Merge branch 'master' into takeover/pr-3015-orphan-policy

src/commands/embed.ts

@@ -581,7 +581,7 @@ async function embedPage(
   for (let j = 0; j < toEmbed.length; j++) {
     embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
   }
-  const updated: ChunkInput[] = chunks.map(c => ({
+  const updated: ChunkInput[] = chunks.map(c => preserveCodeMetadata(c, {
     chunk_index: c.chunk_index,
     chunk_text: c.chunk_text,
     chunk_source: c.chunk_source,
@@ -605,6 +605,31 @@ async function embedPage(
   slog(`${slug}: embedded ${toEmbed.length} chunks`);
 }
 
+/**
+ * Carry code-chunk metadata (language, symbol_name, symbol_type, line range,
+ * parent scope, doc comment, qualified name) from a loaded Chunk back into a
+ * ChunkInput destined for upsertChunks.
+ *
+ * Issue #769: every re-embed used to strip these fields, and upsertChunks
+ * overwrites (does not COALESCE) the metadata columns from EXCLUDED, so
+ * each pass clobbered code-def's primary index to NULL. Pulling the
+ * preservation into one helper keeps the three re-embed call sites
+ * (embedPage, embedAll non-stale, embedAllStale) in lock-step.
+ */
+function preserveCodeMetadata(loaded: any, base: ChunkInput): ChunkInput {
+  return {
+    ...base,
+    language: loaded.language ?? undefined,
+    symbol_name: loaded.symbol_name ?? undefined,
+    symbol_type: loaded.symbol_type ?? undefined,
+    start_line: loaded.start_line ?? undefined,
+    end_line: loaded.end_line ?? undefined,
+    parent_symbol_path: loaded.parent_symbol_path ?? undefined,
+    doc_comment: loaded.doc_comment ?? undefined,
+    symbol_name_qualified: loaded.symbol_name_qualified ?? undefined,
+  };
+}
+
 async function embedAll(
   engine: BrainEngine,
   staleOnly: boolean,
@@ -717,8 +742,10 @@ async function embedAll(
       for (let j = 0; j < toEmbed.length; j++) {
         embeddingMap.set(toEmbed[j].chunk_index, embeddings[j]);
       }
-      // Preserve ALL chunks, only update embeddings for stale ones
-      const updated: ChunkInput[] = chunks.map(c => ({
+      // Preserve ALL chunks, only update embeddings for stale ones.
+      // preserveCodeMetadata threads code-chunk metadata (#769) so re-embed
+      // doesn't clobber language/symbol_name/symbol_type to NULL.
+      const updated: ChunkInput[] = chunks.map(c => preserveCodeMetadata(c, {
         chunk_index: c.chunk_index,
         chunk_text: c.chunk_text,
         chunk_source: c.chunk_source,
@@ -1012,7 +1039,10 @@ async function embedAllStale(
           for (let j = 0; j < stale.length; j++) {
             staleIdxToEmbedding.set(stale[j].chunk_index, embeddings[j]);
           }
-          const merged: ChunkInput[] = existing.map(c => ({
+          // preserveCodeMetadata threads code-chunk metadata (#769) so the
+          // autopilot --stale path doesn't clobber language/symbol_name/etc
+          // to NULL on every cycle.
+          const merged: ChunkInput[] = existing.map(c => preserveCodeMetadata(c, {
             chunk_index: c.chunk_index,
             chunk_text: c.chunk_text,
             chunk_source: c.chunk_source,

