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/commands/autopilot.ts

@@ -151,100 +151,6 @@ export function shouldSpawnAutopilotWorker(args: string[]): boolean {
   return !args.includes('--no-worker');
 }
 
-/**
- * #1525 — positional subcommand translation.
- *
- * Pre-fix, `gbrain autopilot status` silently fell through to "start daemon"
- * because `runAutopilot()` only branched on flag forms (`--status`, etc.).
- * `status` was treated as a stray positional and ignored.
- *
- * This translator maps known positional subcommands to their flag form so
- * `autopilot status` is equivalent to `autopilot --status`, then rejects
- * any unrecognized positional with a fail-loud error before any side
- * effect (lockfile, daemon spawn, sync dispatch) runs.
- *
- * Scope decisions:
- *   - Known aliases: `status` → `--status`, `install` → `--install`,
- *     `uninstall` → `--uninstall`, `start` → (drop; default daemon launch).
- *   - `stop` is intentionally NOT aliased here. Stopping a running daemon
- *     is a new behavior (read PID from lock, SIGTERM, drain) that deserves
- *     its own design and PR. Users typing `gbrain autopilot stop` today get
- *     the unknown-positional error with the canonical alternatives.
- *   - At most one positional allowed; multiple positionals fail loud.
- */
-// Every flag that consumes the NEXT argv token. Missing one here makes the
-// translator misread the flag's value as a positional subcommand and exit 2
-// (e.g. `--install --target linux-cron`). Keep in sync with parseArg call sites.
-const AUTOPILOT_VALUE_FLAGS = new Set(['--repo', '--interval', '--target']);
-const AUTOPILOT_POSITIONAL_ALIASES: Record<string, string | null> = {
-  status: '--status',
-  install: '--install',
-  uninstall: '--uninstall',
-  start: null, // drop the positional; default behavior is daemon launch
-};
-
-export type PositionalTranslation =
-  | { ok: true; args: string[] }
-  | {
-      ok: false;
-      reason: 'unknown_subcommand' | 'multiple_subcommands';
-      message: string;
-    };
-
-export function translatePositionalSubcommands(args: string[]): PositionalTranslation {
-  const out: string[] = [];
-  let positionalSeen = false;
-  let i = 0;
-  while (i < args.length) {
-    const a = args[i];
-    if (AUTOPILOT_VALUE_FLAGS.has(a)) {
-      // Pass through the flag and its value untouched. If the value is
-      // missing at end-of-argv, fall through so the existing parseArg
-      // path can report the broken usage.
-      out.push(a);
-      if (i + 1 < args.length) {
-        out.push(args[i + 1]);
-        i += 2;
-      } else {
-        i += 1;
-      }
-      continue;
-    }
-    if (a.startsWith('-')) {
-      out.push(a);
-      i += 1;
-      continue;
-    }
-    // Positional subcommand.
-    if (positionalSeen) {
-      const known = Object.keys(AUTOPILOT_POSITIONAL_ALIASES).join(', ');
-      return {
-        ok: false,
-        reason: 'multiple_subcommands',
-        message: `Multiple subcommands given. Use only one of: ${known}.`,
-      };
-    }
-    positionalSeen = true;
-    if (a in AUTOPILOT_POSITIONAL_ALIASES) {
-      const alias = AUTOPILOT_POSITIONAL_ALIASES[a];
-      if (alias) out.push(alias);
-      i += 1;
-      continue;
-    }
-    const known = Object.keys(AUTOPILOT_POSITIONAL_ALIASES).join(', ');
-    return {
-      ok: false,
-      reason: 'unknown_subcommand',
-      message:
-        `Unknown subcommand: \`${a}\`.\n` +
-        `Allowed subcommands: ${known}.\n` +
-        `Or use the flag form: --status, --install, --uninstall.\n` +
-        `Run \`gbrain autopilot --help\` for full usage.`,
-    };
-  }
-  return { ok: true, args: out };
-}
-
 export function isPidAlive(pid: number): boolean {
   if (!Number.isFinite(pid) || pid <= 0) return false;
   try {
@@ -457,11 +363,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
       '       gbrain autopilot --install [--repo <path>]\n' +
       '       gbrain autopilot --uninstall\n' +
       '       gbrain autopilot --status [--json]\n\n' +
-      'Subcommand aliases:\n' +
-      '       gbrain autopilot status     → --status\n' +
-      '       gbrain autopilot install    → --install\n' +
-      '       gbrain autopilot uninstall  → --uninstall\n' +
-      '       gbrain autopilot start      → (default daemon launch)\n\n' +
       'Self-maintaining brain daemon. Runs the full maintenance cycle\n' +
       '(lint + backlinks + sync + extract + embed + orphans) on an interval.\n\n' +
       'For a one-shot cron-triggered cycle, see `gbrain dream`.',
@@ -469,16 +370,6 @@ export async function runAutopilot(engine: BrainEngine, args: string[]) {
     return;
   }
 
