fix(evidence): harden unchanged job snapshot

harness/tests/test_corpus_pipeline.py

@@ -1,3 +1,5 @@
+from datetime import UTC, datetime, timedelta
+
 import pytest
 
 from tht.corpus.chunk import ChunkPolicy
@@ -432,12 +434,27 @@ def test_unchanged_documents_skip_acquire_normalize_chunk_and_embed(tmp_path):
 
 
 def test_unchanged_job_reuses_active_generation_without_new_directory(tmp_path):
-    source = Source([(item("one", "a"), "hello")])
+    source = Source([(
+        SourceObject(
+            source_id="fs:one", uri="file:///safe/one.md", fingerprint="sha256:a",
+            modified_at=datetime(2026, 1, 1, tzinfo=UTC),
+            metadata={"media_type": "text/markdown", "size": 5, "nested": {"b": 2, "a": 1}},
+        ),
+        "hello",
+    )])
     candidate = pipeline(tmp_path, source)
     args = dict(workspace_id="demo", workspace_root=tmp_path,
                 config_fingerprint="sha256:" + "1" * 64,
                 input_fingerprint="sha256:" + "2" * 64)
     first = candidate.run_as_job(**args)
+    snapshot = first.manifest.metadata["source_snapshot"]["fs:one"]
+    assert snapshot == {
+        "source_id": "fs:one", "uri": "file:///safe/one.md", "fingerprint": "sha256:a",
+        "modified_at": "2026-01-01T00:00:00Z",
+        "metadata": {"media_type": "text/markdown", "size": 5,
+                     "nested": {"a": 1, "b": 2}},
+        "media_type": "text/markdown", "size": 5,
+    }
     count = len(candidate.store.list_generations())
     second = candidate.run_as_job(**args)
     assert second.generation == first.generation
@@ -445,6 +462,82 @@ def test_unchanged_job_reuses_active_generation_without_new_directory(tmp_path):
     assert len(candidate.store.list_generations()) == count
 
 
+@pytest.mark.parametrize("field", ["uri", "modified_at", "metadata"])
+def test_job_source_snapshot_change_forces_publish_with_same_fingerprint(tmp_path, field):
+    original = SourceObject(
+        source_id="fs:one", uri="file:///safe/one.md", fingerprint="sha256:a",
+        modified_at=datetime(2026, 1, 1, tzinfo=UTC),
+        metadata={"media_type": "text/markdown", "size": 5, "label": "original"},
+    )
+    vectors = Vectors()
+    args = dict(workspace_id="demo", workspace_root=tmp_path,
+                config_fingerprint="sha256:" + "1" * 64,
+                input_fingerprint="sha256:" + "2" * 64)
+    first = pipeline(tmp_path, Source([(original, "hello")]), vectors=vectors).run_as_job(**args)
+    updates = {
+        "uri": "file:///safe/renamed.md",
+        "modified_at": original.modified_at + timedelta(seconds=1),
+        "metadata": {"media_type": "text/markdown", "size": 5, "label": "changed"},
+    }
+    changed = original.model_copy(update={field: updates[field]})
+    source = Source([(changed, "hello")])
+    result = pipeline(tmp_path, source, vectors=vectors).run_as_job(**args)
+    assert result.published is True
+    assert result.generation != first.generation
+    assert source.acquire_calls == ["fs:one"]
+
+
+@pytest.mark.parametrize("fingerprint_name", ["config_fingerprint", "input_fingerprint"])
+def test_job_binding_change_forces_publish(tmp_path, fingerprint_name):
+    vectors = Vectors()
+    source_object = item("one", "a")
+    args = dict(workspace_id="demo", workspace_root=tmp_path,
+                config_fingerprint="sha256:" + "1" * 64,
+                input_fingerprint="sha256:" + "2" * 64)
+    first = pipeline(tmp_path, Source([(source_object, "hello")]), vectors=vectors).run_as_job(**args)
+    args[fingerprint_name] = "sha256:" + "3" * 64
+    source = Source([(source_object, "hello")])
+    result = pipeline(tmp_path, source, vectors=vectors).run_as_job(**args)
+    assert result.published is True
+    assert result.generation != first.generation
+    assert source.acquire_calls == []
+
+
+@pytest.mark.parametrize(
+    "damage", ["legacy_metadata", "corrupt_document_sources", "document", "vector"]
+)
+def test_job_incomplete_active_contract_never_noops(tmp_path, damage):
+    import json
+
+    vectors = Vectors()
+    source_object = item("one", "a")
+    args = dict(workspace_id="demo", workspace_root=tmp_path,
+                config_fingerprint="sha256:" + "1" * 64,
+                input_fingerprint="sha256:" + "2" * 64)
+    candidate = pipeline(tmp_path, Source([(source_object, "hello")]), vectors=vectors)
+    first = candidate.run_as_job(**args)
+    if damage == "legacy_metadata":
+        manifest_path = candidate.store.generation_path(first.generation) / "manifest.json"
+        payload = json.loads(manifest_path.read_text())
+        payload["metadata"].pop("source_snapshot")
+        manifest_path.write_text(json.dumps(payload))
+    elif damage == "corrupt_document_sources":
+        manifest_path = candidate.store.generation_path(first.generation) / "manifest.json"
+        payload = json.loads(manifest_path.read_text())
+        payload["metadata"]["document_sources"] = {}
+        manifest_path.write_text(json.dumps(payload))
+    elif damage == "document":
+        path = candidate.store.resolve_document(first.manifest.documents[0].document_id)
+        path.unlink()
+    else:
+        vectors.records.clear()
+    source = Source([(source_object, "hello")])
+    result = pipeline(tmp_path, source, vectors=vectors).run_as_job(**args)
+    assert result.published is True
+    assert result.generation != first.generation
+    assert source.acquire_calls == ["fs:one"]
+
+
 def test_removed_documents_are_marked_and_absent_from_new_manifest(tmp_path):
     one, two = item("one", "a"), item("two", "b")
     pipeline(tmp_path, Source([(one, "one"), (two, "two")])).run()