feat(llm): 模型配置新增输出上限,支持 max_completion_tokens 字段名

llm_profiles 加两列:max_tokens(单次回复输出上限,0=不发送该字段,由服务端
默认值决定)与 max_tokens_field(上限发哪个键)。后者仅对 format=openai 生效:
空(默认)= max_tokens;'max_completion_tokens' = 新键。OpenAI 推理模型
(o 系列/GPT-5)拒收 max_tokens、只认新键,而多数兼容网关反过来只认旧键,
故做成显式开关而非自动推断。anthropic 与 openai-responses 自带字段名,
存下来只会误导,保存时一并清空。

顺带补上一处从未接通的链路:planner/worker/chat/mainagent/goals 五个
agentcore.Options 都没设过 MaxTokens,恒为 0——openai 系因 omitempty 压根
不发上限,anthropic 走 norma 的 8192 兜底。光加配置项而不接这条线,用户在
后台填了数字也不会生效,故本次一并接通,取值与 streaming 一样按 profile
每轮解析,故障转移切换后下一轮即生效。

task_llm.go 里 nonStreaming() 的来源优先级(agent 绑定 → 任务链激活 profile
→ 全局)抽成共用的 activeCfg(),避免 maxTokens() 再抄一份同样的三级回退。

norma 升到 v0.3.3(该版本提供 Config.MaxTokensField)。go.mod 只动 norma 这
一行:仓库此前未 tidy 过,跑 go mod tidy 会顺带升 x/crypto、x/net 等一批无关
依赖,不混进本次改动。

迁移:新列走 ADD COLUMN IF NOT EXISTS + 默认 0/'',旧库升级后行为完全不变;
CHECK 约束限定取值,每次启动先删后建保持幂等。

agent/goals.go

@@ -98,13 +98,14 @@ func DecomposeGoals(ctx context.Context, prov llm.Provider, dataDir, goalText, d
 	if prov == nil {
 		return nil
 	}
-	return DecomposeGoalsWithProvider(ctx, prov, dataDir, goalText, desc, as, ts, taskID, false, emit)
+	return DecomposeGoalsWithProvider(ctx, prov, dataDir, goalText, desc, as, ts, taskID, false, 0, emit)
 }
 
 // DecomposeGoalsWithProvider is the task-runtime variant used when a task has an
 // ordered provider chain. It preserves the same tools and write behavior while
-// letting the caller own provider selection/failover.
-func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, nonStreaming bool, emit func(db.Activity)) []GoalSpec {
+// letting the caller own provider selection/failover. maxTokens is the profile's
+// per-reply output cap (0 = send none).
+func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir, goalText, desc string, as *db.AssetStore, ts *db.ExplorationStore, taskID int64, nonStreaming bool, maxTokens int, emit func(db.Activity)) []GoalSpec {
 	if prov == nil {
 		return nil
 	}
@@ -147,6 +148,7 @@ func DecomposeGoalsWithProvider(ctx context.Context, prov llm.Provider, dataDir,
 		// 3 步(抽约束 → 登记范围 → 拆目标)各需一次工具调用,给足回合避免收尾前漏调 set_goals。
 		MaxTurns:     8,
 		NonStreaming: nonStreaming, // 该 profile 选非流式时走 Provider.Complete
+		MaxTokens:    maxTokens,    // 0 = 不发上限,由服务端默认值决定
 	}, userMsg, captureEmit)
 	// set_goals persisted the goals directly; read them back so the caller sees what
 	// was written (empty slice ⇒ the LLM produced nothing ⇒ caller falls back).