fix: harden reviewer workflow and memory handling

harness/tests/l2/test_memory_save_one_real.py

@@ -34,9 +34,10 @@ def test_save_one_upserts_to_real_pgvector(l2_env):
 
     record = MemoryRecord(
         id="mem-l2test", ts=datetime.now(), session_id="l2-self-test",
-        decision_seq=999, type="table_promoted", subject="fct_ricoveri",
-        detail="ablazione", rationale="L2 self-test (idempotent)",
-        question_context="ablazione 2025", tables=["fct_ricoveri"], concepts=[],
+        decision_seq=999, type="concept_clarified", subject="ablazione recente",
+        detail="evento di ablazione negli ultimi 15 anni",
+        rationale="L2 self-test (idempotent)",
+        question_context="ablazione 2025", tables=[], concepts=["ablazione recente"],
     )
     from tht.adapters.vector import ThothHttpVectorStore
 

harness/tests/test_adapter_command_regressions.py

@@ -68,9 +68,11 @@ def test_memory_command_writes_through_factory_vector_store(monkeypatch):
     cfg = SimpleNamespace(profile="server", embeddings=object(), vector_write_rest=None)
     manifest = SimpleNamespace(id="s1")
     snapshot = SimpleNamespace(manifest=manifest, decisions=[], artifacts={})
-    record = MemoryRecord(id="m1", ts=datetime(2026, 1, 1), session_id="s1",
-                          decision_seq=7, type="table_promoted", subject="t",
-                          question_context="q")
+    record = MemoryRecord(
+        id="m1", ts=datetime(2026, 1, 1), session_id="s1",
+        decision_seq=7, type="concept_clarified", subject="paziente attivo",
+        detail="flag_attivo = TRUE", question_context="q",
+    )
     monkeypatch.setattr(memory_cmd, "_load_config_or_exit", lambda path: cfg)
     monkeypatch.setattr(memory_cmd, "load_snapshot_or_exit", lambda cfg, session: snapshot)
     monkeypatch.setattr(memory_cmd, "registry_path", lambda cfg: None)

harness/tests/test_decision_min_phase.py

@@ -16,7 +16,7 @@ from tht.workflow import load_workflow
         ("concept_clarified", 1),
         ("memory_rejected", 2),
         ("question_rewritten", 3),
-        ("table_promoted", 2),
+        ("table_promoted", 4),
         ("column_corrected", 4),
         ("evidence_accepted", 4),
         ("value_grounded", 4),

harness/tests/test_memory_metadata.py

@@ -13,9 +13,9 @@ from tht.memory import MemoryRecord, memory_vector_records
 def _record(**kw) -> MemoryRecord:
     base = dict(
         id="mem-x", ts=datetime(2025, 1, 1), session_id="s", decision_seq=1,
-        type="table_promoted", subject="dim_pazienti", detail="promossa",
+        type="concept_clarified", subject="paziente attivo", detail="flag_attivo = TRUE",
         rationale="perche' serve", question_context="dammi pazienti",
-        tables=["t"], concepts=[],
+        tables=[], concepts=["paziente attivo"],
     )
     base.update(kw)
     return MemoryRecord(**base)
@@ -23,17 +23,17 @@ def _record(**kw) -> MemoryRecord:
 
 def test_memory_vector_record_has_subject_detail_rationale_in_metadata():
     vr = memory_vector_records([_record()])[0]
-    assert vr.metadata["subject"] == "dim_pazienti"
-    assert vr.metadata["detail"] == "promossa"
+    assert vr.metadata["subject"] == "paziente attivo"
+    assert vr.metadata["detail"] == "flag_attivo = TRUE"
     assert vr.metadata["rationale"] == "perche' serve"
 
 
 def test_memory_vector_record_metadata_keeps_existing_fields():
     vr = memory_vector_records([_record()])[0]
     # i campi che gia' c'erano restano (backward compat)
-    assert vr.metadata["type"] == "table_promoted"
-    assert vr.metadata["tables"] == ["t"]
-    assert vr.metadata["concepts"] == []
+    assert vr.metadata["type"] == "concept_clarified"
+    assert vr.metadata["tables"] == []
+    assert vr.metadata["concepts"] == ["paziente attivo"]
     assert vr.metadata["session_id"] == "s"
 
