fix(evidence): isolate vectors by workspace

harness/tests/test_corpus_pipeline.py

@@ -1,7 +1,7 @@
 import pytest
 
 from tht.corpus.chunk import ChunkPolicy
-from tht.corpus.pipeline import CorpusPipeline, PipelineError
+from tht.corpus.pipeline import CorpusPipeline, PipelineError, PipelineResult
 from tht.corpus.store import CorpusStore
 from tht.corpus.models import CorpusManifest
 from tht.ports.evidence import AcquiredDocument, SourceObject
@@ -55,17 +55,19 @@ class Vectors:
             value.record.id: value.content_hash for value in self.records
         }
 
-    def delete_generation(self, collection, generation):
+    def delete_generation(self, collection, generation, workspace_id):
         self.records = [
             value for value in self.records
-            if value.record.metadata["vector_generation"] != generation
+            if not (value.record.metadata["vector_generation"] == generation
+                    and value.record.metadata.get("workspace_id") == workspace_id)
         ]
         return 0
 
-    def list_evidence_generations(self, collection):
+    def list_evidence_generations(self, collection, workspace_id):
         return sorted({
             value.record.metadata["vector_generation"] for value in self.records
             if value.record.kind == "evidence"
+            and value.record.metadata.get("workspace_id") == workspace_id
         })
 
 
@@ -128,10 +130,10 @@ def test_retention_keeps_filesystem_when_vector_purge_fails_then_retries(tmp_pat
             super().__init__()
             self.fail_delete = True
 
-        def delete_generation(self, collection, generation):
+        def delete_generation(self, collection, generation, workspace_id):
             if self.fail_delete:
                 raise RuntimeError("credential secret")
-            return super().delete_generation(collection, generation)
+            return super().delete_generation(collection, generation, workspace_id)
 
     vectors = FailingDelete()
     for index in range(2):
@@ -156,13 +158,13 @@ def test_gc_reconciles_vector_only_generation(tmp_path):
     from tht.vectorstore.records import VectorRecord
     vectors.records.append(VectorWriteRecord(
         record=VectorRecord(id="orphan", kind="evidence", ref="doc:x", title="", content="x",
-                            metadata={"vector_generation": orphan}),
+                            metadata={"vector_generation": orphan, "workspace_id": "default"}),
         embedding=[0.0, 0.0, 0.0], content_hash="sha256:" + "0" * 64,
     ))
     candidate = pipeline(tmp_path, Source([]), vectors=vectors, retain=1)
     report = candidate.gc(workspace_root=tmp_path)
     assert report["evicted"] == [orphan]
-    assert vectors.list_evidence_generations("evidence") == []
+    assert vectors.list_evidence_generations("evidence", "default") == []
     assert candidate.gc(workspace_root=tmp_path)["evicted"] == []
 
 
@@ -233,7 +235,7 @@ def test_gc_preserves_vector_dependencies_of_retained_manifests(tmp_path):
         vectors=vectors, retain=2,
     ).run().generation
     assert CorpusStore(tmp_path / "corpus").list_generations() == [second, third]
-    assert first in vectors.list_evidence_generations("evidence")
+    assert first in vectors.list_evidence_generations("evidence", "default")
 
 
 def test_active_searcher_without_active_fails_closed_for_evidence(tmp_path):
@@ -319,6 +321,14 @@ def test_active_evidence_query_holds_lock_against_publish(tmp_path):
     assert published.is_set()
 
 
+def test_pipeline_result_dump_does_not_deepcopy_frozen_metadata():
+    manifest = CorpusManifest(metadata={"nested": {"value": ["safe"]}})
+    payload = PipelineResult(
+        "succeeded", None, False, (), (), (), manifest,
+    ).model_dump(mode="json")
+    assert payload["manifest"]["metadata"] == {"nested": {"value": ["safe"]}}
+
+
 def test_unchanged_documents_skip_acquire_normalize_chunk_and_embed(tmp_path):
     one = item("one", "a")
     first_source = Source([(one, "hello")])