docs(runtime): internal/runtime/loop 44 个文件注释改中文（逐行 1:1，代码零字节改动）

主人 2026-10-03 拍板「注释一律中文」；本模块按长期分批协议（一个模块一次派工）执行。
注释行数 1329 → 1329 逐行 1:1（含中文 33 → 1304）；行尾注释 5 条一并译。
校验：机械校验（非注释改动行）= 空；独立 Go AST 程序（剥注释复渲）证明 44/44 文件代码全等；
全树 go test ./... 58 包 ok / 0 FAIL；交叉编译绿。
gofmt：capacity_loop_test.go / flow.go 改动前即未格式化，按铁律保留原样未动。

internal/runtime/loop/checkpoint.go

@@ -11,29 +11,29 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/persist"
 )
 
-// checkpointStore is the minimal surface the loop needs to persist a durable
-// snapshot and resume an interrupted run (design §1 durable execution).
-// *persist.Store satisfies it in production; tests substitute a fake to assert
-// the resume path without a real database.
+// checkpointStore 是循环持久化持久快照并恢复被中断
+// run 所需的最小表面（design §1 durable execution）。
+// 生产环境由 *persist.Store 满足；测试替换为假实现，以在
+// 没有真实数据库的情况下断言恢复路径。
 type checkpointStore interface {
-	// Save upserts a checkpoint.
+	// Save 插入或更新一个 checkpoint。
 	Save(ctx context.Context, cp *persist.Checkpoint) error
-	// LoadLatest returns the most recently updated checkpoint for agentID, or
-	// ErrNotFound when the agent has none.
+	// LoadLatest 返回 agentID 最近更新的 checkpoint；
+	// 该 agent 没有时返回 ErrNotFound。
 	LoadLatest(ctx context.Context, agentID string) (*persist.Checkpoint, error)
 }
 
-// checkpointState is the durable snapshot persisted at every step. Carrying the
-// message window and the step index lets a long-running task be resumed from the
-// latest checkpoint without losing context (design §1 durable execution).
+// checkpointState 是每步持久化的持久快照。携带消息
+// 窗口与步索引，使长任务能从最新 checkpoint 恢复而不
+// 丢失上下文（design §1 durable execution）。
 type checkpointState struct {
 	Step       int             `json:"step"`
 	Messages   []model.Message `json:"messages"`
 	StopReason string          `json:"stop_reason,omitempty"`
 }
 
-// saveCheckpoint persists a snapshot at step. It never fails the run: a nil store
-// is a no-op and a marshal/save error is logged.
+// saveCheckpoint 在 step 处持久化一个快照。它绝不让 run 失败：
+// store 为 nil 时为空操作，marshal/save 出错则记录日志。
 func (l *Loop) saveCheckpoint(ctx context.Context, sessionID string, step int, messages []model.Message, stopReason string) {
 	if l.checkpoint == nil {
 		return
@@ -59,19 +59,19 @@ func (l *Loop) saveCheckpoint(ctx context.Context, sessionID string, step int, m
 	}
 }
 
-// loadCheckpoint resumes an interrupted run from the session's latest durable
-// checkpoint (design §1 durable execution). It returns the restored message
-// window, the next loop step to run (the checkpoint step + 1) and whether a
-// resumable checkpoint was found. When no checkpoint exists, or the latest one
-// reflects a completed/blocked run, it returns the caller's messages with step 1
-// so the loop starts fresh.
+// loadCheckpoint 从该会话最新的持久 checkpoint 恢复一个被中断的
+// run（design §1 durable execution）。它返回恢复后的消息
+// 窗口、下一个要运行的循环步（checkpoint 步 + 1），以及
+// 是否找到了可恢复的 checkpoint。不存在 checkpoint，或最新的一个
+// 反映的是已完成/被阻塞的 run 时，它返回调用方的消息并置步为 1，
+// 使循环从新开始。
 //
-// Restoration is idempotent: the checkpoint captures the message window exactly as
-// it stood at the end of the last completed step (assistant tool calls plus their
-// results are already fused in), so resuming at the next step rebuilds the model
-// context from history and never re-executes a completed step's tool call nor
-// regenerates a finished answer. Loading is read-only — a resume never deletes the
-// checkpoint, so every step's snapshot stays queryable.
+// 恢复是幂等的：checkpoint 精确保存了上一个已完成步结束时
+// 的消息窗口（assistant 工具调用及其结果已融合在内），
+// 因此从下一步恢复会由历史重建模型
+// 上下文，绝不重跑已完成步的工具调用，也不
+// 重新生成已完成的答案。加载是只读的 —— 恢复绝不删除
+// checkpoint，因此每一步的快照始终可查询。
 func (l *Loop) loadCheckpoint(ctx context.Context, sessionID string, messages []model.Message) ([]model.Message, int, bool) {
 	if l.checkpoint == nil {
 		return messages, 1, false
@@ -96,16 +96,16 @@ func (l *Loop) loadCheckpoint(ctx context.Context, sessionID string, messages []
 	return st.Messages, cp.Step + 1, true
 }
 
-// resumableStopReason reports whether a checkpoint whose last step ended with the
-// given stop reason represents an in-flight run worth resuming. Only a
-// mid-tool-loop step is genuinely "interrupted": the loop folded the tool results
-// into the window and was about to continue, so picking up at the next step
-// rebuilds context without duplicating work. A checkpoint carrying a final answer,
-// a budget exhaustion, a safety-net termination or a guardrail denial is a
-// completed/blocked run — resuming it would regenerate an already-produced answer
-// (a duplicate side effect) — so those fall through to a fresh start (a caller
-// opening a genuinely new conversation uses a distinct sessionID, which misses
-// LoadLatest anyway).
+// resumableStopReason 报告最后一个步以上述 stop reason 结束的
+// checkpoint 是否代表一个值得恢复的在途 run。只有
+// 工具循环中途的步才真正算「被中断」：循环已把工具结果
+// 并入窗口并即将继续，因此从下一步接上
+// 会重建上下文而不重复工作。携带最终答案、
+// 预算耗尽、安全网终止或护栏拒绝的 checkpoint 是
+// 已完成/被阻塞的 run —— 恢复它会重新生成已产出的答案
+// （重复副作用）—— 因此这些一律落到全新开始（打开
+// 真正新会话的调用方使用不同的 sessionID，本就错过
+// LoadLatest）。
 func resumableStopReason(stopReason string) bool {
 	switch stopReason {
 	case "", agent.StopReasonToolCalls: