fix(preprocess): harden crash recovery integrity

.superpowers/sdd/evidence-task-5b-report.md

@@ -41,3 +41,33 @@ green.
 One pre-existing Pydantic serialization warning is exposed by the new end-to-end job test when
 canonical metadata contains frozen tuple values; it does not contaminate CLI stdout. Retention is
 an explicit stable no-op until a retention policy is configured.
+
+## Review fix wave — crash consistency and artifact integrity
+
+Addressed all five follow-up findings:
+
+- `JobRunner` now supports a test-only post-call/pre-checkpoint fault hook. Each stage seals a
+  canonical artifact manifest containing required flat filenames, SHA-256, byte size, producer
+  stage, and the full spec compatibility fingerprint. Resume validates the checkpoint and every
+  sealed artifact before allocating/copying a new run, rejecting missing, tampered, extra, nested,
+  or symlinked state. A sealed `running` stage is promoted after a simulated process crash; a
+  sealed `failed` stage is deliberately retried.
+- Vector intent (exact record IDs and content hashes) is sealed before upsert. Execution reconciles
+  `existing_hashes` and writes only missing/mismatched rows. Crash-after-effect tests prove no
+  duplicate acquire, embed, or vector upsert.
+- Raw upsert, stage, recovery-upsert, recovery-stage, and publish exceptions compensate the exact
+  generation. Compensation markers survive failed checkpoints; resume rotates the generation,
+  refreshes generation-bound artifacts, reconciles vectors, and stages idempotently.
+- `CorpusStore.publish` is idempotent and failure-atomic. If replace succeeds but directory fsync
+  fails, it restores the previous `ACTIVE` value (or removes a newly created pointer), fsyncs the
+  rollback, and re-raises. Pipeline cleanup refuses to discard a generation referenced by ACTIVE.
+- Added crash/resume coverage after all seven ordered stages; corrupt/missing plan, manifest, and
+  embeddings; unsafe extra paths; nonexistent run IDs; raw vector/stage failures; and post-replace
+  ACTIVE rollback.
+
+Fresh fix-wave verification:
+
+- Focused jobs/corpus/CLI/search suite: `82 passed, 17 warnings`.
+- Available harness suite (same sandbox exclusions described above):
+  `579 passed, 5 deselected, 31 warnings`.
+- Scoped Ruff and `git diff --check`: clean.

harness/tests/test_corpus_pipeline.py

@@ -49,6 +49,11 @@ class Vectors:
             raise RuntimeError("partial write")
         return len(records)
 
