feat(history): 会话消息持久化 + 主/次级会话（design/16）

- runtime/history: Store + HistoryStore 接口（Append/List 游标分页/Recent/DeleteSession/NewSubSession/ListSessions），主会话固定 web、次级 web:<uuid>
- db 迁移 0003: chat_sessions + chat_messages + idx_chat_msgs_session_created（事务幂等）
- gateway: GET/POST /sessions、GET /sessions/{id}/messages 热分页、DELETE /sessions/{id}（主不可删 403）；Deps.History 可空注入
- 编排层落库: session_persist.go（submitInternal 写用户消息、completeTurn 写助手消息，拒绝的消息不落库）
- WS: wsEvent 加 sessionId 定向 + history 回放帧（握手上行后回放主会话最新 50 条，thinking 剥离）
- wire.go 装配 History 到 gateway.Deps.History
- 端到端实测: POST /sessions 建次级 201 / GET /sessions 列表 / POST /chat 后 /messages 读到 user+assistant / DELETE 次级成功 / DELETE 主会话 403

internal/services/gateway/ws_history.go

@@ -0,0 +1,34 @@
+package gateway
+
+import (
+	"context"
+)
+
+// wsReplayLimit is the number of most-recent messages replayed to a connecting
+// client. It is intentionally small so a reconnect is cheap; the frontend pulls
+// the full active-session history via REST (design/16 §4.1) and the replay is
+// only there to give a brand-new connection immediate context.
+const wsReplayLimit = 50
+
+// replayHistory sends the most recent messages for sessionID to a single,
+// freshly-registered connection as a ws "history" event. It is best-effort: a
+// nil store or a read error skips the replay silently. Thinking is NOT included
+// (design/16 §6) so the frame never leaks chain-of-thought to a connection whose
+// role the gateway did not restrict on upgrade.
+func (h *hub) replayHistory(c *wsConn, sessionID string) {
+	if h.history == nil {
+		return
+	}
+	msgs, err := h.history.Recent(context.Background(), sessionID, wsReplayLimit)
+	if err != nil {
+		h.log.Warn("history replay", "err", err, "session", sessionID)
+		return
+	}
+	out := make([]historyMessage, 0, len(msgs))
+	for _, m := range msgs {
+		wire := toHistoryMessage(m)
+		wire.Thinking = ""
+		out = append(out, wire)
+	}
+	c.enqueueJSON(wsEvent{Type: "history", SessionID: sessionID, Data: map[string]any{"messages": out}})
+}