diff --git a/paperforge/embedding/build_state.py b/paperforge/embedding/build_state.py
index 7cc174af..2403964c 100644
--- a/paperforge/embedding/build_state.py
+++ b/paperforge/embedding/build_state.py
@@ -3,6 +3,26 @@ from __future__ import annotations
 import json
 from pathlib import Path
 
+def _default_state() -> dict:
+    return {
+        "status": "idle",
+        "current": 0,
+        "total": 0,
+        "paper_id": "",
+        "last_update": "",
+        "started_at": "",
+        "finished_at": "",
+        "resume_supported": True,
+        "mode": "api",
+        "model": "text-embedding-3-small",
+        "message": "",
+        "pid": 0,
+    }
+
+
+def _fallback_state() -> dict:
+    return {"status": "idle", "current": 0, "total": 0, "paper_id": ""}
+
 
 def get_vector_build_state_path(vault: Path) -> Path:
     from paperforge.config import paperforge_paths
@@ -14,41 +34,32 @@ def get_vector_build_state_path(vault: Path) -> Path:
 def read_vector_build_state(vault: Path) -> dict:
     path = get_vector_build_state_path(vault)
     if not path.exists():
-        return {
-            "status": "idle",
-            "current": 0,
-            "total": 0,
-            "paper_id": "",
-            "last_update": "",
-            "started_at": "",
-            "finished_at": "",
-            "resume_supported": True,
-            "mode": "api",
-            "model": "text-embedding-3-small",
-            "message": "",
-            "pid": 0,
-        }
+        return _default_state()
+    raw = path.read_text(encoding="utf-8")
     try:
-        return json.loads(path.read_text(encoding="utf-8"))
+        return json.loads(raw)
     except Exception:
-        return {"status": "idle", "current": 0, "total": 0, "paper_id": ""}
+        # Main file corrupt — try .tmp backup from failed atomic write
+        tmp = path.with_suffix(".tmp")
+        if tmp.exists():
+            try:
+                return json.loads(tmp.read_text(encoding="utf-8"))
+            except Exception:
+                pass
+        return _fallback_state()
 
 
 def write_vector_build_state(vault: Path, state: dict) -> None:
     path = get_vector_build_state_path(vault)
     path.parent.mkdir(parents=True, exist_ok=True)
     tmp = path.with_suffix(".tmp")
-    tmp.write_text(json.dumps(state, ensure_ascii=False, indent=2), encoding="utf-8")
+    content = json.dumps(state, ensure_ascii=False, indent=2)
+    tmp.write_text(content, encoding="utf-8")
     try:
         tmp.replace(path)
     except OSError:
-        # Windows: target file may be locked (e.g., Obsidian plugin reading it).
-        # Fall back to direct write — atomicity is secondary to avoiding crash.
-        path.write_text(tmp.read_text(encoding="utf-8"), encoding="utf-8")
-        try:
-            tmp.unlink()
-        except OSError:
-            pass
+        # Windows: target file locked. Write directly; leave .tmp as backup.
+        path.write_text(content, encoding="utf-8")
 
 def mark_vector_build_state(vault: Path, **fields) -> dict:
     state = read_vector_build_state(vault)
diff --git a/tests/test_embed_integration.py b/tests/test_embed_integration.py
new file mode 100644
index 00000000..42f33ada
--- /dev/null
+++ b/tests/test_embed_integration.py
@@ -0,0 +1,307 @@
+"""Integration tests for the embed pipeline with a real ChromaDB (EphemeralClient).
+
+Tests the actual encode → write → retrieve → delete cycle against an
+in-memory ChromaDB, with the embedding provider mocked to return fixed vectors.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+from unittest.mock import patch
+
+import chromadb
+import pytest
+
+import paperforge.config
+from paperforge.worker._utils import pipeline_paths as _pp
+
+paperforge.config.pipeline_paths = _pp
+
+from paperforge.embedding.builder import (
+    PaperEmbeddingJob,
+    encode_payload,
+    encode_paper_job,
+    write_encoded_payload,
+    prepare_payloads_for_entry,
+    prepare_legacy_payload,
+    prepare_body_payload,
+    prepare_object_payload,
+)
+from paperforge.embedding.search import merge_retrieve
+from paperforge.embedding._chroma import delete_paper_vectors
+
+
+# ---------------------------------------------------------------------------
+# Fixture: ephemeral ChromaDB
+# ---------------------------------------------------------------------------
+
+COLLECTION_NAMES = ["paperforge_fulltext", "paperforge_body", "paperforge_objects"]
+EMBEDDING_DIM = 4  # small dimension for fast tests
+
+
+@pytest.fixture(scope="module")
+def ephe_client():
+    """Shared EphemeralClient for the module — data persists across tests."""
+    return chromadb.EphemeralClient()
+
+
+def _make_fake_get_collection(client):
+    """Factory: returns a get_collection function backed by the given client."""
+    def _fake(vault, name="paperforge_fulltext"):
+        return client.get_or_create_collection(
+            name=name,
+            metadata={"hnsw:space": "cosine"},
+        )
+    return _fake
+
+
+@pytest.fixture(autouse=True)
+def _patch_collections(ephe_client):
+    """Replace all get_collection references to use the ephemeral client."""
