fix(jobs): bind completed effects to checkpoints

harness/tests/test_corpus_pipeline.py

@@ -62,6 +62,22 @@ class Vectors:
         return 0
 
 
+class InterruptingVectors(Vectors):
+    def __init__(self):
+        super().__init__()
+        self.batches = []
+        self.interrupt = True
+
+    def upsert(self, collection, records):
+        self.batches.append([value.record.id for value in records])
+        if self.interrupt:
+            self.interrupt = False
+            self.records.append(records[0])
+            raise KeyboardInterrupt("process interruption after partial write")
+        self.records.extend(records)
+        return len(records)
+
+
 def item(name, fingerprint):
     return SourceObject(
         source_id=f"fs:{name}", uri=f"file:///safe/{name}.md", fingerprint=f"sha256:{fingerprint}"
@@ -309,3 +325,35 @@ def test_job_pipeline_rejects_corrupt_required_artifacts_before_resume(
             input_fingerprint="sha256:" + "2" * 64,
             resume_run_id=crashed.name,
         )
+
+
+def test_vector_intent_is_reconciled_after_process_interruption_without_duplicate_upsert(tmp_path):
+    one = item("one", "a")
+    vectors = InterruptingVectors()
+    candidate = pipeline(
+        tmp_path, Source([(one, "a" * 250)]), vectors=vectors,
+        policy=ChunkPolicy(version="chunk-v1", max_chars=100),
+    )
+    with pytest.raises(KeyboardInterrupt):
+        candidate.run_as_job(
+            workspace_id="demo", workspace_root=tmp_path,
+            config_fingerprint="sha256:" + "1" * 64,
+            input_fingerprint="sha256:" + "2" * 64,
+        )
+    runs = tmp_path / ".tht-jobs" / "evidence" / "runs"
+    interrupted = next(runs.iterdir())
+    checkpoint = __import__("json").loads((interrupted / "checkpoint.json").read_text())
+    vector_stage = checkpoint["stages"][3]
+    assert vector_stage["status"] == "running"
+    assert vector_stage["effect_state"] == "intent"
+    first_written = vectors.batches[0][0]
+
+    result = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+        resume_run_id=interrupted.name,
+    )
+    assert result.status == "succeeded" and result.published is True
+    assert first_written not in vectors.batches[1]
+    assert len(vectors.records) == 3