-  // #1525: translate positional subcommands to their flag form BEFORE any
-  // side effect (lockfile, daemon spawn, sync dispatch). Unknown positionals
-  // fail loud here rather than silently starting the daemon.
-  const translated = translatePositionalSubcommands(args);
-  if (!translated.ok) {
-    console.error(translated.message);
-    process.exit(2);
-  }
-  args = translated.args;
-
   if (args.includes('--install')) {
     await installDaemon(engine, args);
     return;

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/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/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) {

test/autopilot-positional-subcommands.test.ts

@@ -1,197 +0,0 @@
-/**
- * Tests for translatePositionalSubcommands() — the v0.41.x #1525 fix that
- * prevents `gbrain autopilot status` from silently starting the daemon.
- *
- * IRON RULE regression guard: the exact ticket repro (`gbrain autopilot
- * status`) MUST translate to `--status`, not fall through to the default
- * daemon launch. Verified by the "ticket-exact repro" case below.
- */
-
-import { describe, test, expect } from 'bun:test';
-import { translatePositionalSubcommands } from '../src/commands/autopilot.ts';
-
-describe('translatePositionalSubcommands — known aliases', () => {
-  test('IRON RULE — `autopilot status` translates to `--status` (ticket #1525 repro)', () => {
-    const r = translatePositionalSubcommands(['status']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--status']);
-  });
-
-  test('`install` translates to `--install`', () => {
-    const r = translatePositionalSubcommands(['install']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--install']);
-  });
-
-  test('`uninstall` translates to `--uninstall`', () => {
-    const r = translatePositionalSubcommands(['uninstall']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--uninstall']);
-  });
-
-  test('`start` drops the positional (default daemon launch)', () => {
-    const r = translatePositionalSubcommands(['start']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual([]);
-  });
-
-  test('`start --json` drops only the positional, keeps the flag', () => {
-    const r = translatePositionalSubcommands(['start', '--json']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--json']);
-  });
-});
-
-describe('translatePositionalSubcommands — flag/positional interleaving', () => {
-  test('`status --json` preserves the trailing flag', () => {
-    const r = translatePositionalSubcommands(['status', '--json']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--status', '--json']);
-  });
-
-  test('`--json status` preserves the leading flag', () => {
-    const r = translatePositionalSubcommands(['--json', 'status']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--json', '--status']);
-  });
-
-  test('`--repo /foo status` does not mis-classify the path as positional', () => {
-    const r = translatePositionalSubcommands(['--repo', '/foo', 'status']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--repo', '/foo', '--status']);
-  });
-
-  test('`--interval 300 install` does not mis-classify the number as positional', () => {
-    const r = translatePositionalSubcommands(['--interval', '300', 'install']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--interval', '300', '--install']);
-  });
-
-  test('`--install --target linux-cron` does not mis-classify the target as positional', () => {
-    // --target is installDaemon's value flag; its value must never be read
-    // as a positional subcommand (regression guard for the review fix).
-    const r = translatePositionalSubcommands(['--install', '--target', 'linux-cron']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--install', '--target', 'linux-cron']);
-  });
-
-  test('`install --target macos` keeps the alias translation and the target value', () => {
-    const r = translatePositionalSubcommands(['install', '--target', 'macos']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--install', '--target', 'macos']);
-  });
-
-  test('value-flag at end of argv with missing value passes through (so parseArg can report it)', () => {
-    const r = translatePositionalSubcommands(['--repo']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--repo']);
-  });
-
-  test('value-flag whose value looks like an alias is NOT translated', () => {
