feat(loop): Top5#5 上下文溢出压缩 + 每轮 assemble（对齐 OpenClaw context-engine）——Top5 全完成

- 每轮 assemble(transformContext 等价)：run.go 每 step 前 assembleContext——依次 Inject(每轮
  重新 Retriever.Inject 不缓存首轮)、WorkingMemory 前置 Summarize、可插拔 TransformContext 投影/
  过滤；生成的 view 交 step，msgs 保持纯对话供工具结果追加，记忆不重复注入、不写回 checkpoint。
- Compactor：结构分支感知压缩——estimateContextTokens 启发式超 MaxContextTokens 且消息≥3 才动；
  findKeepStart 按 user 消息切整轮从后往前累加(至少保留最后一整轮→尾随工具结果绝不切片/不跨轮)；
  旧轮 user/assistant 对压成一条 RoleSystem 摘要置顶，既有摘要前缀识别合并防叠；capSummary 上限
  4000 rune；ctx 即 abortSignal(取消返回 ctx.Err，压缩失败仅记日志不中断 run)。
- 选项：Options 增 MaxContextTokens/Compactor/TransformContext，均 nil-tolerant(默认=现有行为，
  零值不压缩)；DefaultMaxContextTokens=32768。Compactor 接口预留注入 LLM 摘要。

go build/vet/test -count=1 全绿 + windows/darwin 交叉编译。12 断言覆盖尾随工具结果保留/不跨轮切割/
摘要上限/摘要合并/ctx 取消/每轮重检索/TransformContext 逐轮/超限压缩/默认关闭。-race 受宿主
TSan VMA 限制无法运行(与代码无关)。

**至此 Top5(1~5) 全部完成并推库。**

internal/runtime/loop/options.go

@@ -1,6 +1,7 @@
 package loop
 
 import (
+	"context"
 	"log/slog"
 
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
@@ -27,7 +28,8 @@ type Options struct {
 	Tools *tool.Dispatcher
 	// Memory is the long-term memory store used to persist extracted items.
 	Memory *memory.Store
-	// Retriever injects relevant memories into the first-turn context.
+	// Retriever injects relevant memories into the context, re-evaluated every
+	// round (see assembleContext) so retrieval tracks the latest conversation.
 	Retriever *memory.Retriever
 	// WorkingMemory is the per-session sliding window, provided for the
 	// orchestrator to inspect; the loop does not mutate it.
@@ -62,6 +64,25 @@ type Options struct {
 	// nil-tolerant; these are reserved for "敏感操作代审" (design/11) and the
 	// future asynchronous-tool path. A zero Hooks value is a no-op.
 	Hooks ToolLoopHooks
+	// MaxContextTokens is the context budget that triggers overflow compaction.
+	// Non-positive (the zero value) disables compaction, preserving the loop's
+	// previous behavior. Setting it to a positive value (e.g.
+	// DefaultMaxContextTokens) enables compaction: when the loop's estimated
+	// context exceeds the budget it summarizes older turns into a leading summary
+	// message while keeping the recent tail (including trailing tool results)
+	// intact (design §1.3). Nil-tolerant: leaving it unset changes nothing.
+	MaxContextTokens int
+	// Compactor compacts an over-budget conversation. Nil selects the built-in
+	// structural compactor (no model call); a non-nil value lets a caller inject
+	// an LLM-backed or custom summarizer. It is only invoked when
+	// MaxContextTokens is positive.
+	Compactor Compactor
+	// TransformContext, when non-nil, is called immediately before every model
+	// step with the assembled message list (after memory injection and any
+	// compaction) and the current round number (1-based). It may project, filter
+	// or re-order the messages the model sees. It is the loop's equivalent of
+	// OpenClaw's transformContext (design §1.3). Nil = no projection.
+	TransformContext func(ctx context.Context, messages []model.Message, round int) []model.Message
 	// Log is the base logger. Nil is replaced with a discard logger tagged with
 	// source "runtime/loop".
 	Log *slog.Logger