src/core/contextual-retrieval-service.ts

@@ -54,6 +54,7 @@ import {
 import {
   generatePerChunkSynopsis,
   SYNOPSIS_PROMPT_VERSION,
+  SYNOPSIS_DOC_MAX_CHARS,
   type GeneratePerChunkSynopsisResult,
 } from './page-summary.ts';
 import {
@@ -103,8 +104,17 @@ function getEmbeddingModelTag(): string {
 export function computeCorpusGeneration(args: {
   crMode: CRMode;
   haikuModel: string;
+  /**
+   * Resolved `SYNOPSIS_DOC_MAX_CHARS` for per_chunk_synopsis runs. When
+   * present, folded into the hash so changes to
+   * `GBRAIN_SYNOPSIS_DOC_MAX_CHARS` invalidate the prior cache cleanly.
+   * Omit for `crMode !== 'per_chunk_synopsis'` — title / none modes
+   * don't consult the cap and the field stays out of the hash for
+   * back-compat with pre-cap embeddings.
+   */
+  synopsisDocMaxChars?: number;
 }): string {
-  return createHash('sha256')
+  const h = createHash('sha256')
     .update(args.crMode)
     .update('|')
     .update(String(SYNOPSIS_PROMPT_VERSION))
@@ -113,9 +123,11 @@ export function computeCorpusGeneration(args: {
     .update('|')
     .update(String(TITLE_WRAPPER_VERSION))
     .update('|')
-    .update(getEmbeddingModelTag())
-    .digest('hex')
-    .slice(0, 16);
+    .update(getEmbeddingModelTag());
+  if (args.synopsisDocMaxChars !== undefined) {
+    h.update('|doc_cap=').update(String(args.synopsisDocMaxChars));
+  }
+  return h.digest('hex').slice(0, 16);
 }
 
 /**
@@ -253,7 +265,11 @@ export async function reembedPageWithContextualRetrieval(
       args.pageSlug,
       args.sourceId,
       resolution.mode,
-      computeCorpusGeneration({ crMode: resolution.mode, haikuModel: args.haikuModel ?? DEFAULT_HAIKU_MODEL }),
+      computeCorpusGeneration({
+        crMode: resolution.mode,
+        haikuModel: args.haikuModel ?? DEFAULT_HAIKU_MODEL,
+        synopsisDocMaxChars: resolution.mode === 'per_chunk_synopsis' ? SYNOPSIS_DOC_MAX_CHARS : undefined,
+      }),
     );
     return { kind: 'skipped', reason: 'no_chunks' };
   }
@@ -282,6 +298,7 @@ export async function reembedPageWithContextualRetrieval(
       const corpus_generation = computeCorpusGeneration({
         crMode: attemptMode,
         haikuModel,
+        synopsisDocMaxChars: attemptMode === 'per_chunk_synopsis' ? SYNOPSIS_DOC_MAX_CHARS : undefined,
       });
 
       // ── PHASE 2: single DB transaction ───────────────────────────

src/core/import-file.ts

@@ -733,6 +733,11 @@ export async function importFromContent(
       : computeCorpusGeneration({
           crMode: effectiveCRMode,
           haikuModel: 'anthropic:claude-haiku-4-5-20251001',
+          // Inline import-file path never uses per_chunk_synopsis (refuses
+          // upstream); pass undefined so the doc-cap field stays out of
+          // the hash here. Per_chunk_synopsis runs through the Minion
+          // backfill handler which threads SYNOPSIS_DOC_MAX_CHARS through
+          // the service layer.
         });
 
   // Transaction wraps all DB writes. Every per-page tx call carries the

src/core/link-extraction.ts

@@ -489,7 +489,22 @@ export async function extractPageLinks(
       // text inside `[[...]]` before any `|`), NOT the display alias
       // (ref.name = match[2]). `[[struktura|the project]]` must resolve
       // `struktura`, not "the project". The display text is for context only.
-      const matches = await resolver.resolveBasenameMatches(ref.slug);
+      //
+      // The literal may be path-qualified (`[[notes/struktura]]`). The FS
+      // path (resolveSlugAll) strips the dirname before its basename lookup,
+      // but this path passed the raw literal to an index keyed by final
+      // segments only — so every slash-containing wikilink outside
+      // DIR_PATTERN silently resolved to nothing. Query by the final
+      // segment, then use the written path as a disambiguation filter
+      // (the analogue of the FS ancestor walk honoring the written path):
+      // a match must end with the literal, so `[[notes/struktura]]` can
+      // resolve to `vault/notes/struktura` but never to `wiki/struktura`.
+      const slashIdx = ref.slug.lastIndexOf('/');
+      const basename = slashIdx === -1 ? ref.slug : ref.slug.slice(slashIdx + 1);
+      let matches = await resolver.resolveBasenameMatches(basename);
+      if (slashIdx !== -1) {
+        matches = matches.filter(m => m === ref.slug || m.endsWith(`/${ref.slug}`));
+      }
       if (matches.length === 0) continue;
       const idx = content.indexOf(ref.slug);
       const context = idx >= 0 ? excerpt(content, idx, 240) : ref.name;

