fix(embed): extend #1717 model labeling to the embed-backfill stale path

src/core/embed-stale.ts (used by the embed-backfill minion handler) built
its merged ChunkInput without a model field, so every chunk on a touched
page — re-embedded AND preserved — was relabeled to the engine default on
each backfill pass. Mirror the embed.ts semantics: stamp the resolved
gateway label on re-embedded chunks, carry the existing label on untouched
ones. Pinned by a PGLite test that fails without the fix.

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/embed.ts

@@ -1,5 +1,5 @@
 import type { BrainEngine } from '../core/engine.ts';
-import { embedBatch, currentEmbeddingSignature } from '../core/embedding.ts';
+import { embedBatch, currentEmbeddingSignature, resolveEmbeddingModelLabel } from '../core/embedding.ts';
 import type { ChunkInput } from '../core/types.ts';
 import { chunkText } from '../core/chunkers/recursive.ts';
 import { createProgress, type ProgressReporter } from '../core/progress.ts';
@@ -581,11 +581,16 @@ async function embedPage(
   for (let j = 0; j < toEmbed.length; j++) {
     embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
   }
+  // #1717: label each (re)embedded chunk with the model that actually
+  // produced its vector. Preserved chunks (not re-embedded this pass) keep
+  // their existing model so a mixed-model page isn't relabeled wholesale.
+  const embedModelLabel = resolveEmbeddingModelLabel();
   const updated: ChunkInput[] = chunks.map(c => ({
     chunk_index: c.chunk_index,
     chunk_text: c.chunk_text,
     chunk_source: c.chunk_source,
     embedding: embeddingMap.get(c.chunk_index),
+    model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
     token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
   }));
 
@@ -717,12 +722,16 @@ async function embedAll(
       for (let j = 0; j < toEmbed.length; j++) {
         embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
       }
+      // #1717: stamp the resolved embedding model on (re)embedded chunks;
+      // preserve the existing model on chunks left untouched.
+      const embedModelLabel = resolveEmbeddingModelLabel();
       // Preserve ALL chunks, only update embeddings for stale ones
       const updated: ChunkInput[] = chunks.map(c => ({
         chunk_index: c.chunk_index,
         chunk_text: c.chunk_text,
         chunk_source: c.chunk_source,
         embedding: embeddingMap.get(c.chunk_index) ?? undefined,
+        model: embeddingMap.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
         token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
       }));
       await observed(pacer, () => engine.upsertChunks(page.slug, updated, pageOpts));
@@ -1012,11 +1021,15 @@ async function embedAllStale(
           for (let j = 0; j < stale.length; j++) {
             staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
           }
+          // #1717: label the re-embedded (stale) chunks with the resolved
+          // model; preserve the existing model on the non-stale chunks.
+          const embedModelLabel = resolveEmbeddingModelLabel();
           const merged: ChunkInput[] = existing.map(c => ({
             chunk_index: c.chunk_index,
             chunk_text: c.chunk_text,
             chunk_source: c.chunk_source,
             embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
+            model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
             token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
           }));
           await observed(pacer, () => engine.upsertChunks(slug, merged, { sourceId: keySourceId }));

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/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/embed-stale.ts

@@ -20,6 +20,7 @@
 import type { BrainEngine } from './engine.ts';
 import type { ChunkInput } from './types.ts';
 import { embedBatchWithBackoff } from '../commands/embed.ts';
+import { resolveEmbeddingModelLabel } from './embedding.ts';
 import { type DbPacer, createNoopPacer, observed } from './db-pacer.ts';
 import { AbortError } from './abort-check.ts';
 
@@ -200,11 +201,17 @@ export async function embedStaleForSource(
         for (let j = 0; j < stale.length; j++) {
           staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
         }
+        // #1717: label re-embedded chunks with the model that produced the
+        // vector; preserved chunks keep their existing model. Without this,
+        // upsertChunks falls back to DEFAULT_EMBEDDING_MODEL for every chunk
+        // (the same mislabel the embed.ts paths fixed).
+        const embedModelLabel = resolveEmbeddingModelLabel();
         const merged: ChunkInput[] = existing.map((c) => ({
           chunk_index: c.chunk_index,
           chunk_text: c.chunk_text,
           chunk_source: c.chunk_source,
           embedding: staleIdxToEmbedding.get(c.chunk_index) ?? undefined,
+          model: staleIdxToEmbedding.has(c.chunk_index) && embedModelLabel ? embedModelLabel : c.model,
           token_count: c.token_count || Math.ceil(c.chunk_text.length / 4),
           // Carry through per-chunk metadata. upsertChunks writes these as
           // EXCLUDED.<col> (not COALESCE), so omitting them here resets image

src/core/embedding.ts

@@ -113,6 +113,21 @@ export async function embedBatch(
   return results;
 }
 
+/**
+ * Resolve the embedding model label (`provider:model`) to stamp onto
+ * `content_chunks.model`, so each chunk records the model that actually
+ * produced its vector instead of the engine's hardcoded default (#1717).
+ * Returns undefined if the gateway is unconfigured; callers then fall back
+ * to the chunk's existing model rather than mislabeling it.
+ */
+export function resolveEmbeddingModelLabel(): string | undefined {
+  try {
+    return gatewayGetModel();
+  } catch {
+    return undefined;
+  }
+}
+
 /** Currently-configured embedding model (short form without provider prefix). */
 export function getEmbeddingModelName(): string {
   return gatewayGetModel().split(':').slice(1).join(':') || 'text-embedding-3-large';

