feat(evidence): evaluate retrieval with a small fixture

harness/tests/test_corpus_pipeline.py

@@ -106,7 +106,7 @@ def item(name, fingerprint):
 
 
 def pipeline(tmp_path, source, *, embedder=None, vectors=None, model="model-a", policy=None,
-             retain=3):
+             retain=3, candidate_evaluator=None):
     return CorpusPipeline(
         store=CorpusStore(tmp_path / "corpus"), sources=[source],
         embedder=embedder or Embedder(), vector_store=vectors or Vectors(),
@@ -114,6 +114,7 @@ def pipeline(tmp_path, source, *, embedder=None, vectors=None, model="model-a",
         chunk_policy=policy or ChunkPolicy(version="chunk-v1", max_chars=100),
         pipeline_version="evidence-v1",
         retain_published_generations=retain,
+        candidate_evaluator=candidate_evaluator,
     )
 
 
@@ -853,6 +854,55 @@ def test_dimension_mismatch_fails_before_vector_write_and_publish(tmp_path):
     assert candidate.store.active_generation() is None
 
 
+def test_failed_candidate_evaluation_never_switches_the_active_generation(tmp_path):
+    from types import SimpleNamespace
+
+    vectors = Vectors()
+    active = pipeline(tmp_path, Source([(item("one", "a"), "old")]), vectors=vectors).run().generation
+    candidate = pipeline(
+        tmp_path,
+        Source([(item("one", "b"), "new")]),
+        vectors=vectors,
+        candidate_evaluator=lambda manifest: SimpleNamespace(passed=False),
+    )
+
+    with pytest.raises(PipelineError, match="candidate retrieval evaluation failed"):
+        candidate.run()
+
+    assert candidate.store.active_generation() == active
+    assert {record.record.metadata["vector_generation"] for record in vectors.records} == {active}
+
+
+def test_job_failed_candidate_evaluation_never_switches_the_active_generation(tmp_path):
+    from types import SimpleNamespace
+
+    vectors = Vectors()
+    active = pipeline(tmp_path, Source([(item("one", "a"), "old")]), vectors=vectors).run_as_job(
+        workspace_id="demo",
+        workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+    ).generation
+    candidate = pipeline(
+        tmp_path,
+        Source([(item("one", "b"), "new")]),
+        vectors=vectors,
+        candidate_evaluator=lambda manifest: SimpleNamespace(passed=False),
+    )
+
+    result = candidate.run_as_job(
+        workspace_id="demo",
+        workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "3" * 64,
+    )
+
+    assert result.status == "failed"
+    assert result.published is False
+    assert candidate.store.active_generation() == active
+    assert {record.record.metadata["vector_generation"] for record in vectors.records} == {active}
+
+
 def test_pipeline_marks_each_evidence_fragment_for_server_side_italian_bm25(tmp_path):
     vectors = Vectors()