fix(dream): stamp incremental extraction watermark (#2636)

The Dream cycle disables sync's inline extraction and routes changed
slugs through extractForSlugs, which flushed link/timeline batches but
never stamped links_extracted_at — so incrementally extracted pages
stayed permanently visible to `extract --stale` / doctor.

Collect processedRefs per successfully processed page and stamp them
via stampExtracted (best-effort) after both batch flushes, non-dry-run
mode 'all' only. Source-id threading from the original PR #2637 already
landed on master via #1503/#1747, so this rebase carries only the
missing watermark stamp plus regression tests.

Takeover of #2637.

Co-authored-by: JavanC <JavanC@users.noreply.github.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/commands/embed.ts

@@ -1,5 +1,5 @@
 import type { BrainEngine } from '../core/engine.ts';
-import { embedBatch, currentEmbeddingSignature, resolveEmbeddingModelLabel } from '../core/embedding.ts';
+import { embedBatch, currentEmbeddingSignature } 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';
@@ -581,16 +581,11 @@ 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),
   }));
 
@@ -722,16 +717,12 @@ 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));
@@ -1021,15 +1012,11 @@ 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/extract.ts

@@ -1025,6 +1025,10 @@ async function extractForSlugs(
   let linksCreated = 0;
   let timelineCreated = 0;
   let pagesProcessed = 0;
+  // #2636: successfully processed pages get their extraction watermark
+  // stamped after the final flush (mode 'all' only — a partial-mode run
+  // hasn't done the full extraction the watermark asserts).
+  const processedRefs: Array<{ slug: string; source_id: string }> = [];
 
   // Issue #972: read the basename flag once per extract run.
   const globalBasename = await isGlobalBasenameEnabled(engine);
@@ -1113,6 +1117,7 @@ async function extractForSlugs(
         }
 
         pagesProcessed++;
+        if (!dryRun) processedRefs.push({ slug, source_id: sourceId ?? 'default' });
       } catch { /* skip unreadable */ }
       progress.tick(1);
     },
@@ -1120,6 +1125,13 @@ async function extractForSlugs(
 
   await flushLinks();
   await flushTimeline();
+  // #2636: the Dream cycle disables sync's inline extraction and routes
+  // changed slugs through this incremental path — without a stamp here,
+  // those pages never get links_extracted_at and stay permanently visible
+  // to `extract --stale` / doctor. Stamp only after BOTH batches flushed.
+  if (!dryRun && mode === 'all') {
+    await stampExtracted(engine, processedRefs);
+  }
   progress.finish();
 
   if (!jsonMode) {

src/core/embed-stale.ts

@@ -20,7 +20,6 @@
 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';
 
@@ -201,17 +200,11 @@ 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

@@ -113,21 +113,6 @@ 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, resolveEmbeddingModelLabel } from './embedding.ts';
+import { embedBatch, embedMultimodal, currentEmbeddingSignature } from './embedding.ts';
 import { slugifyPath, slugifyCodePath, isCodeFilePath } from './sync.ts';
 import type { ChunkInput, PageInput, PageType } from './types.ts';
 import { computeEffectiveDate } from './effective-date.ts';
@@ -716,12 +716,8 @@ 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);
@@ -1145,10 +1141,7 @@ 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);
@@ -1160,12 +1153,9 @@ 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/embed-stale.serial.test.ts

@@ -15,7 +15,6 @@ 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;
@@ -277,49 +276,4 @@ 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

@@ -37,8 +37,6 @@ 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.
@@ -805,34 +803,3 @@ 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-incremental.test.ts

@@ -60,6 +60,51 @@ async function seedPage(slug: string, body: string): Promise<void> {
 }
 
 describe('runExtractCore — incremental cycle path (#417)', () => {
+  test('Dream incremental all-mode stamps the source-scoped extraction watermark (#2636)', async () => {
+    await engine.executeRaw(
+      `INSERT INTO sources (id, name, local_path) VALUES ($1, $2, $3)`,
+      ['repo-a', 'repo-a', tempDir],
+    );
+    await engine.putPage('people/alice-example', {
+      type: 'person',
+      title: 'alice-example',
+      compiled_truth: '# alice',
+      timeline: '',
+      frontmatter: {},
+      content_hash: 'h',
+    }, { sourceId: 'repo-a' });
+    writeFileSync(join(tempDir, 'people/alice-example.md'), '# alice');
+
+    await runExtractCore(engine as unknown as BrainEngine, {
+      mode: 'all',
+      dir: tempDir,
+      slugs: ['people/alice-example'],
+      sourceId: 'repo-a',
+    });
+
+    const rows = await engine.executeRaw<{ links_extracted_at: string | null }>(
+      `SELECT links_extracted_at FROM pages WHERE slug = $1 AND source_id = $2`,
+      ['people/alice-example', 'repo-a'],
+    );
+    expect(rows[0]?.links_extracted_at).not.toBeNull();
+    expect(await engine.countStalePagesForExtraction({ sourceId: 'repo-a' })).toBe(0);
+  });
+
+  test('Dream incremental dry-run does NOT stamp the watermark', async () => {
+    await seedPage('people/alice-example', '# alice');
+    await runExtractCore(engine as unknown as BrainEngine, {
+      mode: 'all',
+      dir: tempDir,
+      slugs: ['people/alice-example'],
+      dryRun: true,
+    });
+    const rows = await engine.executeRaw<{ links_extracted_at: string | null }>(
+      `SELECT links_extracted_at FROM pages WHERE slug = $1`,
+      ['people/alice-example'],
+    );
+    expect(rows[0]?.links_extracted_at ?? null).toBeNull();
+  });
+
   test('1. slugs: [] returns immediately with zero counts (early-return path)', async () => {
     await seedPage('people/alice-example', '# alice');
     const result = await runExtractCore(engine as unknown as BrainEngine, {

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

@@ -73,21 +73,4 @@ 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');
-  });
 });