feat(preprocess): resume evidence jobs by run id

harness/tests/test_corpus_pipeline.py

@@ -49,6 +49,13 @@ class Vectors:
             raise RuntimeError("partial write")
         return len(records)
 
+    def delete_generation(self, collection, generation):
+        self.records = [
+            value for value in self.records
+            if value.record.metadata["vector_generation"] != generation
+        ]
+        return 0
+
 
 def item(name, fingerprint):
     return SourceObject(
@@ -130,3 +137,39 @@ def test_dry_run_and_failed_acquire_never_change_active(tmp_path):
     with pytest.raises(PipelineError):
         pipeline(tmp_path, Source([(changed, RuntimeError("boom"))])).run()
     assert CorpusStore(tmp_path / "corpus").active_generation() == active
+
+
+def test_job_pipeline_uses_ordered_plan_and_returns_run_id(tmp_path):
+    one = item("one", "a")
+    candidate = pipeline(tmp_path, Source([(one, "hello")]))
+    result = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+    )
+    assert result.status == "succeeded"
+    assert result.run_id and len(result.run_id) == 32
+    checkpoint = tmp_path / ".tht-jobs" / "evidence" / "runs" / result.run_id / "checkpoint.json"
+    payload = __import__("json").loads(checkpoint.read_text())
+    assert [stage["name"] for stage in payload["stages"]] == [
+        "discover", "acquire_normalize_chunk", "embed", "vector_upsert",
+        "stage_validate", "publish", "retention_cleanup",
+    ]
+
+
+def test_job_pipeline_dry_run_only_discovers_and_reports_changes(tmp_path):
+    one = item("one", "a")
+    source = Source([(one, "hello")])
+    embedder = Embedder()
+    vectors = Vectors()
+    result = pipeline(tmp_path, source, embedder=embedder, vectors=vectors).run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+        dry_run=True,
+    )
+    assert result.changed == ("fs:one",)
+    assert source.acquire_calls == []
+    assert embedder.calls == []
+    assert vectors.records == []
+    assert result.generation is None and result.published is False