+    fake = _make_fake_get_collection(ephe_client)
+    # Patch at the source and every module that does `from _chroma import get_collection`
+    with patch("paperforge.embedding._chroma.get_collection", fake), \
+         patch("paperforge.embedding.builder.get_collection", fake), \
+         patch("paperforge.embedding.search.get_collection", fake):
+        yield
+
+
+@pytest.fixture(autouse=True)
+def _clear_collections(ephe_client):
+    """Clear all collections before each test."""
+    for name in COLLECTION_NAMES:
+        try:
+            ephe_client.delete_collection(name)
+        except Exception:
+            pass
+    yield
+
+
+@pytest.fixture(autouse=True)
+def _patch_providers(mock_provider):
+    """Replace OpenAICompatibleProvider in both builder and search modules."""
+    with patch("paperforge.embedding.builder.OpenAICompatibleProvider", return_value=mock_provider), \
+         patch("paperforge.embedding.search.OpenAICompatibleProvider", return_value=mock_provider):
+        yield
+# Mock provider factory
+# ---------------------------------------------------------------------------
+
+
+class FixedProvider:
+    """Provider that returns deterministic embeddings for any text."""
+
+    def __init__(self, vault: Path | None = None):
+        self.vault = vault
+
+    def encode(self, texts: list[str]) -> list[list[float]]:
+        import hashlib
+        result = []
+        for t in texts:
+            h = hashlib.md5(t.encode()).digest()
+            # Convert first EMBEDDING_DIM bytes to floats in [0, 1]
+            vec = [b / 255.0 for b in h[:EMBEDDING_DIM]]
+            result.append(vec)
+        return result
+
+    def encode_single(self, text: str) -> list[float]:
+        return self.encode([text])[0]
+
+
+@pytest.fixture
+def mock_provider():
+    return FixedProvider()
+
+
+@pytest.fixture(autouse=True)
+def _patch_provider(mock_provider):
+    """Replace OpenAICompatibleProvider with FixedProvider in builder."""
+    with patch("paperforge.embedding.builder.OpenAICompatibleProvider", return_value=mock_provider):
+        yield
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def make_body_units(paper_id: str, n: int = 2) -> list[dict]:
+    return [
+        {
+            "unit_id": f"{paper_id}:body:n{n}:1-1:1-1",
+            "paper_id": paper_id,
+            "section_path": f"s{n}",
+            "section_level": 1,
+            "section_title": f"Section {n}",
+            "unit_text": f"This is body text {n} for paper {paper_id}.",
+            "unit_kind": "body",
+            "part_ordinal": 0,
+            "token_estimate": 10,
+        }
+        for n in range(1, n + 1)
+    ]
+
+
+def make_object_units(paper_id: str, n: int = 2) -> list[dict]:
+    return [
+        {
+            "unit_id": f"{paper_id}:object:f{n}:1-1:1-1",
+            "paper_id": paper_id,
+            "section_path": "results",
+            "object_kind": "figure",
+            "object_label": f"Figure {n}",
+            "caption_text": f"Caption for figure {n}.",
+            "nearby_body_text": f"Nearby text for figure {n}.",
+            "token_estimate": 15,
+        }
+        for n in range(1, n + 1)
+    ]
+
+
+# ---------------------------------------------------------------------------
+# Tests: basic payload preparation
+# ---------------------------------------------------------------------------
+
+class TestPayloadPrep:
+    """Verify payload creation works (lightweight, not mock-dependent)."""
+
+    def test_body_payload_has_correct_structure(self):
+        units = make_body_units("p1", 2)
+        payload = prepare_body_payload("p1", units)
+        assert payload.collection_name == "paperforge_body"
+        assert len(payload.texts) == 2
+        assert payload.ids == [u["unit_id"] for u in units]
+        assert all(m["unit_kind"] == "body" for m in payload.metadatas)
+
+    def test_object_payload_joins_texts(self):
+        units = make_object_units("p1", 1)
+        payload = prepare_object_payload("p1", units)
+        assert payload.collection_name == "paperforge_objects"
+        assert "Figure 1" in payload.texts[0]
+        assert "Caption for figure 1" in payload.texts[0]
+
+    def test_legacy_payload_uses_fulltext_collection(self, tmp_path):
+        chunks = [{"text": "Hello", "chunk_index": 0, "section": "intro",
+                    "page_number": 1, "token_estimate": 5}]
+        payload = prepare_legacy_payload("k1", chunks)
+        assert payload.collection_name == "paperforge_fulltext"
+        assert payload.texts == ["Hello"]
+
+
+# ---------------------------------------------------------------------------
+# Tests: encode → write → retrieve cycle with real ChromaDB
+# ---------------------------------------------------------------------------
+
+class TestEmbedRoundTrip:
+    """Integration tests against EphemeralClient."""
+
+    def test_write_and_retrieve_body_units(self, tmp_path, mock_provider):
+        """Write body units to ChromaDB, retrieve them back."""