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/import-file.ts

@@ -8,7 +8,7 @@ import { chunkText } from './chunkers/recursive.ts';
 import { chunkCodeText, chunkCodeTextFull, detectCodeLanguage, CHUNKER_VERSION } from './chunkers/code.ts';
 import { findChunkForOffset } from './chunkers/edge-extractor.ts';
 import { extractCodeRefs, imageOfCandidates } from './link-extraction.ts';
-import { embedBatch, embedMultimodal, currentEmbeddingSignature } from './embedding.ts';
+import { embedBatch, embedMultimodal, currentEmbeddingSignature, resolveEmbeddingModelLabel } from './embedding.ts';
 import { slugifyPath, slugifyCodePath, isCodeFilePath } from './sync.ts';
 import type { ChunkInput, PageInput, PageType } from './types.ts';
 import { computeEffectiveDate } from './effective-date.ts';
@@ -716,8 +716,12 @@ export async function importFromContent(
       ? chunks.map((c) => wrapChunkForEmbedding(c.chunk_text, prefix, c.chunk_source))
       : chunks.map((c) => c.chunk_text);
     const embeddings = await embedBatch(wrappedTexts);
+    // #1717: label each chunk with the model that actually produced its
+    // vector, not the engine's hardcoded default.
+    const embedModelLabel = resolveEmbeddingModelLabel();
     for (let i = 0; i < chunks.length; i++) {
       chunks[i].embedding = embeddings[i];
+      if (embedModelLabel) chunks[i].model = embedModelLabel;
       // token_count tracks the wrapped string length so cost reporting
       // reflects what we actually sent to the embedder.
       chunks[i].token_count = Math.ceil(wrappedTexts[i].length / 4);
@@ -1141,7 +1145,10 @@ export async function importCodeFile(
     const matched = existingByKey.get(key);
     if (matched && matched.embedding) {
       // Reuse the existing embedding verbatim. No API call, no cost.
+      // #1717: carry the existing model label along with the reused vector
+      // so the upsert doesn't relabel it with the engine default.
       chunks[i]!.embedding = matched.embedding as Float32Array;
+      chunks[i]!.model = matched.model ?? undefined;
       chunks[i]!.token_count = matched.token_count ?? undefined;
     } else {
       needsEmbedIndexes.push(i);
@@ -1153,9 +1160,12 @@ export async function importCodeFile(
     try {
       const textsToEmbed = needsEmbedIndexes.map((i) => chunks[i]!.chunk_text);
       const embeddings = await embedBatch(textsToEmbed);
+      // #1717: stamp the model that produced these vectors.
+      const embedModelLabel = resolveEmbeddingModelLabel();
       for (let j = 0; j < needsEmbedIndexes.length; j++) {
         const i = needsEmbedIndexes[j]!;
         chunks[i]!.embedding = embeddings[j]!;
+        if (embedModelLabel) chunks[i]!.model = embedModelLabel;
         chunks[i]!.token_count = Math.ceil(chunks[i]!.chunk_text.length / 4);
       }
     } catch (e: unknown) {

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/embed-stale.serial.test.ts

@@ -15,6 +15,7 @@ import { describe, test, expect, beforeAll, afterAll, beforeEach } from 'bun:tes
 import { PGLiteEngine } from '../src/core/pglite-engine.ts';
 import { resetPgliteState } from './helpers/reset-pglite.ts';
 import { embedStaleForSource } from '../src/core/embed-stale.ts';
+import { configureGateway, resetGateway } from '../src/core/ai/gateway.ts';
 import type { ChunkInput } from '../src/core/types.ts';
 