 
@@ -62,6 +62,6 @@ def test_save_one_memory_preserves_subject_through_upsert_row():
     save_one_memory([_record(decision_seq=1)], decision_seq=1, store=writer, embedder=embedder)
     row = writer.upsert.call_args[0][1][0]
     md = row.record.metadata
-    assert md["subject"] == "dim_pazienti"
-    assert md["detail"] == "promossa"
+    assert md["subject"] == "paziente attivo"
+    assert md["detail"] == "flag_attivo = TRUE"
     assert md["rationale"] == "perche' serve"

harness/tests/test_memory_promotion.py

@@ -9,7 +9,13 @@ la rende un passo del workflow. Questi test fissano il contratto harness-side:
 from datetime import datetime
 
 from tht.decisions import DecisionRecord, append_decision
-from tht.memory import declined_promotion_seqs, reusable_promotions
+from tht.memory import (
+    MemoryRecord,
+    declined_promotion_seqs,
+    memory_vector_records,
+    promote,
+    reusable_promotions,
+)
 from tht.session.models import SessionManifest
 from tht.workflow import load_workflow
 
@@ -62,3 +68,48 @@ def test_reusable_promotions_exclude_declined(tmp_path):
                     subject="fact_a", detail="seq:1")                         # seq 3
     cand = reusable_promotions(tmp_path, _manifest(), tmp_path / "registry.jsonl")
     assert [c.decision_seq for c in cand] == [2]
+
+
+def test_reusable_promotions_deduplicate_identical_memory_content(tmp_path):
+    append_decision(tmp_path, type="concept_clarified", subject="paziente attivo",
+                    detail="flag_attivo = TRUE", rationale="scelta reviewer")
+    append_decision(tmp_path, type="concept_clarified", subject="paziente attivo",
+                    detail="flag_attivo = TRUE", rationale="scelta reviewer")
+
+    cand = reusable_promotions(tmp_path, _manifest(), tmp_path / "registry.jsonl")
+
+    assert [c.decision_seq for c in cand] == [1]
+
+
+def test_only_concept_clarified_is_proposed_or_promoted(tmp_path):
+    append_decision(tmp_path, type="table_promoted", subject="fact_a",
+                    detail="tabella principale")
+    append_decision(tmp_path, type="table_excluded", subject="fact_b",
+                    detail="tabella non pertinente")
+    append_decision(tmp_path, type="concept_clarified", subject="paziente attivo",
+                    detail="flag_attivo = TRUE")
+    registry = tmp_path / "registry.jsonl"
+
+    candidates = reusable_promotions(tmp_path, _manifest(), registry)
+    promoted = promote(tmp_path, _manifest(), seqs=[1, 2, 3], registry_path=registry)
+
+    assert [(c.decision_seq, c.type) for c in candidates] == [(3, "concept_clarified")]
+    assert [(c.decision_seq, c.type) for c in promoted] == [(3, "concept_clarified")]
+
+
+def test_legacy_table_records_are_not_published_as_memory_vectors():
+    records = [
+        MemoryRecord(
+            id="mem-0001", ts=datetime(2026, 1, 1), session_id="s1",
+            decision_seq=1, type="table_promoted", subject="fact_a",
+        ),
+        MemoryRecord(
+            id="mem-0002", ts=datetime(2026, 1, 1), session_id="s1",
+            decision_seq=2, type="concept_clarified", subject="paziente attivo",
+            detail="flag_attivo = TRUE",
+        ),
+    ]
+
+    vectors = memory_vector_records(records)
+
+    assert [record.ref for record in vectors] == ["mem-0002"]

