record cassettes against new fixtures, fix stale assertions

tests/ingester/test_pollers.py

@@ -265,7 +265,11 @@ async def test_storage_options_thread_through_to_job_extra(jobs, sync):
     poller = _periodic(source, cfg, jobs, sync)
     await poller._sweep_once()
     queued = await jobs.list_jobs(source_id="bucket")
-    assert queued[0].extra == {"storage_options": {"endpoint": "http://seaweed:8333"}}
+    # _otel is also threaded into extra so the worker's `ingester.job` span
+    # can nest under the sweep that enqueued it; assert the source-specific
+    # keys we care about and ignore the trace context payload.
+    assert queued[0].extra is not None
+    assert queued[0].extra["storage_options"] == {"endpoint": "http://seaweed:8333"}
 
 
 @pytest.mark.asyncio
@@ -282,7 +286,8 @@ async def test_http_headers_thread_through_to_job_extra(jobs, sync):
     poller = _periodic(source, cfg, jobs, sync)
     await poller._sweep_once()
     queued = await jobs.list_jobs(source_id="auth")
-    assert queued[0].extra == {"headers": {"Authorization": "Bearer abc"}}
+    assert queued[0].extra is not None
+    assert queued[0].extra["headers"] == {"Authorization": "Bearer abc"}
 
 
 # --- PollerManager lifecycle ---

tests/ingester/test_revision_round_trip.py

@@ -4,6 +4,8 @@ as unchanged. Catches the FS-specific bug where revision was lost in the
 pipeline and every periodic sweep re-enqueued every file forever.
 """
 
+from pathlib import Path
+
 import pytest
 
 from haiku.rag.client import HaikuRAG
@@ -11,7 +13,13 @@ from haiku.rag.ingester.sources.base import SourceEventKind
 from haiku.rag.ingester.sources.fs import FSSource
 
 
+@pytest.fixture(scope="module")
+def vcr_cassette_dir():
+    return str(Path(__file__).parent.parent / "cassettes" / "test_revision_round_trip")
+
+
 @pytest.mark.asyncio
+@pytest.mark.vcr()
 async def test_fs_ingest_writes_source_revision_to_metadata(temp_db_path, tmp_path):
     file_path = tmp_path / "doc.md"
     file_path.write_text("hello")
@@ -24,6 +32,7 @@ async def test_fs_ingest_writes_source_revision_to_metadata(temp_db_path, tmp_pa
 
 
 @pytest.mark.asyncio
+@pytest.mark.vcr()
 async def test_fs_second_sweep_emits_unchanged_after_ingest(temp_db_path, tmp_path):
     """The full round-trip: ingest a file, build a sync_state-shaped snapshot
     from document.metadata, hand it to FSSource.discover() — must see
@@ -46,6 +55,7 @@ async def test_fs_second_sweep_emits_unchanged_after_ingest(temp_db_path, tmp_pa
 
 
 @pytest.mark.asyncio
+@pytest.mark.vcr()
 async def test_fs_second_sweep_emits_upsert_when_file_changes(temp_db_path, tmp_path):
     """Counterpart to the unchanged test: a file modified after ingest still
     triggers UPSERT. Ensures the round-trip doesn't accidentally over-skip."""
@@ -72,6 +82,7 @@ async def test_fs_second_sweep_emits_upsert_when_file_changes(temp_db_path, tmp_
 
 
 @pytest.mark.asyncio
+@pytest.mark.vcr()
 async def test_fs_head_short_circuit_skips_fetch_for_unchanged_revision(
     temp_db_path, tmp_path, monkeypatch
 ):

tests/test_client_analyze.py

@@ -89,8 +89,10 @@ class TestClientAnalysisIntegration:
 
     @pytest.mark.asyncio
     @pytest.mark.vcr()
-    async def test_analyze_search_and_extract(self, allow_model_requests, temp_db_path):
-        pdf_path = Path("tests/data/doclaynet.pdf")
+    async def test_analyze_search_and_extract(
+        self, allow_model_requests, temp_db_path, doclaynet_first_page_pdf
+    ):
+        pdf_path = doclaynet_first_page_pdf
         config = AppConfig()
         config.processing.conversion_options.do_ocr = False
 
@@ -122,6 +124,11 @@ class TestClientAnalysisIntegration:
                 for label in expected_labels
                 if label in answer_lower or label.replace("-", " ") in answer_lower
             ]
-            assert len(found_labels) >= 6, (
-                f"Expected at least 6 labels, found {len(found_labels)}: {found_labels}"
+            # Page 0 of the DocLayNet paper (title + abstract) names fewer
+            # element types than the older single-page sample did. The
+            # assertion still checks that analyze() pulled real content out
+            # of the doc; ≥3 of the canonical 11 labels is a reasonable
+            # floor for a single-page extract.
+            assert len(found_labels) >= 3, (
+                f"Expected at least 3 labels, found {len(found_labels)}: {found_labels}"
             )

tests/test_context_enhancement.py

@@ -522,20 +522,20 @@ async def test_expand_context_no_base64_images(temp_db_path):
 
 
 @pytest.mark.vcr()
-async def test_expand_context_no_base64_images_docling_local(temp_db_path):
+async def test_expand_context_no_base64_images_docling_local(
+    temp_db_path, doclaynet_first_page_pdf
+):
     """Ensure expanded context from real PDF does not contain base64 image data.
 
-    Tests end-to-end with doclaynet.pdf using docling-local converter.
+    Tests end-to-end with a single page of doclaynet.pdf using docling-local converter.
     """
-    from pathlib import Path
-
     config = AppConfig()
     config.processing.converter = "docling-local"
     config.processing.chunker = "docling-local"
     config.processing.conversion_options.do_ocr = False
 
     async with HaikuRAG(temp_db_path, config=config, create=True) as client:
-        pdf_path = Path(__file__).parent / "data" / "doclaynet.pdf"
+        pdf_path = doclaynet_first_page_pdf
         result = await client.create_document_from_source(pdf_path)
         doc = result if not isinstance(result, list) else result[0]
         assert doc.id is not None