+        key = "p1"
+        units = make_body_units(key, 2)
+        payload = prepare_body_payload(key, units)
+        encoded = encode_payload(tmp_path, payload)  # uses mock_provider via patch
+        write_encoded_payload(tmp_path, encoded)
+
+        # Retrieve across all collections
+        results = merge_retrieve(tmp_path, "body text", limit=5)
+        assert len(results) >= 1
+        assert results[0]["paper_id"] == key
+        assert results[0]["source"] == "body_unit"
+
+    def test_write_and_retrieve_object_units(self, tmp_path, mock_provider):
+        """Write object units, retrieve them."""
+        key = "p1"
+        units = make_object_units(key, 2)
+        payload = prepare_object_payload(key, units)
+        encoded = encode_payload(tmp_path, payload)
+        write_encoded_payload(tmp_path, encoded)
+
+        results = merge_retrieve(tmp_path, "figure caption", limit=5)
+        assert len(results) >= 1
+        assert results[0]["source"] == "object_unit"
+
+    def test_write_and_retrieve_legacy_chunks(self, tmp_path, mock_provider):
+        """Write legacy chunks, retrieve them."""
+        key = "p1"
+        chunks = [
+            {"text": f"Chunk {i}", "chunk_index": i, "section": "intro",
+             "page_number": 1, "token_estimate": 5}
+            for i in range(3)
+        ]
+        payload = prepare_legacy_payload(key, chunks)
+        encoded = encode_payload(tmp_path, payload)
+        write_encoded_payload(tmp_path, encoded)
+
+        results = merge_retrieve(tmp_path, "Chunk", limit=5)
+        assert len(results) >= 1
+        assert results[0]["source"] == "legacy_chunk"
+
+    def test_multiple_papers_retrieve_respects_per_paper_cap(self, tmp_path, mock_provider):
+        """merge_retrieve caps at 2 results per paper."""
+        for i in range(3):
+            key = f"p{i}"
+            payload = prepare_legacy_payload(key, [
+                {"text": f"Paper {i} chunk", "chunk_index": 0, "section": "intro",
+                 "page_number": 1, "token_estimate": 5},
+                {"text": f"Paper {i} more", "chunk_index": 1, "section": "methods",
+                 "page_number": 1, "token_estimate": 5},
+                {"text": f"Paper {i} extra", "chunk_index": 2, "section": "results",
+                 "page_number": 1, "token_estimate": 5},
+            ])
+            encoded = encode_payload(tmp_path, payload)
+            write_encoded_payload(tmp_path, encoded)
+
+        results = merge_retrieve(tmp_path, "Paper", limit=10)
+        # Each paper should have at most 2 results
+        from collections import Counter
+        counts = Counter(r["paper_id"] for r in results)
+        assert all(v <= 2 for v in counts.values())
+
+    def test_delete_paper_vectors_removes_data(self, tmp_path, mock_provider):
+        """After delete, retrieval returns no results for that paper."""
+        key = "p1"
+        units = make_body_units(key, 1)
+        payload = prepare_body_payload(key, units)
+        encoded = encode_payload(tmp_path, payload)
+        write_encoded_payload(tmp_path, encoded)
+
+        # Confirm it's there
+        assert len(merge_retrieve(tmp_path, "body", limit=5)) >= 1
+
+        # Delete
+        delete_paper_vectors(tmp_path, key)
+
+        # Confirm it's gone
+        results = merge_retrieve(tmp_path, "body", limit=5)
+        pid_results = [r for r in results if r["paper_id"] == key]
+        assert len(pid_results) == 0
+
+    def test_prepare_payloads_for_entry_returns_body_and_object(self, tmp_path, mock_provider):
+        """prepare_payloads_for_entry returns correct payload types."""
+        key = "p1"
+        body_units = make_body_units(key, 1)
+        object_units = make_object_units(key, 1)
+
+        payloads = prepare_payloads_for_entry(
+            tmp_path, key, has_body=True, has_object=True,
+            body_units=body_units, object_units=object_units,
+        )
+        assert payloads is not None
+        assert len(payloads) == 2
+        names = [p.collection_name for p in payloads]
+        assert "paperforge_body" in names
+        assert "paperforge_objects" in names
+
+    def test_encode_paper_job_processes_all_payloads(self, tmp_path, mock_provider):
+        """encode_paper_job encodes all payloads for a paper."""
+        key = "p1"
+        body_payload = prepare_body_payload(key, make_body_units(key, 2))
+        obj_payload = prepare_object_payload(key, make_object_units(key, 1))
+        job = PaperEmbeddingJob(paper_id=key, payloads=[body_payload, obj_payload])
+
+        bundle = encode_paper_job(tmp_path, job)
+        assert bundle.paper_id == key
+        assert len(bundle.payloads) == 2
+        assert bundle.chunk_count == 3  # 2 body + 1 object
+        # Verify all payloads have embeddings
+        for p in bundle.payloads:
+            assert len(p.embeddings) > 0
+            assert len(p.embeddings[0]) == EMBEDDING_DIM