src/core/page-summary.ts

@@ -44,6 +44,33 @@ const HAIKU_MAX_TOKENS = 200;
 /** Default model when caller doesn't override. Resolves through the gateway. */
 const DEFAULT_SYNOPSIS_MODEL = 'anthropic:claude-haiku-4-5-20251001';
 
+/**
+ * Hard cap on `documentText` length (chars) before send.
+ *
+ * 2026-05-25 fix wave: small local chat models (Gemma 4 E2B, Qwen3 4B) get
+ * dramatically slower on long contexts even with 131K-token windows declared.
+ * A 73K-char page synopsis on Gemma 4 E2B takes 60-120s, exceeding the
+ * worker's default 30s `lockDuration` and tripping `lock-lost` errors.
+ *
+ * Truncate to a budget that fits a small model's effective throughput while
+ * preserving enough document context for the synopsis to be useful. Truncates
+ * the TAIL because the head (title, frontmatter, intro) carries the
+ * document-level anchor the synopsis needs.
+ *
+ * Override per workload via `GBRAIN_SYNOPSIS_DOC_MAX_CHARS`. Default 32768
+ * (~8K tokens at 4 chars/tok) keeps small-model synopsis under ~30s.
+ * Anthropic Haiku is unaffected at this cap; bump higher when running
+ * frontier models if you want richer document anchoring.
+ */
+export const SYNOPSIS_DOC_MAX_CHARS = (() => {
+  const env = process.env.GBRAIN_SYNOPSIS_DOC_MAX_CHARS;
+  if (env && /^\d+$/.test(env)) {
+    const n = parseInt(env, 10);
+    if (n >= 512 && n <= 1_048_576) return n;
+  }
+  return 32768;
+})();
+
 /**
  * Synopsis prompt version. Folded into corpus_generation so prompt edits
  * invalidate prior embeddings via the v0.40.3.0 query_cache.page_generations
@@ -188,11 +215,19 @@ function buildUserPrompt(
   documentText: string,
   chunkText: string,
 ): string {
+  // Tail-truncate `documentText` to `SYNOPSIS_DOC_MAX_CHARS` so small local
+  // chat models don't stall on >100KB pages. Head preserved (title block,
+  // frontmatter, intro paragraphs carry the document-level anchor).
+  let trimmedDoc = documentText;
+  if (documentText.length > SYNOPSIS_DOC_MAX_CHARS) {
+    trimmedDoc = documentText.slice(0, SYNOPSIS_DOC_MAX_CHARS) +
+      `\n\n[... ${documentText.length - SYNOPSIS_DOC_MAX_CHARS} chars truncated for synopsis budget ...]`;
+  }
   return [
     `<page_title>${pageTitle}</page_title>`,
     '',
     '<full_document>',
-    documentText,
+    trimmedDoc,
     '</full_document>',
     '',
     '<chunk>',

