reland: fix(embed): stamp gateway-resolved model in content_chunks.model, not compiled default (#2846) (#3343)

* fix(embed): stamp gateway-resolved model in content_chunks.model, not compiled default (#2846)

upsertChunks fell back to the compile-time DEFAULT_EMBEDDING_MODEL
('zeroentropyai:zembed-1') when a ChunkInput carried no explicit `model`.
The embed pipeline (src/commands/embed.ts) builds ChunkInputs without a
`model` field, so rows whose vectors were produced by the config-resolved
model (e.g. openai:text-embedding-3-large) were mislabeled with the
hardcoded default — corrupting the provenance that signature-drift
staleness and dimension-migration logic depend on.

Both engines now resolve the gateway's runtime embedding model once per
upsert and use it as the fallback, mirroring the existing resolve-then-
default pattern used for schema sizing. Regression test added (pglite);
verified via negative control that it fails against the old fallback.

This is a write-path change (upsertChunks), not a search-path change, so
retrieval eval replay is not applicable.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* test: Lane A.7 pins gateway-resolved chunk model, not compiled default

#2846 changed upsertChunks' fallback from DEFAULT_EMBEDDING_MODEL to the
gateway-resolved runtime model. Lane A.7 still pinned the old fallback,
and the test preload (test/helpers/legacy-embedding-preload.ts) pins the
gateway to openai:text-embedding-3-large for every test process — so the
original #2846 landing failed this test deterministically and got batch-
reverted. The test now asserts the resolved model (the intended #2846
semantics) while keeping the CDX2-4 bare-literal regression guard.

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

---------

Co-authored-by: SailorJoe6 <SailorJoe6@Gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Garry Tan <garrytan@gmail.com>

src/core/pglite-engine.ts

@@ -2312,6 +2312,18 @@ export class PGLiteEngine implements BrainEngine {
     const params: unknown[] = [];
     let paramIdx = 1;
 
+    // Provenance fallback for chunks without an explicit `model`: resolve the
+    // gateway's runtime model, not the compile-time DEFAULT_EMBEDDING_MODEL.
+    // See postgres-engine.ts _upsertChunksOnce for the full rationale — pglite
+    // mirrors it for parity.
+    let resolvedModel: string = DEFAULT_EMBEDDING_MODEL;
+    try {
+      const gw = await import('./ai/gateway.ts');
+      resolvedModel = gw.getEmbeddingModel() || resolvedModel;
+    } catch {
+      // Gateway unconfigured (unit tests / pre-connect): keep the default.
+    }
+
     for (const chunk of chunks) {
       const embeddingStr = chunk.embedding
         ? '[' + Array.from(chunk.embedding).join(',') + ']'
@@ -2344,7 +2356,7 @@ export class PGLiteEngine implements BrainEngine {
       if (embeddingImageStr) params.push(embeddingImageStr);
       params.push(
         pageId, chunk.chunk_index, chunk.chunk_text, chunk.chunk_source,
-        chunk.model || DEFAULT_EMBEDDING_MODEL, chunk.token_count || null,
+        chunk.model || resolvedModel, chunk.token_count || null,
         chunk.language || null, chunk.symbol_name || null, chunk.symbol_type || null,
         chunk.start_line ?? null, chunk.end_line ?? null,
         parentPath, chunk.doc_comment || null, chunk.symbol_name_qualified || null,

src/core/postgres-engine.ts

@@ -2438,6 +2438,23 @@ export class PostgresEngine implements BrainEngine {
     const params: unknown[] = [];
     let paramIdx = 1;
 
+    // Provenance fallback for chunks that don't carry an explicit `model`:
+    // resolve the model the gateway ACTUALLY uses at runtime, not the
+    // compile-time DEFAULT_EMBEDDING_MODEL constant. Callers like `embed`
+    // build ChunkInputs without a `model` field (src/commands/embed.ts), so
+    // the old `chunk.model || DEFAULT_EMBEDDING_MODEL` fallback stamped the
+    // hardcoded default (e.g. zeroentropyai:zembed-1) onto rows whose vectors
+    // were produced by a different, config-resolved model — corrupting the
+    // provenance that signature-drift staleness + dim-migration logic trust.
+    // Mirrors the resolve-then-fallback pattern used for schema sizing above.
+    let resolvedModel: string = DEFAULT_EMBEDDING_MODEL;
+    try {
+      const gw = await import('./ai/gateway.ts');
+      resolvedModel = gw.getEmbeddingModel() || resolvedModel;
+    } catch {
+      // Gateway unconfigured (unit tests / pre-connect): keep the default.
+    }
+
     for (const chunk of chunks) {
       const embeddingStr = chunk.embedding
         ? '[' + Array.from(chunk.embedding).join(',') + ']'
@@ -2467,7 +2484,7 @@ export class PostgresEngine implements BrainEngine {
       if (embeddingImageStr) params.push(embeddingImageStr);
       params.push(
         pageId, chunk.chunk_index, chunk.chunk_text, chunk.chunk_source,
-        chunk.model || DEFAULT_EMBEDDING_MODEL, chunk.token_count || null,
+        chunk.model || resolvedModel, chunk.token_count || null,
         chunk.language || null, chunk.symbol_name || null, chunk.symbol_type || null,
         chunk.start_line ?? null, chunk.end_line ?? null,
         parentPath, chunk.doc_comment || null, chunk.symbol_name_qualified || null,

test/e2e/embedding-column-pglite.test.ts

@@ -216,6 +216,44 @@ describe('hybridSearch + resolver — unknown column at entry (D11)', () => {
   });
 });
 
