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

harness/tests/test_job_runner.py

@@ -64,6 +64,27 @@ def test_failed_stage_is_resumable_and_skips_completed_stage(tmp_path):
     assert calls == [("discover", False), ("acquire", False), ("recovered", False)]
 
 
+def test_resume_carries_successful_stage_artifacts_into_new_run(tmp_path):
+    def discover(context):
+        artifacts = context.run_dir / "artifacts"
+        artifacts.mkdir()
+        (artifacts / "discovery.json").write_text('{"source":"one"}')
+
+    first = run_job(
+        _spec(tmp_path, stage_ids=("discover", "acquire")),
+        [discover, lambda _context: (_ for _ in ()).throw(RuntimeError("crash"))],
+    )
+
+    def acquire(context):
+        assert (context.run_dir / "artifacts" / "discovery.json").read_text() == '{"source":"one"}'
+
+    resumed = run_job(
+        _spec(tmp_path, resume_run_id=first.run_id, stage_ids=("discover", "acquire")),
+        [discover, acquire],
+    )
+    assert resumed.status == "succeeded"
+
+
 def test_successful_job_is_idempotently_resumable(tmp_path):
     calls = []
 

harness/tests/test_preprocess_cli.py

@@ -30,3 +30,27 @@ def test_preprocess_failure_is_structured_and_nonzero(monkeypatch, tmp_path):
     assert response.exit_code != 0
     assert json.loads(response.output) == {"status": "failed", "error": "preprocessing failed"}
     assert "secret detail" not in response.output
+
+
+def test_preprocess_resume_rejects_generation_id_before_configuration(monkeypatch, tmp_path):
+    import tht.cli.preprocess_cmd as command
+
+    called = False
+
+    def forbidden(*args, **kwargs):
+        nonlocal called
+        called = True
+
+    monkeypatch.setattr(command, "run_from_config", forbidden)
+    response = CliRunner().invoke(
+        app,
+        [
+            "preprocess", "evidence", "--resume", "gen:" + "a" * 32,
+            "--json", "-c", str(tmp_path / "workspace.yaml"),
+        ],
+    )
+    assert response.exit_code != 0
+    assert json.loads(response.output) == {
+        "status": "failed", "error": "resume requires a preprocessing run id"
+    }
+    assert called is False