src/core/pglite-engine.ts

@@ -2324,6 +2324,10 @@ export class PGLiteEngine implements BrainEngine {
     // v0.40.3.0 D24 NULL→non-NULL race fix mirrors postgres-engine.ts. Two writers
     // racing on the same chunk previously raced last-write-wins; the fix lets the
     // fresher `embedded_at` win in the text-unchanged branch.
+    //
+    // Code-chunk metadata columns follow the same chunk_text-gated CASE pattern as `embedding`
+    // (#769). Re-chunk trusts EXCLUDED outright; pure re-embed COALESCEs so a caller carrying
+    // only embedding-shaped fields doesn't clobber metadata to NULL.
     await this.db.query(
       `INSERT INTO content_chunks ${cols} VALUES ${rowParts.join(', ')}
        ON CONFLICT (page_id, chunk_index) DO UPDATE SET
@@ -2347,14 +2351,14 @@ export class PGLiteEngine implements BrainEngine {
                 THEN EXCLUDED.embedded_at
            ELSE content_chunks.embedded_at
          END,
-         language = EXCLUDED.language,
-         symbol_name = EXCLUDED.symbol_name,
-         symbol_type = EXCLUDED.symbol_type,
-         start_line = EXCLUDED.start_line,
-         end_line = EXCLUDED.end_line,
-         parent_symbol_path = EXCLUDED.parent_symbol_path,
-         doc_comment = EXCLUDED.doc_comment,
-         symbol_name_qualified = EXCLUDED.symbol_name_qualified,
+         language = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.language ELSE COALESCE(EXCLUDED.language, content_chunks.language) END,
+         symbol_name = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_name ELSE COALESCE(EXCLUDED.symbol_name, content_chunks.symbol_name) END,
+         symbol_type = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_type ELSE COALESCE(EXCLUDED.symbol_type, content_chunks.symbol_type) END,
+         start_line = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.start_line ELSE COALESCE(EXCLUDED.start_line, content_chunks.start_line) END,
+         end_line = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.end_line ELSE COALESCE(EXCLUDED.end_line, content_chunks.end_line) END,
+         parent_symbol_path = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.parent_symbol_path ELSE COALESCE(EXCLUDED.parent_symbol_path, content_chunks.parent_symbol_path) END,
+         doc_comment = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.doc_comment ELSE COALESCE(EXCLUDED.doc_comment, content_chunks.doc_comment) END,
+         symbol_name_qualified = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_name_qualified ELSE COALESCE(EXCLUDED.symbol_name_qualified, content_chunks.symbol_name_qualified) END,
          modality = EXCLUDED.modality,
          embedding_image = COALESCE(EXCLUDED.embedding_image, content_chunks.embedding_image)`,
       params

src/core/postgres-engine.ts

@@ -2475,6 +2475,13 @@ export class PostgresEngine implements BrainEngine {
     //   - new is fresher (embedded_at > existing.embedded_at) → take new
     //   - otherwise → keep existing (slower writer with stale embedding loses)
     // Mirrored in pglite-engine.ts; pinned by test/e2e/concurrent-embed-race.test.ts.
+    //
+    // Code-chunk metadata columns (language / symbol_name / symbol_type / line range /
+    // parent_symbol_path / doc_comment / symbol_name_qualified) follow the SAME chunk_text-gated
+    // CASE pattern as `embedding` (#769). Re-chunk (chunk_text changed) trusts EXCLUDED outright;
+    // pure re-embed (chunk_text unchanged) COALESCEs so a caller that only carries embedding
+    // doesn't clobber metadata to NULL. Without this, every embed --stale pass nuked code-def's
+    // primary index for thousands of chunks at once.
     await sql.unsafe(
       `INSERT INTO content_chunks ${cols} VALUES ${rows.join(', ')}
        ON CONFLICT (page_id, chunk_index) DO UPDATE SET
@@ -2498,14 +2505,14 @@ export class PostgresEngine implements BrainEngine {
                 THEN EXCLUDED.embedded_at
            ELSE content_chunks.embedded_at
          END,
-         language = EXCLUDED.language,
-         symbol_name = EXCLUDED.symbol_name,
-         symbol_type = EXCLUDED.symbol_type,
-         start_line = EXCLUDED.start_line,
-         end_line = EXCLUDED.end_line,
-         parent_symbol_path = EXCLUDED.parent_symbol_path,
-         doc_comment = EXCLUDED.doc_comment,
-         symbol_name_qualified = EXCLUDED.symbol_name_qualified,
+         language = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.language ELSE COALESCE(EXCLUDED.language, content_chunks.language) END,
+         symbol_name = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_name ELSE COALESCE(EXCLUDED.symbol_name, content_chunks.symbol_name) END,
+         symbol_type = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_type ELSE COALESCE(EXCLUDED.symbol_type, content_chunks.symbol_type) END,
+         start_line = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.start_line ELSE COALESCE(EXCLUDED.start_line, content_chunks.start_line) END,
+         end_line = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.end_line ELSE COALESCE(EXCLUDED.end_line, content_chunks.end_line) END,
+         parent_symbol_path = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.parent_symbol_path ELSE COALESCE(EXCLUDED.parent_symbol_path, content_chunks.parent_symbol_path) END,
+         doc_comment = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.doc_comment ELSE COALESCE(EXCLUDED.doc_comment, content_chunks.doc_comment) END,
+         symbol_name_qualified = CASE WHEN EXCLUDED.chunk_text != content_chunks.chunk_text THEN EXCLUDED.symbol_name_qualified ELSE COALESCE(EXCLUDED.symbol_name_qualified, content_chunks.symbol_name_qualified) END,
          modality = EXCLUDED.modality,
          embedding_image = COALESCE(EXCLUDED.embedding_image, content_chunks.embedding_image)`,
       params as Parameters<typeof sql.unsafe>[1],