+    def existing_hashes(self, collection, kinds):
+        return {
+            value.record.id: value.content_hash for value in self.records
+        }
+
     def delete_generation(self, collection, generation):
         self.records = [
             value for value in self.records
@@ -173,3 +178,134 @@ def test_job_pipeline_dry_run_only_discovers_and_reports_changes(tmp_path):
     assert embedder.calls == []
     assert vectors.records == []
     assert result.generation is None and result.published is False
+
+
+@pytest.mark.parametrize("crash_stage", [
+    "discover", "acquire_normalize_chunk", "embed", "vector_upsert",
+    "stage_validate", "publish", "retention_cleanup",
+])
+def test_job_pipeline_crash_after_each_stage_resumes_without_duplicate_effects(tmp_path, crash_stage):
+    one = item("one", "a")
+    source = Source([(one, "hello")])
+    embedder = Embedder()
+    vectors = Vectors()
+    candidate = pipeline(tmp_path, source, embedder=embedder, vectors=vectors)
+
+    class Crash(BaseException):
+        pass
+
+    def fault(_context, stage):
+        if stage == crash_stage:
+            raise Crash()
+
+    with pytest.raises(Crash):
+        candidate.run_as_job(
+            workspace_id="demo", workspace_root=tmp_path,
+            config_fingerprint="sha256:" + "1" * 64,
+            input_fingerprint="sha256:" + "2" * 64,
+            after_stage_return=fault,
+        )
+    runs = tmp_path / ".tht-jobs" / "evidence" / "runs"
+    crashed = next(runs.iterdir()).name
+    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=crashed,
+    )
+    assert result.status == "succeeded"
+    assert source.acquire_calls == ["fs:one"]
+    assert len(embedder.calls) == 1
+    assert len(vectors.records) == 1
+
+
+def test_job_pipeline_raw_upsert_failure_compensates_and_resumes_with_new_generation(tmp_path):
+    one = item("one", "a")
+    vectors = Vectors(fail=True)
+    candidate = pipeline(tmp_path, Source([(one, "hello")]), vectors=vectors)
+    first = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+    )
+    assert first.status == "failed"
+    assert vectors.records == []
+    old_generation = first.generation
+    vectors.fail = False
+    resumed = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+        resume_run_id=first.run_id,
+    )
+    assert resumed.status == "succeeded", resumed
+    assert resumed.generation != old_generation
+    assert candidate.store.active_generation() == resumed.generation
+
+
+def test_job_pipeline_raw_stage_failure_compensates_vectors_and_resumes(tmp_path, monkeypatch):
+    one = item("one", "a")
+    vectors = Vectors()
+    candidate = pipeline(tmp_path, Source([(one, "hello")]), vectors=vectors)
+    real_stage = candidate.store.stage
+    calls = 0
+
+    def fail_once(*args, **kwargs):
+        nonlocal calls
+        calls += 1
+        if calls == 1:
+            raise OSError("raw stage failure")
+        return real_stage(*args, **kwargs)
+
+    monkeypatch.setattr(candidate.store, "stage", fail_once)
+    first = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+    )
+    assert first.status == "failed" and vectors.records == []
+    resumed = candidate.run_as_job(
+        workspace_id="demo", workspace_root=tmp_path,
+        config_fingerprint="sha256:" + "1" * 64,
+        input_fingerprint="sha256:" + "2" * 64,
+        resume_run_id=first.run_id,
+    )
+    assert resumed.status == "succeeded", resumed
+
+
+@pytest.mark.parametrize("stage,filename", [
+    ("discover", "plan.json"),
+    ("acquire_normalize_chunk", "manifest.json"),
+    ("embed", "embeddings.json"),
+])
+@pytest.mark.parametrize("mutation", ["missing", "tampered"])
+def test_job_pipeline_rejects_corrupt_required_artifacts_before_resume(
+    tmp_path, stage, filename, mutation,
+):
+    one = item("one", "a")
+    candidate = pipeline(tmp_path, Source([(one, "hello")]))
+
+    class Crash(BaseException):
+        pass
+
+    with pytest.raises(Crash):
+        candidate.run_as_job(
+            workspace_id="demo", workspace_root=tmp_path,
+            config_fingerprint="sha256:" + "1" * 64,
+            input_fingerprint="sha256:" + "2" * 64,
+            after_stage_return=lambda _context, name: (
+                (_ for _ in ()).throw(Crash()) if name == stage else None
+            ),
+        )
+    runs = tmp_path / ".tht-jobs" / "evidence" / "runs"
+    crashed = next(runs.iterdir())
+    target = crashed / "artifacts" / filename
+    target.unlink() if mutation == "missing" else target.write_text("tampered")
+    from tht.jobs.runner import CorruptCheckpointError
+    with pytest.raises(CorruptCheckpointError, match="artifact"):
+        candidate.run_as_job(
+            workspace_id="demo", workspace_root=tmp_path,
+            config_fingerprint="sha256:" + "1" * 64,
+            input_fingerprint="sha256:" + "2" * 64,
+            resume_run_id=crashed.name,
+        )

harness/tests/test_corpus_publish.py

@@ -41,3 +41,18 @@ def test_active_pointer_cannot_escape_generation_root(tmp_path):
     store.active_path.write_text("../outside\n")
     with pytest.raises(UnsafeCorpusPath):
         store.active_manifest()
+
+
+def test_publish_restores_previous_active_when_directory_fsync_fails_after_replace(tmp_path, monkeypatch):
+    store = CorpusStore(tmp_path / "corpus")
+    first = store.stage(CorpusManifest(), {})
+    second = store.stage(CorpusManifest(), {})
+    store.publish(first)
+    def fail_once():
+        store._fsync_directory = store._sync_root
+        raise OSError("post replace crash")
+
+    store._fsync_directory = fail_once
+    with pytest.raises(OSError, match="post replace"):
+        store.publish(second)
+    assert store.active_generation() == first

harness/tests/test_job_runner.py

@@ -4,7 +4,7 @@ import pytest
 from pydantic import ValidationError
 
 from tht.jobs.models import JobSpec
-from tht.jobs.runner import CorruptCheckpointError, run_job
+from tht.jobs.runner import CorruptCheckpointError, StageArtifacts, run_job
 import tht.jobs.runner as runner_module
 