-    // `--repo status` means "use repo path 'status'", not "show status".
-    // Translator must not destructure the value of --repo.
-    const r = translatePositionalSubcommands(['--repo', 'status']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--repo', 'status']);
-  });
-});
-
-describe('translatePositionalSubcommands — pass-through cases', () => {
-  test('empty args returns empty args', () => {
-    const r = translatePositionalSubcommands([]);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual([]);
-  });
-
-  test('flag-only invocation passes through unchanged', () => {
-    const r = translatePositionalSubcommands(['--status', '--json']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['--status', '--json']);
-  });
-
-  test('short flag `-h` passes through unchanged', () => {
-    const r = translatePositionalSubcommands(['-h']);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(['-h']);
-  });
-
-  test('all known bare flags pass through unchanged', () => {
-    const flags = ['--help', '--install', '--uninstall', '--status', '--json', '--inline', '--no-worker'];
-    const r = translatePositionalSubcommands(flags);
-    expect(r.ok).toBe(true);
-    if (r.ok) expect(r.args).toEqual(flags);
-  });
-});
-
-describe('translatePositionalSubcommands — rejection of unknown positionals', () => {
-  test('unknown positional `foo` fails with reason=unknown_subcommand + structured message', () => {
-    const r = translatePositionalSubcommands(['foo']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('unknown_subcommand');
-      expect(r.message).toContain('Unknown subcommand: `foo`');
-      expect(r.message).toContain('status');
-      expect(r.message).toContain('install');
-      expect(r.message).toContain('uninstall');
-      expect(r.message).toContain('--help');
-    }
-  });
-
-  test('unknown positional `stop` fails with reason=unknown_subcommand (NOT silently aliased)', () => {
-    // Stop is mentioned in the ticket but deliberately NOT aliased in this
-    // PR — stopping a running daemon is a new behavior, not just an alias.
-    // Until that feature lands separately, `stop` must fail loud rather
-    // than starting the daemon (the bug we're fixing).
-    const r = translatePositionalSubcommands(['stop']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('unknown_subcommand');
-      expect(r.message).toContain('Unknown subcommand: `stop`');
-    }
-  });
-
-  test('unknown positional `status-detail` (close-but-not-matching) fails', () => {
-    const r = translatePositionalSubcommands(['status-detail']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('unknown_subcommand');
-      expect(r.message).toContain('Unknown subcommand: `status-detail`');
-    }
-  });
-
-  test('multiple positionals fail with reason=multiple_subcommands (`start install`)', () => {
-    const r = translatePositionalSubcommands(['start', 'install']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('multiple_subcommands');
-      expect(r.message).toContain('Multiple subcommands');
-    }
-  });
-
-  test('multiple positionals fail even when both are known aliases (`status install`)', () => {
-    const r = translatePositionalSubcommands(['status', 'install']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('multiple_subcommands');
-      expect(r.message).toContain('Multiple subcommands');
-    }
-  });
-
-  test('known-then-unknown rejects with multiple_subcommands (first-positional-wins)', () => {
-    // First positional is known, second is not. Rejection comes from the
-    // multiple-positional rule, which fires before the unknown check; the
-    // intent is "only one subcommand allowed."
-    const r = translatePositionalSubcommands(['status', 'garbage']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) expect(r.reason).toBe('multiple_subcommands');
-  });
-
-  test('unknown-then-known rejects on the unknown (unknown fires before second-positional check)', () => {
-    const r = translatePositionalSubcommands(['garbage', 'status']);
-    expect(r.ok).toBe(false);
-    if (!r.ok) {
-      expect(r.reason).toBe('unknown_subcommand');
-      expect(r.message).toContain('garbage');
-    }
-  });
-});

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/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');
+  });
 });