test/e2e/global-basename-pglite.test.ts

@@ -172,6 +172,59 @@ describe('issue #972 — DB-source (gbrain extract links --source db)', () => {
     expect(strk!.link_type).toBe('wikilink_basename');
   });
 
+  test('flag ON → path-qualified wikilink outside DIR_PATTERN resolves via DB path', async () => {
+    // `[[notes/struktura]]` — `notes` is not in DIR_PATTERN, so the ref
+    // reaches the generic pass with its dirname intact. Regression: the DB
+    // path queried the basename index with the raw literal (which is keyed
+    // by final segments only), so path-qualified wikilinks outside
+    // DIR_PATTERN silently produced zero edges while the FS path resolved
+    // the identical content.
+    await engine.putPage('notes/struktura', {
+      type: 'concept' as any, title: 'Struktura Notes',
+      compiled_truth: '', timeline: '',
+    });
+    await engine.putPage('concepts/knowledge-graph', {
+      type: 'concept', title: 'Knowledge Graph',
+      compiled_truth: 'Background in [[notes/struktura]].', timeline: '',
+    });
+    await engine.setConfig('link_resolution.global_basename', 'true');
+
+    await runExtract(engine, ['links', '--source', 'db']);
+
+    const outLinks = await engine.getLinks('concepts/knowledge-graph');
+    const strk = outLinks.find(l => l.to_slug === 'notes/struktura');
+    expect(strk).toBeDefined();
+    expect(strk!.link_type).toBe('wikilink_basename');
+    expect(strk!.link_source).toBe('wikilink-resolved');
+  });
+
+  test('path-qualified wikilink never attaches to a basename-only sibling', async () => {
+    // Both notes/struktura and wiki/struktura exist. The author wrote
+    // `[[notes/struktura]]` — the written path must exclude wiki/struktura
+    // (a bare `[[struktura]]` would legitimately match both).
+    await engine.putPage('notes/struktura', {
+      type: 'concept' as any, title: 'Struktura Notes',
+      compiled_truth: '', timeline: '',
+    });
+    await engine.putPage('wiki/struktura', {
+      type: 'concept' as any, title: 'Struktura Wiki',
+      compiled_truth: '', timeline: '',
+    });
+    await engine.putPage('concepts/x', {
+      type: 'concept', title: 'X',
+      compiled_truth: 'See [[notes/struktura]].', timeline: '',
+    });
+    await engine.setConfig('link_resolution.global_basename', 'true');
+
+    await runExtract(engine, ['links', '--source', 'db']);
+
+    const outLinks = await engine.getLinks('concepts/x');
+    const basenameLinks = outLinks
+      .filter(l => l.link_type === 'wikilink_basename')
+      .map(l => l.to_slug);
+    expect(basenameLinks).toEqual(['notes/struktura']);
+  });
+
   test('flag OFF → no basename edges via DB path (back-compat)', async () => {
     await engine.putPage('projects/struktura', {
       type: 'project', title: 'Struktura',