 
@@ -69,6 +69,7 @@ def test_resume_carries_successful_stage_artifacts_into_new_run(tmp_path):
         artifacts = context.run_dir / "artifacts"
         artifacts.mkdir()
         (artifacts / "discovery.json").write_text('{"source":"one"}')
+        return StageArtifacts(("discovery.json",))
 
     first = run_job(
         _spec(tmp_path, stage_ids=("discover", "acquire")),
@@ -85,6 +86,65 @@ def test_resume_carries_successful_stage_artifacts_into_new_run(tmp_path):
     assert resumed.status == "succeeded"
 
 
+def test_crash_after_stage_effect_resumes_without_repeating_stage(tmp_path):
+    calls = []
+
+    def stage(context):
+        calls.append("stage")
+        artifacts = context.run_dir / "artifacts"
+        artifacts.mkdir()
+        (artifacts / "effect.json").write_text("ok")
+        return StageArtifacts(("effect.json",))
+
+    class Crash(BaseException):
+        pass
+
+    with pytest.raises(Crash):
+        run_job(
+            _spec(tmp_path), [stage],
+            after_stage_return=lambda *_: (_ for _ in ()).throw(Crash()),
+        )
+    runs = tmp_path / ".tht-jobs" / "evidence" / "runs"
+    crashed_run = next(runs.iterdir()).name
+    resumed = run_job(_spec(tmp_path).with_resume(crashed_run), [stage])
+    assert resumed.status == "succeeded"
+    assert calls == ["stage"]
+
+
+def test_resume_rejects_tampered_successful_stage_artifact(tmp_path):
+    def stage(context):
+        artifacts = context.run_dir / "artifacts"
+        artifacts.mkdir()
+        (artifacts / "effect.json").write_text("ok")
+        return StageArtifacts(("effect.json",))
+
+    report = run_job(_spec(tmp_path), [stage])
+    path = tmp_path / ".tht-jobs" / "evidence" / "runs" / report.run_id / "artifacts" / "effect.json"
+    path.write_text("tampered")
+    with pytest.raises(CorruptCheckpointError, match="artifact"):
+        run_job(_spec(tmp_path).with_resume(report.run_id), [stage])
+
+
+def test_resume_rejects_extra_symlink_before_any_stage(tmp_path):
+    report = run_job(_spec(tmp_path), [lambda _context: StageArtifacts()])
+    artifacts = tmp_path / ".tht-jobs" / "evidence" / "runs" / report.run_id / "artifacts"
+    (artifacts / "unsafe").symlink_to(tmp_path)
+    called = False
+
+    def forbidden(_context):
+        nonlocal called
+        called = True
+
+    with pytest.raises(CorruptCheckpointError, match="artifact"):
+        run_job(_spec(tmp_path).with_resume(report.run_id), [forbidden])
+    assert called is False
+
+
+def test_nonexistent_well_formed_resume_run_id_is_rejected(tmp_path):
+    with pytest.raises(CorruptCheckpointError, match="checkpoint"):
+        run_job(_spec(tmp_path).with_resume("a" * 32), [lambda _context: None])
+
+
 def test_successful_job_is_idempotently_resumable(tmp_path):
     calls = []
 

harness/tht/corpus/pipeline.py

@@ -16,7 +16,7 @@ from tht.ports.evidence import EvidenceSource, SourceObject
 from tht.ports.vector import VectorStore, VectorWriteRecord
 from tht.vectorstore.records import VectorRecord
 from tht.jobs.models import JobSpec
-from tht.jobs.runner import JobContext, run_job
+from tht.jobs.runner import JobContext, StageArtifacts, run_job, seal_stage_artifacts
 