harness/tests/test_memory_save_one.py

@@ -16,8 +16,9 @@ from tht.memory import MemoryRecord, memory_vector_record_for_decision, save_one
 def _record(seq: int = 7, **kw) -> MemoryRecord:
     base = dict(
         id="mem-0007", ts=datetime(2025, 1, 1), session_id="s1", decision_seq=seq,
-        type="table_promoted", subject="pazienti", detail="promossa", rationale="r",
-        question_context="dammi i pazienti", tables=["pazienti"], concepts=[],
+        type="concept_clarified", subject="paziente attivo",
+        detail="flag_attivo = TRUE", rationale="r",
+        question_context="dammi i pazienti", tables=[], concepts=["paziente attivo"],
     )
     base.update(kw)
     return MemoryRecord(**base)

harness/tests/test_psd_local_compose_contract.py

@@ -54,6 +54,10 @@ def test_psd_overlay_uses_generated_workspace_for_default_and_named_commands():
     assert core["networks"]["default"]["aliases"] == ["core", "thothii-core"]
     frontend = compose["services"]["frontend"]
     assert frontend["ports"] == ["127.0.0.1:8099:8080"]
+    assert frontend["build"]["args"] == {
+        "VITE_BASE": "/",
+        "VITE_BACKEND_URL": "/api",
+    }
     assert frontend["networks"] == {
         "default": {"aliases": ["frontend", "thothii-frontend"]}
     }

harness/tests/test_solved_search_cli.py

@@ -7,10 +7,12 @@ puro (`[]` in modalita' --json) ed exit 0, cosi' il modello prosegue senza
 exemplar. Il finalize-hook gestisce gia' lo stesso scenario in modo analogo.
 """
 import json
+from datetime import datetime
 
 from typer.testing import CliRunner
 
 from tht.cli import app
+from tht.memory import MemoryRecord, save_registry
 from tht.ports.vector import VectorReadUnavailable
 from tht.vectorstore.rest_client import VectorRestError
 from tht.vectorstore.store import VectorHit
@@ -97,3 +99,44 @@ def test_solved_search_json_maps_hit_metadata(tmp_path, monkeypatch):
         "session_id": "s1", "question": "quante ablazioni nel 2023",
         "sql": "SELECT 1", "tables": ["fact_seeablazione"], "score": 0.91,
     }]
+
+
+def test_memory_search_excludes_legacy_table_records(tmp_path, monkeypatch):
+    records = [
+        MemoryRecord(
+            id="mem-0001", ts=datetime(2026, 1, 1), session_id="s1",
+            decision_seq=1, type="table_promoted", subject="fact_pazienti",
+        ),
+        MemoryRecord(
+            id="mem-0002", ts=datetime(2026, 1, 1), session_id="s1",
+            decision_seq=2, type="concept_clarified", subject="paziente attivo",
+            detail="flag_attivo = TRUE",
+        ),
+    ]
+    cfg = _cfg(tmp_path)
+    save_registry(records, tmp_path / "a" / "memory" / "registry.jsonl")
+
+    class FakeSearcher:
+        def search(self, vec, top_n=10, kinds=None):
+            return [
+                VectorHit(
+                    id=f"memory:{record.id}", kind="memory", ref=record.id,
+                    title=record.subject, content=record.detail, metadata={},
+                    similarity=0.9,
+                )
+                for record in records
+            ]
+
+    class FakeEmbedder:
+        def embed_query(self, text):
+            return [0.1] * 8
+
+    monkeypatch.setattr("tht.cli.vector_cmd.open_searcher", lambda workspace: FakeSearcher())
+    monkeypatch.setattr("tht.cli.vector_cmd.make_embedder", lambda embeddings: FakeEmbedder())
+
+    res = CliRunner().invoke(
+        app, ["memory", "search", "pazienti", "--json", "-c", str(cfg)]
+    )
+
+    assert res.exit_code == 0, res.output
+    assert [record["id"] for record in json.loads(res.stdout)] == ["mem-0002"]