test/embed.serial.test.ts

@@ -803,3 +803,107 @@ describe('embedAllStale --source threading (D7)', () => {
     expect((firstCallOpts as { sourceId?: string }).sourceId).toBe('media-corpus');
   });
 });
+
+// ────────────────────────────────────────────────────────────────
+// Code metadata preservation across re-embed (regression for #769)
+// ────────────────────────────────────────────────────────────────
+//
+// gbrain v0.30.1 and earlier silently clobbered code-chunk metadata
+// (language, symbol_name, symbol_type, start_line, end_line,
+// parent_symbol_path, doc_comment, symbol_name_qualified) on every
+// re-embed pass. The chunker populated those columns at import time,
+// but embed.ts loaded chunks via getChunks then mapped them to a
+// stripped ChunkInput carrying only 5 fields. upsertChunks then
+// OVERWROTE (not COALESCEd) the metadata columns from EXCLUDED, so
+// re-embed wiped them to NULL. End result on a real brain: 4875 code
+// pages, 47866 chunks, all with NULL language/symbol_name/symbol_type;
+// code-def returned 0 hits across every indexed repo.
+//
+// All three runEmbed paths (--stale autopilot, --all, --slugs) must
+// thread metadata through the re-upsert. Tests below assert that the
+// engine.upsertChunks call carries the same metadata it loaded.
+
+describe('runEmbed preserves code-chunk metadata across re-embed (regression for #769)', () => {
+  const fullCodeChunk = {
+    chunk_index: 0,
+    chunk_text: '[Java] foo/Bar.java:10-20 method baz',
+    chunk_source: 'compiled_truth' as const,
+    embedded_at: null,
+    token_count: 12,
+    language: 'java',
+    symbol_name: 'baz',
+    symbol_type: 'function',
+    start_line: 10,
+    end_line: 20,
+    parent_symbol_path: ['Bar'],
+    doc_comment: 'does the thing',
+    symbol_name_qualified: 'Bar.baz',
+  };
+
+  function metadataOf(chunk: any) {
+    return {
+      language: chunk.language,
+      symbol_name: chunk.symbol_name,
+      symbol_type: chunk.symbol_type,
+      start_line: chunk.start_line,
+      end_line: chunk.end_line,
+      parent_symbol_path: chunk.parent_symbol_path,
+      doc_comment: chunk.doc_comment,
+      symbol_name_qualified: chunk.symbol_name_qualified,
+    };
+  }
+
+  test('--stale (autopilot path) carries code metadata into upsertChunks', async () => {
+    const stale = [{
+      slug: 'code-page',
+      chunk_index: 0,
+      chunk_text: fullCodeChunk.chunk_text,
+      chunk_source: 'compiled_truth',
+      model: null,
+      token_count: 12,
+    }];
+    let upsertChunkArgs: any[] | null = null;
+    const engine = mockEngine({
+      countStaleChunks: async () => 1,
+      listStaleChunks: async () => stale,
+      getChunks: async () => [fullCodeChunk],
+      upsertChunks: async (_slug: string, chunks: any[]) => { upsertChunkArgs = chunks; },
+    });
+
+    await runEmbed(engine, ['--stale']);
+
+    expect(upsertChunkArgs).not.toBeNull();
+    expect(upsertChunkArgs!).toHaveLength(1);
+    expect(metadataOf(upsertChunkArgs![0])).toEqual(metadataOf(fullCodeChunk));
+  });
+
+  test('--all (full re-embed) carries code metadata into upsertChunks', async () => {
+    let upsertChunkArgs: any[] | null = null;
+    const engine = mockEngine({
+      listPages: async () => [{ slug: 'code-page' }],
+      getChunks: async () => [fullCodeChunk],
+      upsertChunks: async (_slug: string, chunks: any[]) => { upsertChunkArgs = chunks; },
+    });
+
+    await runEmbed(engine, ['--all']);
+
+    expect(upsertChunkArgs).not.toBeNull();
+    expect(upsertChunkArgs!).toHaveLength(1);
+    expect(metadataOf(upsertChunkArgs![0])).toEqual(metadataOf(fullCodeChunk));
+  });
+
+  test('--slugs (per-page embed) carries code metadata into upsertChunks', async () => {
+    let upsertChunkArgs: any[] | null = null;
+    const engine = mockEngine({
+      getPage: async () => ({ slug: 'code-page', compiled_truth: 'x', timeline: '' }),
+      getChunks: async () => [fullCodeChunk],
+      upsertChunks: async (_slug: string, chunks: any[]) => { upsertChunkArgs = chunks; },
+    });
+
+    await runEmbed(engine, ['--slugs', 'code-page']);
+
+    expect(upsertChunkArgs).not.toBeNull();
+    expect(upsertChunkArgs!).toHaveLength(1);
+    expect(metadataOf(upsertChunkArgs![0])).toEqual(metadataOf(fullCodeChunk));
+  });
+});

test/link-extraction.test.ts

@@ -403,6 +403,77 @@ describe('extractPageLinks', () => {
     expect(candidates).toEqual([]);
   });
 