 
 EVIDENCE_STAGE_IDS = (
@@ -96,6 +96,7 @@ class CorpusPipeline:
         input_fingerprint: str,
         dry_run: bool = False,
         resume_run_id: str | None = None,
+        after_stage_return=None,
     ) -> PipelineResult:
         """Execute preprocessing through the durable shared job envelope."""
         discovered = self._discover()
@@ -159,10 +160,11 @@ class CorpusPipeline:
                 "removed": removed,
                 "previous": previous.model_dump(mode="json") if previous else None,
             })
+            return StageArtifacts(("plan.json",))
 
         def acquire_stage(context: JobContext) -> None:
             if context.dry_run:
-                return
+                return StageArtifacts()
             plan = read(context, "plan.json")
             previous = CorpusManifest.model_validate(plan["previous"]) if plan["previous"] else None
             prior = {doc.source_id: doc for doc in previous.documents} if previous else {}
@@ -194,10 +196,11 @@ class CorpusPipeline:
                 },
             )
             write(context, "manifest.json", manifest.model_dump(mode="json"))
+            return StageArtifacts(("manifest.json",))
 
         def embed_stage(context: JobContext) -> None:
             if context.dry_run:
-                return
+                return StageArtifacts()
             plan = read(context, "plan.json")
             manifest = CorpusManifest.model_validate(read(context, "manifest.json"))
             changed_docs = {doc.document_id for doc in manifest.documents if doc.source_id in plan["changed"]}
@@ -208,6 +211,7 @@ class CorpusPipeline:
             ):
                 raise PipelineError("embedding output is incompatible")
             write(context, "embeddings.json", embeddings)
+            return StageArtifacts(("embeddings.json",))
 
         def records(context: JobContext):
             plan = read(context, "plan.json")
@@ -220,7 +224,8 @@ class CorpusPipeline:
 
         def compensate(context: JobContext) -> None:
             generation = read(context, "plan.json")["generation"]
-            self.store.discard(generation)
+            if self.store.active_generation() != generation:
+                self.store.discard(generation)
             try:
                 self.vector_store.delete_generation("evidence", generation)
             except Exception:
@@ -241,62 +246,101 @@ class CorpusPipeline:
             for document in manifest.documents:
                 if document.source_id in changed and generations.get(document.document_id) == old:
                     generations[document.document_id] = plan["generation"]
-            metadata = dict(manifest.metadata)
-            metadata["document_generations"] = generations
-            manifest = manifest.model_copy(update={
-                "vector_generation": plan["generation"], "metadata": metadata,
-            })
+            manifest_payload = manifest.model_dump(mode="json")
+            manifest_payload["metadata"]["document_generations"] = generations
+            manifest_payload["vector_generation"] = plan["generation"]
+            manifest = CorpusManifest.model_validate(manifest_payload)
             write(context, "manifest.json", manifest.model_dump(mode="json"))
             marker.unlink()
 
         def vector_stage(context: JobContext) -> None:
             if context.dry_run:
-                return
+                return StageArtifacts()
             rotate_compensated_generation(context)
             values = records(context)
+            write(context, "vector-intent.json", {
+                "generation": read(context, "plan.json")["generation"],
+                "records": {value.record.id: value.content_hash for value in values},
+            })
+            seal_stage_artifacts(
+                context, "vector_upsert",
+                ("plan.json", "manifest.json", "vector-intent.json"), spec,
+            )
             try:
-                if values and self.vector_store.upsert("evidence", values) != len(values):
+                existing = self.vector_store.existing_hashes("evidence", ["evidence"])
+                missing = [
+                    value for value in values
+                    if existing.get(value.record.id) != value.content_hash
+                ]
+                if missing and self.vector_store.upsert("evidence", missing) != len(missing):
                     raise PipelineError("vector write count mismatch")
             except Exception:
                 compensate(context)
                 raise
+            return StageArtifacts(("plan.json", "manifest.json", "vector-intent.json"))
 
         def stage_stage(context: JobContext) -> None:
             if context.dry_run:
-                return
+                return StageArtifacts()
             plan = read(context, "plan.json")
             manifest = CorpusManifest.model_validate(read(context, "manifest.json"))
+            recovered = False
             try:
-                self.store.stage(
-                    manifest, {doc.document_id: doc.content for doc in manifest.documents},
-                    generation=plan["generation"],
-                )
+                if artifact(context, "compensated.json").exists():
+                    recovered = True
+                    rotate_compensated_generation(context)
+                    values = records(context)
+                    existing = self.vector_store.existing_hashes("evidence", ["evidence"])
+                    missing = [value for value in values if existing.get(value.record.id) != value.content_hash]
+                    if missing and self.vector_store.upsert("evidence", missing) != len(missing):
+                        raise PipelineError("vector write count mismatch")
+                    plan = read(context, "plan.json")
+                    manifest = CorpusManifest.model_validate(read(context, "manifest.json"))
+                if not self.store.generation_path(plan["generation"]).exists():
+                    self.store.stage(
+                        manifest, {doc.document_id: doc.content for doc in manifest.documents},
+                        generation=plan["generation"],
+                    )
                 self.store.manifest(plan["generation"])
             except Exception:
                 compensate(context)
                 raise
+            return StageArtifacts(
+                ("plan.json", "manifest.json", "vector-intent.json") if recovered else ()
+            )
 
