Fix ai-service Dockerfile: bake in drug_entities.json, override its path

apps/ai-service/adapters/postgres.py

@@ -93,3 +93,59 @@ class PostgresTraceRepository:
             decision=row[4], reason=row[5], resolved_drug_id=row[6],
             citations=tuple(row[7]), created_at=row[8],
         )
+
+
+class PostgresConversationStore:
+    """Durable, cross-worker replacement for `RagAgent`'s in-process
+    `dict[str, list[str]]` history — the F-08 gap named in ADR 0008: history
+    was lost on restart and not shared across workers.
+
+    Append-only, one connection per call — the same tradeoffs as
+    `PostgresTraceRepository` (F-09: real pooling is a further improvement
+    not made here) and the same `connect_timeout=5` for the same reason
+    (an unreachable-but-not-refusing host hangs on the OS TCP timeout
+    otherwise, defeating a caller's fail-open try/except just as completely
+    as no try/except at all).
+
+    Raises on any error rather than swallowing it — `RagAgent` is the
+    caller that decides fail-open (lose this turn's memory, not the
+    response), matching how `routers/rag.py` already wraps
+    `PostgresTraceRepository.save()` rather than the repository hiding its
+    own failures.
+    """
+
+    def __init__(self, dsn: str) -> None:
+        self._dsn = dsn
+
+    def migrate(self, migration_path: Path) -> None:
+        import psycopg
+
+        statement = migration_path.read_text(encoding="utf-8")
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
+            connection.execute(statement)
+
+    def recent(self, conversation_id: str, limit: int) -> list[str]:
+        """The last `limit` lines, oldest first — matches the ordering the
+        understanding LLM prompt already expects ("LỊCH SỬ HỘI THOẠI (cũ ->
+        mới)")."""
+        import psycopg
+
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
+            rows = connection.execute(
+                """
+                SELECT line FROM rag_conversation_turn
+                WHERE conversation_id = %s ORDER BY id DESC LIMIT %s
+                """,
+                (conversation_id, limit),
+            ).fetchall()
+        return [row[0] for row in reversed(rows)]
+
+    def append(self, conversation_id: str, line: str) -> None:
+        import psycopg
+
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
+            connection.execute(
+                "INSERT INTO rag_conversation_turn (conversation_id, line) "
+                "VALUES (%s, %s)",
+                (conversation_id, line),
+            )