fix(facts): seed phase B backfill row_nums past the DB max (#2044 class)

phaseBFenceFacts appended at fence-max+1 with no DB seeding, so on a
diverged page (stamped DB rows + missing fence at the resolved path —
the same divergence class the writeFactsToFence fix handles) the
post-rename UPDATE hit idx_facts_fence_key, the page failed, and the
already-renamed fence disagreed with the DB while the row_num-NULL
backlog persisted. Seed appended row_nums from MAX(fence max, DB max)
per (source_id, slug), mirroring the fence-write fix. Repro test added.

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', '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', 'facts', '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.
@@ -466,11 +466,6 @@ async function main() {
     const result = JSON.parse(JSON.stringify(rawResult, bigintToStringReplacer));
     const output = formatResult(op.name, result);
     if (output) process.stdout.write(output);
-    // #1484 — invisible-miss hint: a bare query/search that hit zero results
-    // on a multi-source brain tells the user (stderr) which source it
-    // actually searched and how to widen the scope.
-    const hint = await sourceScopeHint(op.name, params, ctx.sourceId, engine, result);
-    if (hint) console.error(hint);
   } catch (e: unknown) {
     // v0.42.20.0 (codex D4): on error, set exitCode + return so the `finally`
     // STILL runs (drains every background-work sink + disconnects). A bare
@@ -842,44 +837,6 @@ async function makeContext(engine: BrainEngine, params: Record<string, unknown>)
   };
 }
 
-/**
- * #1484 — a bare `gbrain query`/`search` silently scopes to the resolved
- * source (usually 'default'); on a multi-source brain a zero-hit run looks
- * identical to "the brain doesn't know this" even when the answer lives in
- * another source. Returns a stderr hint when (a) the op is query/search,
- * (b) it returned zero results, (c) the caller did NOT scope explicitly
- * (--source / --source-id / --all-sources), and (d) the brain has >1
- * registered source. Best-effort: any lookup failure returns null.
- *
- * Exported for tests (same import-safety contract as formatResult).
- */
-export async function sourceScopeHint(
-  opName: string,
-  params: Record<string, unknown>,
-  sourceId: string,
-  engine: BrainEngine,
-  result: unknown,
-): Promise<string | null> {
-  if (opName !== 'query' && opName !== 'search') return null;
-  if (!Array.isArray(result) || result.length > 0) return null;
-  // Explicit scoping (flag tier) = user intent; don't second-guess it.
-  if (params.source || params.source_id || params.all_sources) return null;
-  if (sourceId === '__all__') return null;
-  try {
-    const rows = await engine.executeRaw<{ n: number }>(
-      `SELECT count(*)::int AS n FROM sources`,
-    );
-    const n = Number(rows[0]?.n ?? 0);
-    if (n <= 1) return null;
-    return (
-      `Hint: this brain has ${n} sources; you searched only "${sourceId}". ` +
-      `Retry with --source-id __all__ (all sources) or --source-id <id>.`
-    );
-  } catch {
-    return null; // hint is best-effort; never fail the query over it
-  }
-}
-
 // Exported for tests (same import-safety contract as cliAliases/printOpHelp).
 export function formatResult(opName: string, result: unknown): string {
   switch (opName) {
@@ -1034,6 +991,8 @@ 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)
@@ -1069,6 +1028,7 @@ 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.)',
@@ -1898,6 +1858,13 @@ 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
@@ -2378,6 +2345,7 @@ 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

@@ -0,0 +1,48 @@
+/**
+ * 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

@@ -39,6 +39,7 @@ 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 {
@@ -148,9 +149,16 @@ function isLocalPathDirty(localPath: string): boolean {
   }
 }
 
-async function phaseBFenceFacts(
+/**
+ * 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(
   engine: BrainEngine | null,
-  opts: OrchestratorOpts,
+  opts: Pick<OrchestratorOpts, 'dryRun'>,
 ): Promise<OrchestratorPhaseResult> {
   if (opts.dryRun) {
     // Dry-run: report what WOULD happen without touching FS or DB.
@@ -238,7 +246,11 @@ async function phaseBFenceFacts(
     for (const [key, group] of groups) {
       const [sourceId, entitySlug] = key.split('\0');
       const localPath = localPathById.get(sourceId)!;
-      const filePath = join(localPath, `${entitySlug}.md`);
+      // resolvePageFilePath, NOT a bare join — non-default sources fence
+      // into `<local_path>/.sources/<id>/<slug>.md`, the same path the
+      // fence-write backstop and put_page write-through compute. A bare
+      // join here diverges fence and DB for non-default sources (#2044).
+      const filePath = resolvePageFilePath(localPath, entitySlug, sourceId);
       const tmpPath = `${filePath}.tmp`;
 
       try {
@@ -269,6 +281,21 @@ 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${row.source ?? ''}`;
@@ -291,6 +318,7 @@ async function phaseBFenceFacts(
                 .toISOString().slice(0, 10)
             : undefined;
           const { body: updated, rowNum } = upsertFactRow(body, {
+            rowNum:     nextRowNum++,
             claim:      row.fact,
             kind:       row.kind,
             confidence: row.confidence,
@@ -381,7 +409,7 @@ async function phaseCVerify(
     for (const g of groups) {
       const localPath = localPathById.get(g.source_id);
       if (!localPath) continue;
-      const filePath = join(localPath, `${g.source_markdown_slug}.md`);
+      const filePath = resolvePageFilePath(localPath, g.source_markdown_slug, g.source_id);
       if (!existsSync(filePath)) {
         mismatches.push(`${g.source_markdown_slug} (file missing)`);
         continue;