+describe('upsertChunks — model provenance uses gateway-resolved model, not compiled default', () => {
+  // Regression (zbrain-rfi): when a caller builds ChunkInputs without an
+  // explicit `model` (as src/commands/embed.ts does), the engine used to
+  // stamp the compile-time DEFAULT_EMBEDDING_MODEL ('zeroentropyai:zembed-1')
+  // onto content_chunks.model — even though the vector was produced by the
+  // config-resolved model. That corrupted provenance the signature-drift +
+  // dim-migration logic trusts. The engine must fall back to the model the
+  // gateway ACTUALLY resolves at write time.
+  test('unspecified chunk.model records the resolved model, not zeroentropyai:zembed-1', async () => {
+    configureGateway({
+      embedding_model: 'openai:text-embedding-3-large',
+      embedding_dimensions: 1536,
+      env: { OPENAI_API_KEY: 'sk-test' },
+    });
+
+    await engine.putPage('docs/provenance-page', {
+      type: 'concept',
+      title: 'Provenance test page',
+      compiled_truth: 'Chunk whose model column must reflect the resolved model.',
+    });
+    // No `model` field on the input — the write-side fallback must fill it.
+    await engine.upsertChunks('docs/provenance-page', [
+      { chunk_index: 0, chunk_text: 'provenance chunk', chunk_source: 'compiled_truth' },
+    ]);
+
+    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 = 'docs/provenance-page'`,
+    );
+    expect(rows.length).toBe(1);
+    expect(rows[0].model).toBe('openai:text-embedding-3-large');
+    expect(rows[0].model).not.toBe('zeroentropyai:zembed-1');
+
+    resetGateway();
+  });
+});
+
 describe('buildVectorCastFragment — engine SQL composer (D3)', () => {
   test('vector descriptor emits $1::vector', () => {
     const r: ResolvedColumn = {

test/v0_37_gap_fill.serial.test.ts

@@ -30,11 +30,15 @@ import { configureGateway, resetGateway, __setEmbedTransportForTests } from '../
 import { withEnv } from './helpers/with-env.ts';
 
 // ─────────────────────────────────────────────────────────────────────
-// Lane A.7 — Chunk-row INSERT model default tracks defaults.ts constant
-// (not stale OpenAI literal). Pre-fix `chunk.model || 'text-embedding-3-large'`
-// in both engines; post-fix `chunk.model || DEFAULT_EMBEDDING_MODEL`.
+// Lane A.7 — Chunk-row INSERT model default tracks the gateway-resolved
+// model (not a stale OpenAI literal, not the compile-time constant).
+// Pre-fix `chunk.model || 'text-embedding-3-large'` in both engines;
+// v0.37 fix `chunk.model || DEFAULT_EMBEDDING_MODEL`; #2846 tightened it
+// to the gateway's runtime model so provenance matches the vector's
+// actual producer (falls back to DEFAULT_EMBEDDING_MODEL only when the
+// gateway is unconfigured).
 // ─────────────────────────────────────────────────────────────────────
-describe('Lane A.7 — chunk-row INSERT default tracks ai/defaults.ts constant', () => {
+describe('Lane A.7 — chunk-row INSERT default tracks the gateway-resolved model', () => {
   let engine: PGLiteEngine;
 
   beforeAll(async () => {
@@ -47,8 +51,11 @@ describe('Lane A.7 — chunk-row INSERT default tracks ai/defaults.ts constant',
     await engine.disconnect();
   });
 
-  test('upsertChunks without explicit model: row stores DEFAULT_EMBEDDING_MODEL', async () => {
-    const { DEFAULT_EMBEDDING_MODEL } = await import('../src/core/ai/defaults.ts');
+  test('upsertChunks without explicit model: row stores the gateway-resolved model', async () => {
+    // The test preload (test/helpers/legacy-embedding-preload.ts) pins the
+    // gateway to 'openai:text-embedding-3-large', so that's what the write
+    // site must stamp — NOT the compile-time DEFAULT_EMBEDDING_MODEL (#2846).
+    const { getEmbeddingModel } = await import('../src/core/ai/gateway.ts');
     await engine.putPage('test/a7', { type: 'note', title: 'A.7', compiled_truth: 'hello' });
     await engine.upsertChunks('test/a7', [
       { chunk_index: 0, chunk_text: 'hello', chunk_source: 'compiled_truth' },
@@ -57,9 +64,9 @@ describe('Lane A.7 — chunk-row INSERT default tracks ai/defaults.ts constant',
     const rows = await engine.executeRaw<{ model: string }>(
       `SELECT model FROM content_chunks WHERE chunk_index = 0 LIMIT 1`,
     );
-    expect(rows[0]?.model).toBe(DEFAULT_EMBEDDING_MODEL);
+    expect(rows[0]?.model).toBe(getEmbeddingModel());
     // CDX2-4 regression: would have been 'text-embedding-3-large'
-    // (a literal pre-fix; production write site that was never tested).
+    // (a bare literal pre-fix; provider-prefixed form is required).
     expect(rows[0]?.model).not.toBe('text-embedding-3-large');
   });
 });