fix: address PR review comments (round 2)

Human reviewer (zmanian):
- H1: Accept configurable embedding dimension in LanceDbVectorStore::new()
  instead of hardcoding 1536. Dimension is sourced from
  EmbeddingProvider::dimension() at init time.
- H2: Skip double-write of embeddings to DB when external vector store
  is active (pass None to insert_chunk for embedding column).
- H3: Update PR title from "refactor" to "feat" (net-new feature).
- H4: Document non-atomic update_embedding in struct doc comment.

Bot reviewer (Copilot):
- Cache LanceDB table handle via tokio::sync::OnceCell (avoid
  open_table per operation).
- Cache Arc<Schema> in struct (avoid rebuilding per insert).
- Fix error variants: ChunkingFailed → EmbeddingFailed for LanceDB
  store/delete operations.
- Propagate store_embedding errors in reindex_document instead of
  warn-only (prevents silent data loss).
- Prefetch document metadata map in backfill_embeddings to avoid N+1
  queries.
- Add lancedb feature + protoc to CI test matrix so LanceDB tests
  actually run on Linux.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

.github/workflows/test.yml

@@ -14,7 +14,7 @@ jobs:
       matrix:
         include:
           - name: all-features
-            flags: "--features postgres,libsql,html-to-markdown"
+            flags: "--features postgres,libsql,lancedb,html-to-markdown"
           - name: default
             flags: ""
           - name: libsql-only
@@ -26,6 +26,11 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
         with:
           targets: wasm32-wasip2
+      - name: Install protoc (for lancedb)
+        if: matrix.name == 'all-features'
+        uses: arduino/setup-protoc@v3
+        with:
+          repo-token: ${{ secrets.GITHUB_TOKEN }}
       - uses: Swatinem/rust-cache@v2
         with:
           key: ${{ matrix.name }}