 let engine: PGLiteEngine;
@@ -276,4 +277,49 @@ describe('embedStaleForSource', () => {
     // The stale text row actually got its embedding.
     expect(txtRow.embedded_at).not.toBeNull();
   });
+
+  // #1717: the backfill path must label re-embedded chunks with the model
+  // that produced the vector, and preserve the existing label on chunks it
+  // did not touch (before the fix, both were reset to the engine default).
+  test('labels re-embedded chunks with the gateway model, preserves untouched labels (#1717)', async () => {
+    configureGateway({
+      embedding_model: 'openai:text-embedding-3-large',
+      env: { OPENAI_API_KEY: 'sk-test-embed-stale-1717' },
+    });
+    try {
+      await engine.putPage('notes/model-label', {
+        type: 'note',
+        title: 'model-label',
+        compiled_truth: '# model-label\n\nseeded',
+      });
+      await engine.upsertChunks('notes/model-label', [
+        {
+          chunk_index: 0,
+          chunk_text: 'already embedded elsewhere',
+          chunk_source: 'compiled_truth',
+          embedding: new Float32Array(1536).fill(0.01),
+          model: 'voyage:voyage-3',
+          token_count: 4,
+        },
+        {
+          chunk_index: 1,
+          chunk_text: 'stale chunk needing embed',
+          chunk_source: 'compiled_truth',
+          token_count: 5,
+          embedding: undefined, // stale
+        },
+      ]);
+
+      const result = await embedStaleForSource(engine, 'default', { embedFn: fakeEmbedFn });
+      expect(result.embedded).toBe(1);
+
+      const after = await engine.getChunks('notes/model-label');
+      const preserved = after.find((c) => c.chunk_index === 0)!;
+      const reembedded = after.find((c) => c.chunk_index === 1)!;
+      expect(reembedded.model).toBe('openai:text-embedding-3-large');
+      expect(preserved.model).toBe('voyage:voyage-3');
+    } finally {
+      resetGateway();
+    }
+  });
 });

test/embed.serial.test.ts

@@ -37,6 +37,8 @@ mock.module('../src/core/embedding.ts', () => ({
   // setPageEmbeddingSignature / invalidateStaleSignatureEmbeddings resolve to
   // null via the Proxy default, so the signature value is inert here.
   currentEmbeddingSignature: () => 'test:model:1536',
+  // #1717: embed paths stamp this label on (re)embedded chunks.
+  resolveEmbeddingModelLabel: () => 'openai:text-embedding-3-large',
 }));
 
 // Import AFTER mocking.
@@ -803,3 +805,34 @@ describe('embedAllStale --source threading (D7)', () => {
     expect((firstCallOpts as { sourceId?: string }).sourceId).toBe('media-corpus');
   });
 });
+
+// #1717: content_chunks.model must record the model that actually produced
+// each vector, not the gateway/engine default.
+describe('content_chunks.model labeling (#1717)', () => {
+  test('stamps the resolved embedding model on re-embedded chunks, preserves it on untouched chunks', async () => {
+    let upserted: any[] | undefined;
+    // Chunk 0 is stale (no embedded_at) → gets re-embedded this pass.
+    // Chunk 1 is already embedded with a DIFFERENT model → must be preserved,
+    // not relabeled to the current model.
+    const chunks = [
+      { chunk_index: 0, chunk_text: 'a', chunk_source: 'compiled_truth', embedded_at: null, model: 'zeroentropyai:zembed-1', token_count: 1 },
+      { chunk_index: 1, chunk_text: 'b', chunk_source: 'compiled_truth', embedded_at: '2026-01-01', embedding: new Float32Array(1536), model: 'voyage:voyage-3', token_count: 1 },
+    ];
+    const engine = mockEngine({
+      getPage: async () => ({ slug: 'notes/x', compiled_truth: 'a', timeline: '', source_id: 'default' }),
+      getChunks: async () => chunks,
+      upsertChunks: async (_slug: string, c: any[]) => { upserted = c; },
+      setPageEmbeddingSignature: async () => null,
+    });
+
+    await runEmbedCore(engine, { slugs: ['notes/x'] });
+
+    expect(upserted).toBeDefined();
+    const byIdx = Object.fromEntries(upserted!.map(c => [c.chunk_index, c]));
+    // Re-embedded chunk carries the model that produced its vector (was
+    // mislabeled with the default before the fix).
+    expect(byIdx[0].model).toBe('openai:text-embedding-3-large');
+    // Untouched chunk keeps its original model — no wholesale relabel.
+    expect(byIdx[1].model).toBe('voyage:voyage-3');
+  });
+});

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/import-signature-stamp.serial.test.ts

@@ -73,4 +73,21 @@ describe('importFromContent embedding_signature stamping (F1)', () => {
     await importFromContent(engine, 'concepts/unstamped', '# Unstamped\n\nbody content.', { noEmbed: true });
     expect(await signatureOf('concepts/unstamped')).toBeNull();
   });
+
+  // #1717: content_chunks.model must record the model that produced the
+  // vector (the configured gateway model), not the engine's hardcoded
+  // default. The gateway here is configured to openai:text-embedding-3-large,
+  // which differs from DEFAULT_EMBEDDING_MODEL — so this fails without the
+  // import-path model stamping.
+  test('inline embed labels content_chunks.model with the configured model (#1717)', async () => {
+    await importFromContent(engine, 'concepts/labeled', '# Labeled\n\nsome body content to chunk and embed.', {});
+    const rows = await engine.executeRaw<{ model: string }>(
+      `SELECT cc.model FROM content_chunks cc
+        JOIN pages p ON p.id = cc.page_id
+       WHERE p.slug = $1 AND p.source_id = 'default'`,
+      ['concepts/labeled'],
+    );
+    expect(rows.length).toBeGreaterThan(0);
+    for (const r of rows) expect(r.model).toBe('openai:text-embedding-3-large');
+  });
 });

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 () => {