src/core/ai/gateway.ts

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

src/core/cycle/extract-facts.ts

@@ -176,9 +176,11 @@ export async function runExtractFacts(
   if (legacyCount > 0) {
     result.guardTriggered = true;
     result.warnings.push(
-      `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.`,
+      `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.`,
     );
     return result;
   }

src/core/engine.ts

@@ -1739,13 +1739,14 @@ export interface BrainEngine {
    * single-row supersede flow because fence reconciliation is the canonical
    * source-of-truth direction, not the consolidator path.
    *
-   * 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)`).
+   * 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.
    *
-   * Returns the inserted ids in input-order so callers can correlate
-   * fence-row → DB-id without a separate lookup.
+   * Returns the inserted ids in input-order (colliding rows omitted) 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

@@ -218,11 +218,27 @@ export async function writeFactsToFence(
       }
 
       // 2. Upsert each fact onto the fence in input order. row_num
-      //    monotonically increases (max-existing + 1 per call, append-only).
+      //    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;
+
       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/operations.ts

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

src/core/pglite-engine.ts

@@ -4102,11 +4102,13 @@ export class PGLiteEngine implements BrainEngine {
   ): Promise<{ inserted: number; ids: number[] }> {
     if (rows.length === 0) return { inserted: 0, ids: [] };
 
-    // 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.
+    // 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).
     const ids = await this.db.transaction(async (tx) => {
       const out: number[] = [];
       for (const input of rows) {
@@ -4149,7 +4151,11 @@ export class PGLiteEngine implements BrainEngine {
                  $14, $15,
                  $16, $17, $18, $19,
                  $20
-               ) RETURNING id`
+               )
+               ON CONFLICT (source_id, source_markdown_slug, row_num)
+                 WHERE row_num IS NOT NULL
+                 DO NOTHING
+               RETURNING id`
             : `INSERT INTO facts (
                  source_id, entity_slug, fact, kind, visibility, notability, context,
                  valid_from, valid_until, source, source_session, confidence,
@@ -4163,12 +4169,16 @@ export class PGLiteEngine implements BrainEngine {
                  $15, $16,
                  $17, $18, $19, $20,
                  $21
-               ) RETURNING id`,
+               )
+               ON CONFLICT (source_id, source_markdown_slug, row_num)
+                 WHERE row_num IS NOT NULL
+                 DO NOTHING
+               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],
         );
-        out.push(ins.rows[0].id);
+        if (ins.rows[0]) out.push(ins.rows[0].id);
       }
       return out;
     });

src/core/postgres-engine.ts

@@ -4302,10 +4302,12 @@ 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 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).
+    // 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.
     // 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) => {
@@ -4346,9 +4348,13 @@ export class PostgresEngine implements BrainEngine {
             ${input.row_num}, ${input.source_markdown_slug},
             ${claimMetric}, ${claimValue}, ${claimUnit}, ${claimPeriod},
             ${eventType}
-          ) RETURNING id
+          )
+          ON CONFLICT (source_id, source_markdown_slug, row_num)
+            WHERE row_num IS NOT NULL
+            DO NOTHING
+          RETURNING id
         `;
-        out.push(Number(ins[0].id));
+        if (ins[0]) out.push(Number(ins[0].id));
       }
       return out;
     });

src/core/search/hybrid.ts

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

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

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

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

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

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

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

test/extract-facts-phase.test.ts

@@ -304,7 +304,9 @@ describe('runExtractFacts — empty-fence guard (Codex R2-#7)', () => {
     expect(r.legacyRowsPending).toBe(1);
     expect(r.factsInserted).toBe(0);
     expect(r.factsDeleted).toBe(0);
-    expect(r.warnings.some(w => w.includes('apply-migrations'))).toBe(true);
+    // #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);
 
     // Legacy row was NOT touched.
     // eslint-disable-next-line @typescript-eslint/no-explicit-any

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

@@ -0,0 +1,133 @@
+/**
+ * #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

@@ -292,6 +292,49 @@ 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/hybrid-vector-arm-warn.serial.test.ts

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

test/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 index rolls back the whole batch on a collision
+ *   - v51 partial UNIQUE collision skips only the colliding row (#2044)
  *   - 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)
@@ -135,30 +135,29 @@ describe('engine.insertFacts — batch insert', () => {
     });
   });
 
-  test('v51 partial UNIQUE index rolls back the whole batch on collision', async () => {
+  test('v51 partial UNIQUE collision skips ONLY the colliding row (#2044)', async () => {
     // Seed row #1 first.
     await engine.insertFacts([fixtureFact(1, { fact: 'seeded' })], { source_id: 'default' });
 
-    // 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);
+    // 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);
 
-    // Verify the transaction rolled back — only the seeded row should remain.
+    // The seeded row survives untouched; the colliding claim is skipped.
     // 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']);
+    expect(rows.rows.map((r: { fact: string }) => r.fact)).toEqual(['seeded', 'second', 'third']);
   });
 
   test('different source_markdown_slug values DO NOT collide on the same row_num', async () => {

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

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

test/voyage-response-cap.test.ts

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