+  test('path-qualified wikilink outside DIR_PATTERN queries by final segment', async () => {
+    // `[[notes/struktura]]` (dir not in DIR_PATTERN) falls to the generic
+    // pass. The resolver's basename index is keyed by final path segments,
+    // so the lookup must strip the dirname — mirroring the FS path
+    // (resolveSlugAll). Regression: the raw literal was passed through,
+    // which never matched, so these links silently dropped.
+    const seen: string[] = [];
+    const resolver: SlugResolver = {
+      resolve: async () => null,
+      resolveBasenameMatches: async (name) => {
+        seen.push(name);
+        return name === 'struktura' ? ['notes/struktura'] : [];
+      },
+    };
+    const { candidates } = await extractPageLinks(
+      'concepts/x', 'See [[notes/struktura]].',
+      {}, 'concept', resolver, { globalBasename: true },
+    );
+    expect(seen).toContain('struktura');
+    expect(seen).not.toContain('notes/struktura');
+    expect(candidates.map(c => c.targetSlug)).toEqual(['notes/struktura']);
+    expect(candidates[0].linkType).toBe('wikilink_basename');
+    expect(candidates[0].linkSource).toBe('wikilink-resolved');
+  });
+
+  test('path-qualified wikilink keeps only matches ending with the written path', async () => {
+    // The written path disambiguates: `[[notes/struktura]]` must never
+    // attach to `wiki/struktura` even though both share the basename.
+    const resolver: SlugResolver = {
+      resolve: async () => null,
+      resolveBasenameMatches: async (name) =>
+        name === 'struktura' ? ['notes/struktura', 'wiki/struktura'] : [],
+    };
+    const { candidates } = await extractPageLinks(
+      'concepts/x', 'See [[notes/struktura]].',
+      {}, 'concept', resolver, { globalBasename: true },
+    );
+    expect(candidates.map(c => c.targetSlug)).toEqual(['notes/struktura']);
+  });
+
+  test('path-qualified wikilink matches a deeper real slug by path suffix', async () => {
+    // The page lives at vault/notes/struktura; the author wrote the shorter
+    // tail `[[notes/struktura]]`. Suffix matching connects them, while the
+    // basename-only sibling `wiki/struktura` stays excluded.
+    const resolver: SlugResolver = {
+      resolve: async () => null,
+      resolveBasenameMatches: async (name) =>
+        name === 'struktura' ? ['vault/notes/struktura', 'wiki/struktura'] : [],
+    };
+    const { candidates } = await extractPageLinks(
+      'concepts/x', 'See [[notes/struktura]].',
+      {}, 'concept', resolver, { globalBasename: true },
+    );
+    expect(candidates.map(c => c.targetSlug)).toEqual(['vault/notes/struktura']);
+  });
+
+  test('path-qualified self-link is dropped like the bare form', async () => {
+    // `[[notes/struktura]]` written on notes/struktura itself must not
+    // produce a self-loop (same guard as the bare `[[own-tail]]` case).
+    const resolver: SlugResolver = {
+      resolve: async () => null,
+      resolveBasenameMatches: async (name) =>
+        name === 'struktura' ? ['notes/struktura'] : [],
+    };
+    const { candidates } = await extractPageLinks(
+      'notes/struktura', 'See [[notes/struktura]].',
+      {}, 'concept', resolver, { globalBasename: true },
+    );
+    expect(candidates).toEqual([]);
+  });
+
   test('bare wikilink resolution does not interfere with DIR_PATTERN wikilinks', async () => {
     // 2b refs (people/alice) take the verb-inferred type;
     // 2c refs (struktura) take wikilink_basename. Same call.