docs(comments): 注释中文化 · internal/runtime/orchestration（29 文件，纯替换）

- 只翻注释/文档字符串：剥离注释后代码行逐字节不变（机械校验只剩允许的行尾注释形态）
- 裁决编号与 design 节号（02.4.3 / 11.4.2 / design/05 §9 等）逐字保留
- 复核：go build 绿；orchestration 及其子包（audit/comm）测试绿；gofmt 干净

internal/runtime/orchestration/spawnflow_test.go

@@ -17,11 +17,10 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
 )
 
-// TestSubagentBranchFoldsIntoParentGraph is the end-to-end wiring test for the
-// branch visualization: a real loop driven by the scheduler under a parent
-// session must appear in that session's graph as a subagent lane root whose inner
-// llm/tool steps hang off it, must not start a round (so the frontend's single
-// flowGraph is never replaced), and must not open a graph under its own id.
+// TestSubagentBranchFoldsIntoParentGraph 是分支可视化的端到端接线测试：由调度器在父会话
+// 之下驱动的真实循环，必须作为子代理泳道根出现在该会话的图里、其内部 llm/tool 步骤挂在
+// 它下面；必须不开启新一轮（否则前端唯一的 flowGraph 会被替换）；也必须不以其自身 id
+// 另开一张图。
 func TestSubagentBranchFoldsIntoParentGraph(t *testing.T) {
 	t.Parallel()
 	capture := &graphCapture{}
@@ -64,8 +63,7 @@ func TestSubagentBranchFoldsIntoParentGraph(t *testing.T) {
 	if subNode.Kind != "subagent" || subNode.Status != trace.FlowStatusDone {
 		t.Fatalf("subagent node = %+v, want kind subagent / status done", subNode)
 	}
-	// The branch's steps chain inside its own lane: llm (asks for the tool), tool,
-	// llm (answers).
+	// 该分支的步骤在自己的泳道内成链：llm（请求工具）、tool、llm（作答）。
 	chain := flowLaneChain(t, g, subNode)
 	if len(chain) != 3 || chain[0].Kind != "llm" || chain[1].Kind != "tool" || chain[2].Kind != "llm" {
 		t.Fatalf("lane chain = %s, want llm>tool>llm", chainKinds(chain))
@@ -75,7 +73,7 @@ func TestSubagentBranchFoldsIntoParentGraph(t *testing.T) {
 		t.Fatalf("input node label = %q, want the parent round's seed text", in.Label)
 	}
 
-	// Node logs: every executed step must carry a real event, and nothing else.
+	// 节点日志：每个已执行的步骤都必须带一条真实事件，且不带别的。
 	llmEvent := flowLogEvent(t, llmNode, loop.EventLLMDone)
 	if got := fmt.Sprint(llmEvent.Props["model"]); got != "mock-1" {
 		t.Fatalf("llm log model = %q, want mock-1 (props: %+v)", got, llmEvent.Props)
@@ -97,7 +95,7 @@ func TestSubagentBranchFoldsIntoParentGraph(t *testing.T) {
 		t.Fatalf("tool log args_keys = %q, want the argument names", got)
 	}
 
-	// Red line: no credential from the tool call or its result may be published.
+	// 红线：工具调用或其结果中的任何凭据都不得被发布。
 	blob, merr := json.Marshal(g)
 	if merr != nil {
 		t.Fatalf("marshal graph: %v", merr)
@@ -113,10 +111,9 @@ func TestSubagentBranchFoldsIntoParentGraph(t *testing.T) {
 	}
 }
 
-// TestSpawnParallelBranchesStaySeparate asserts two sub-agents running at the same
-// time land in two disjoint swimlanes of the parent graph: each lane has its own
-// llm→tool chain and nothing in one lane is reachable from the other (design/19 §2
-// 泳道, the cross-linking failure the lane cursor exists to prevent).
+// TestSpawnParallelBranchesStaySeparate 断言同时运行的两个子代理落在父图的两条互不相交
+// 的泳道里：每条泳道有自己的 llm→tool 链，且一条泳道里没有任何东西可从另一条到达
+// （design/19 §2 泳道，也就是泳道游标存在的意义——防交叉链接失效）。
 func TestSpawnParallelBranchesStaySeparate(t *testing.T) {
 	t.Parallel()
 	capture := &graphCapture{}
@@ -169,17 +166,15 @@ func TestSpawnParallelBranchesStaySeparate(t *testing.T) {
 			t.Fatalf("node %s is reachable from both lanes (cross-linked edges: %+v)", id, g.Edges)
 		}
 	}
-	// The trunk cursor must still be the parent's seed, not the last branch node.
+	// 主干游标必须仍是父轮的种子，而不是最后那个分支节点。
 	if !hasFlowEdge(g, "input", nodeA.ID) || !hasFlowEdge(g, "input", nodeB.ID) {
 		t.Fatalf("edges = %+v, want both branch roots hanging off the parent's input node", g.Edges)
 	}
 }
 
-// TestSubagentRunDoesNotStreamToFrontSession asserts a sub-agent's inner stream is
-// muted: the loop still drives the model (its llm step appears in the parent's flow
-// lane), but no delta reaches the caller's OnStream, so a background sub-agent's
-// monologue never leaks into the front conversation (design/05 §8, design/04 §2
-// 不污染前台).
+// TestSubagentRunDoesNotStreamToFrontSession 断言子代理的内部流被静音：循环仍驱动模型
+// （其 llm 步骤出现在父图的泳道里），但没有任何增量抵达调用方的 OnStream，因此后台子代理
+// 的独白绝不会泄漏进前台对话（design/05 §8, design/04 §2 不污染前台）。
 func TestSubagentRunDoesNotStreamToFrontSession(t *testing.T) {
 	t.Parallel()
 	capture := &graphCapture{}
@@ -228,7 +223,7 @@ func TestSubagentRunDoesNotStreamToFrontSession(t *testing.T) {
 	if len(got) != 0 {
 		t.Fatalf("sub-agent pushed %d stream events into the front session: %+v", len(got), got)
 	}
-	// The work is still visible where it belongs: as a lane of the parent's graph.
+	// 该工作仍在它该在的地方可见：作为父图的一条泳道。
 	g := capture.graph(t, parentSession)
 	subNode := flowSubagentNode(t, g, sub.ID)
 	if chain := flowLaneChain(t, g, subNode); len(chain) == 0 {
@@ -236,9 +231,8 @@ func TestSubagentRunDoesNotStreamToFrontSession(t *testing.T) {
 	}
 }
 
-// TestSpawnWithoutParentSessionStaysOutOfEveryGraph asserts the flow attribution is
-// opt-in: a spawn with no ParentSession publishes no graph at all, so an
-// unattributed background agent can never surface as a stray lane.
+// TestSpawnWithoutParentSessionStaysOutOfEveryGraph 断言流图归属是选择性的：未带
+// ParentSession 的派发根本不发布任何图，因此未归属的后台 agent 绝不会作为游离泳道浮现。
 func TestSpawnWithoutParentSessionStaysOutOfEveryGraph(t *testing.T) {
 	t.Parallel()
 	capture := &graphCapture{}