         def publish_stage(context: JobContext) -> None:
             if context.dry_run:
-                return
+                return StageArtifacts()
             if artifact(context, "compensated.json").exists():
                 rotate_compensated_generation(context)
                 values = records(context)
-                if values and self.vector_store.upsert("evidence", values) != len(values):
+                try:
+                    existing = self.vector_store.existing_hashes("evidence", ["evidence"])
+                    missing = [value for value in values if existing.get(value.record.id) != value.content_hash]
+                    if missing and self.vector_store.upsert("evidence", missing) != len(missing):
+                        raise PipelineError("vector write count mismatch")
+                except Exception:
                     compensate(context)
-                    raise PipelineError("vector write count mismatch")
+                    raise
                 manifest = CorpusManifest.model_validate(read(context, "manifest.json"))
                 generation = read(context, "plan.json")["generation"]
-                self.store.stage(
-                    manifest, {doc.document_id: doc.content for doc in manifest.documents},
-                    generation=generation,
-                )
+                try:
+                    if not self.store.generation_path(generation).exists():
+                        self.store.stage(
+                            manifest, {doc.document_id: doc.content for doc in manifest.documents},
+                            generation=generation,
+                        )
+                except Exception:
+                    compensate(context)
+                    raise
             generation = read(context, "plan.json")["generation"]
             try:
                 self.store.publish(generation)
             except Exception:
                 compensate(context)
                 raise
+            return StageArtifacts(("plan.json", "manifest.json", "vector-intent.json"))
 
         def retention_stage(context: JobContext) -> None:
             # Retention policy is intentionally a stable no-op until configured.
@@ -305,7 +349,7 @@ class CorpusPipeline:
         report = run_job(spec, [
             discover_stage, acquire_stage, embed_stage, vector_stage,
             stage_stage, publish_stage, retention_stage,
-        ])
+        ], after_stage_return=after_stage_return)
         run_dir = workspace_root / ".tht-jobs" / "evidence" / "runs" / report.run_id
         plan = json.loads((run_dir / "artifacts" / "plan.json").read_text())
         if dry_run:

harness/tht/corpus/store.py

@@ -46,6 +46,7 @@ class CorpusStore:
         self.root = Path(root)
         self.active_path = self.root / "ACTIVE"
         self._replace = os.replace
+        self._fsync_directory = self._sync_root
         self._ensure_root()
 
     def _ensure_root(self) -> None:
@@ -110,15 +111,35 @@ class CorpusStore:
         manifest = self.manifest(generation)
         if manifest.manifest_id != generation:
             raise UnsafeCorpusPath("manifest generation mismatch")
+        if self.active_generation() == generation:
+            return generation
+        previous = self.active_generation()
         temporary = self.active_path.with_name(f".ACTIVE.{uuid.uuid4().hex}.tmp")
-        _atomic_write(temporary, (generation + "\n").encode())
-        self._replace(temporary, self.active_path)
+        replaced = False
+        try:
+            _atomic_write(temporary, (generation + "\n").encode())
+            self._replace(temporary, self.active_path)
+            replaced = True
+            self._fsync_directory()
+        except BaseException:
+            temporary.unlink(missing_ok=True)
+            if replaced:
+                if previous is None:
+                    self.active_path.unlink(missing_ok=True)
+                else:
+                    rollback = self.active_path.with_name(f".ACTIVE.rollback.{uuid.uuid4().hex}.tmp")
+                    _atomic_write(rollback, (previous + "\n").encode())
+                    self._replace(rollback, self.active_path)
+                self._sync_root()
+            raise
+        return generation
+
+    def _sync_root(self) -> None:
         directory = os.open(self.root, os.O_RDONLY | os.O_DIRECTORY | os.O_NOFOLLOW)
         try:
             os.fsync(directory)
         finally:
             os.close(directory)
-        return generation
 
     def active_generation(self) -> str | None:
         try:

harness/tht/jobs/runner.py

@@ -23,6 +23,11 @@ class CorruptCheckpointError(RuntimeError):
     pass
 
 
+@dataclass(frozen=True)
+class StageArtifacts:
+    required: tuple[str, ...] = ()
+
+
 @dataclass(frozen=True)
 class JobContext:
     run_id: str
@@ -35,6 +40,84 @@ class JobContext:
 Stage = Callable[[JobContext], Any]
 
