feat(vector): add direct pgvector adapter

harness/tht/config.py

@@ -18,6 +18,7 @@ class ConfigError(Exception):
 
 def _expand_env(value: Any) -> Any:
     if isinstance(value, str):
+
         def repl(m: re.Match) -> str:
             var = m.group(1)
             if var not in os.environ:
@@ -85,7 +86,16 @@ DwhResourceConfig = Annotated[
 
 class PgvectorDirectConfig(BaseModel):
     type: Literal["pgvector_direct"]
-    connection: DatabaseConfig
+    reader: DatabaseConfig | None = None
+    writer: DatabaseConfig | None = None
+    # Deprecated compatibility: a single direct connection historically meant read-only.
+    connection: DatabaseConfig | None = None
+
+    @model_validator(mode="after")
+    def validate_connections(self):
+        if self.reader is None and self.writer is None and self.connection is None:
+            raise ValueError("pgvector_direct requires a reader or writer connection")
+        return self
 
 
 class ThothVectorHttpConfig(BaseModel):
@@ -134,9 +144,9 @@ class LshConfig(BaseModel):
 
 class EligibilityConfig(BaseModel):
     # Soglie del principio di column eligibility (testo ampio ignorato ovunque).
-    max_declared_len: int = 128   # char/varchar dichiarati <= soglia: eligible senza campionare
-    max_avg_length: int = 40      # fallback data-driven: lunghezza media valori campionati
-    max_sampled_len: int = 200    # fallback data-driven: lunghezza massima valore campionato
+    max_declared_len: int = 128  # char/varchar dichiarati <= soglia: eligible senza campionare
+    max_avg_length: int = 40  # fallback data-driven: lunghezza media valori campionati
+    max_sampled_len: int = 200  # fallback data-driven: lunghezza massima valore campionato
     # Colonne di servizio sempre ignorate per nome (match case-insensitive), a prescindere
     # dal tipo: metadati ETL/audit non analitici (es. timestamp di ultimo aggiornamento).
     ignore_columns: list[str] = ["etl_last_update"]
@@ -168,7 +178,7 @@ class VectorConfig(BaseModel):
 
 class SearchConfig(BaseModel):
     rrf_k: int = 60
-    top_schema_tables: int = 12   # default `--top` per `tht search --kind schema` (n. tabelle)
+    top_schema_tables: int = 12  # default `--top` per `tht search --kind schema` (n. tabelle)
     schema_chunk_pool: int = 150  # chunk tabella/colonna fusi prima dell'aggregazione a tabella
 
 
@@ -181,9 +191,17 @@ class ExecutionConfig(BaseModel):
     max_aggregate_cells: int = 20
     max_export_rows: int = 100000
     forbidden_functions: list[str] = [
-        "setval", "nextval", "pg_advisory_lock", "pg_advisory_xact_lock",
-        "dblink", "dblink_exec", "pg_terminate_backend", "pg_cancel_backend",
-        "lo_import", "lo_export", "pg_reload_conf",
+        "setval",
+        "nextval",
+        "pg_advisory_lock",
+        "pg_advisory_xact_lock",
+        "dblink",
+        "dblink_exec",
+        "pg_terminate_backend",
+        "pg_cancel_backend",
+        "lo_import",
+        "lo_export",
+        "pg_reload_conf",
     ]
 
 
@@ -302,7 +320,10 @@ def _populate_legacy_views(raw: dict[str, Any]) -> None:
     vectors = raw.get("vectors")
     if isinstance(vectors, dict):
         if vectors.get("type") == "pgvector_direct":
-            raw.setdefault("vector_db", vectors["connection"])
+            raw.setdefault(
+                "vector_db",
+                vectors.get("writer") or vectors.get("reader") or vectors.get("connection"),
+            )
         elif vectors.get("type") == "thoth_vector_http":
             raw.setdefault("vector_rest", vectors.get("reader"))
             raw.setdefault("vector_write_rest", vectors.get("writer"))