.map((r: { fact: string }) => r.fact)).toEqual(['seeded', 'second', 'third']);
+    expect(rows.rows.map((r: { fact: string }) => r.fact)).toEqual(['seeded']);
   });
 
   test('different source_markdown_slug values DO NOT collide on the same row_num', async () => {