 
+def _artifact_digest(path: Path) -> dict[str, Any]:
+    payload = path.read_bytes()
+    return {"sha256": hashlib.sha256(payload).hexdigest(), "size": len(payload)}
+
+
+def _seal_artifacts(context: JobContext, stage: str, result: Any, spec: JobSpec) -> None:
+    required = result.required if isinstance(result, StageArtifacts) else ()
+    root = context.run_dir / "artifacts"
+    root.mkdir(exist_ok=True)
+    manifest_path = root / "artifact-manifest.json"
+    manifest = json.loads(manifest_path.read_text()) if manifest_path.exists() else {
+        "schema_version": 1,
+        "spec_fingerprint": _compatibility_fingerprint(spec, list(spec.stage_ids)),
+        "stages": {},
+    }
+    files = {}
+    for relative in required:
+        candidate = root / relative
+        if Path(relative).is_absolute() or ".." in Path(relative).parts or candidate.is_symlink():
+            raise CorruptCheckpointError("artifact path is unsafe")
+        if not candidate.is_file():
+            raise CorruptCheckpointError("required stage artifact is missing")
+        files[relative] = _artifact_digest(candidate)
+        for prior in manifest["stages"].values():
+            if relative in prior.get("required", []):
+                prior["required"].remove(relative)
+                prior["files"].pop(relative, None)
+    manifest["stages"][stage] = {"required": list(required), "files": files}
+    _atomic_write(manifest_path, json.dumps(manifest, sort_keys=True, separators=(",", ":")) + "\n")
+
+
+def seal_stage_artifacts(
+    context: JobContext, stage: str, required: tuple[str, ...], spec: JobSpec,
+) -> None:
+    """Durably record external-effect intent before a stage performs that effect."""
+    _seal_artifacts(context, stage, StageArtifacts(required), spec)
+
+
+def _validate_artifacts(run_dir: Path, spec: JobSpec, source: JobRun) -> set[str]:
+    root = run_dir / "artifacts"
+    manifest_path = root / "artifact-manifest.json"
+    successful = {stage.name for stage in source.stages if stage.status == "succeeded"}
+    if not successful and not manifest_path.exists():
+        return set()
+    try:
+        manifest = json.loads(manifest_path.read_text())
+        if manifest["spec_fingerprint"] != _compatibility_fingerprint(spec, list(spec.stage_ids)):
+            raise ValueError
+        sealed = set(manifest["stages"])
+        allowed = {"artifact-manifest.json"}
+        for stage, record in manifest["stages"].items():
+            for relative in record["required"]:
+                candidate = root / relative
+                if Path(relative).is_absolute() or ".." in Path(relative).parts or candidate.is_symlink():
+                    raise ValueError
+                if not candidate.is_file() or _artifact_digest(candidate) != record["files"][relative]:
+                    raise ValueError
+                allowed.add(relative)
+        if not successful.issubset(sealed):
+            raise ValueError
+        entries = list(root.iterdir())
+        if any(path.is_symlink() or not path.is_file() for path in entries):
+            raise ValueError
+        actual = {path.name for path in entries}
+        incomplete = next((stage for stage in source.stages if stage.status != "succeeded"), None)
+        marker = root / "compensated.json"
+        if incomplete is not None and incomplete.status in {"failed", "running"} and marker.is_file():
+            payload = json.loads(marker.read_text())
+            if not isinstance(payload.get("generation"), str):
+                raise ValueError
+            allowed.add("compensated.json")
+        if actual != allowed:
+            raise ValueError
+        return sealed
+    except (OSError, KeyError, TypeError, ValueError, json.JSONDecodeError) as error:
+        raise CorruptCheckpointError("resume artifact manifest is invalid") from error
+
+
 def _atomic_write(path: Path, payload: str) -> None:
     temporary = path.with_name(f".{path.name}.{uuid.uuid4().hex}.tmp")
     fd = os.open(temporary, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
@@ -91,7 +174,10 @@ def _new_run(spec: JobSpec, run_id: str, stages: Sequence[Stage]) -> JobRun:
     )
 
 
-def _resume_run(spec: JobSpec, run_id: str, stages: Sequence[Stage], source: JobRun) -> JobRun:
+def _resume_run(
+    spec: JobSpec, run_id: str, stages: Sequence[Stage], source: JobRun,
+    effect_completed: set[str] | None = None,
+) -> JobRun:
     requested_names = list(spec.stage_ids)
     if len(requested_names) != len(stages) or len(requested_names) != len(set(requested_names)):
         raise CorruptCheckpointError("resume checkpoint is incompatible with requested stages")
@@ -107,11 +193,15 @@ def _resume_run(spec: JobSpec, run_id: str, stages: Sequence[Stage], source: Job
     resumed_stages = []
     for name in requested_names:
         previous = source_by_name.get(name)
-        resumed_stages.append(
-            previous
-            if previous is not None and previous.status == "succeeded"
-            else StageRun(name=name)
-        )
+        if previous is not None and previous.status == "succeeded":
+            resumed_stages.append(previous)
+        elif previous is not None and previous.status == "running" and name in (effect_completed or set()):
+            resumed_stages.append(StageRun(
+                name=name, status="succeeded", started_at=previous.started_at or utc_now(),
+                finished_at=utc_now(),
+            ))
+        else:
+            resumed_stages.append(StageRun(name=name))
     return JobRun(
         run_id=run_id,
         compatibility_fingerprint=source.compatibility_fingerprint,
@@ -129,7 +219,10 @@ def _resume_run(spec: JobSpec, run_id: str, stages: Sequence[Stage], source: Job
     )
 
 
-def run_job(spec: JobSpec, stages: Sequence[Stage]) -> JobReport:
+def run_job(
+    spec: JobSpec, stages: Sequence[Stage], *,
+    after_stage_return: Callable[[JobContext, str], Any] | None = None,
+) -> JobReport:
     """Run stages once, returning a terminal report instead of leaking stage exceptions."""
     with WorkspaceJobLock(spec.workspace_root, spec.workspace_id, spec.job_type):
         jobs_root = spec.workspace_root / ".tht-jobs" / spec.job_type / "runs"
@@ -147,6 +240,7 @@ def run_job(spec: JobSpec, stages: Sequence[Stage]) -> JobReport:
                     source_path = matches[0]
             source = _load_checkpoint(source_path)
             _validate_resume_source(spec, stages, source)
+            effect_completed = _validate_artifacts(source_path.parent, spec, source)
 
         run_id = uuid.uuid4().hex
         run_dir = jobs_root / run_id
@@ -155,7 +249,7 @@ def run_job(spec: JobSpec, stages: Sequence[Stage]) -> JobReport:
         if source is None:
             run = _new_run(spec, run_id, stages)
         else:
-            run = _resume_run(spec, run_id, stages, source)
+            run = _resume_run(spec, run_id, stages, source, effect_completed)
             source_artifacts = jobs_root / source.run_id / "artifacts"
             if source_artifacts.exists():
                 shutil.copytree(source_artifacts, run_dir / "artifacts")
@@ -173,7 +267,7 @@ def run_job(spec: JobSpec, stages: Sequence[Stage]) -> JobReport:
             )
             _persist(checkpoint_path, run)
             try:
-                stage_callable(context)
+                stage_result = stage_callable(context)
             except Exception:
                 failed = stage.model_copy(
                     update={
@@ -191,6 +285,9 @@ def run_job(spec: JobSpec, stages: Sequence[Stage]) -> JobReport:
                 )
                 _persist(checkpoint_path, run)
                 break
+            _seal_artifacts(context, stage.name, stage_result, spec)
+            if after_stage_return is not None:
+                after_stage_return(context, stage.name)
             succeeded = stage.model_copy(update={"status": "succeeded", "finished_at": utc_now()})
             run = run.model_copy(
                 update={"stages": run.stages[:index] + (succeeded,) + run.stages[index + 1 :]}