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/memory.go

@@ -7,11 +7,12 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
 )
 
-// injectMemory retrieves the persona's relevant long-term memories and injects
-// them into the first-turn context as a leading system message, so the model can
-// ground its answer in recalled context (design §6 / §2). It returns the input
-// messages unchanged when no retriever/persona is configured, retrieval fails, or
-// nothing relevant is found.
+// injectMemory retrieves the persona's relevant long-term memories and prepends
+// them as a leading system message, so the model can ground its answer in recalled
+// context (design §6 / §2). It is re-run every round via assembleContext so the
+// retrieval reflects the latest conversation rather than a first-turn snapshot. It
+// returns the input messages unchanged when no retriever/persona is configured,
+// retrieval fails, or nothing relevant is found.
 func (l *Loop) injectMemory(ctx context.Context, messages []model.Message) []model.Message {
 	if l.retriever == nil || l.personaID == "" {
 		return messages
@@ -30,6 +31,50 @@ func (l *Loop) injectMemory(ctx context.Context, messages []model.Message) []mod
 	return out
 }
 
+// assembleContext builds the model-visible message list for one round (design
+// §1.3 每轮 assemble, the loop's equivalent of OpenClaw's transformContext): it
+// re-injects the persona's relevant long-term memory — retrieval is re-run every
+// round from the current conversation, not cached from the first turn, so it
+// supports cross-round refresh — folds in the per-session working-memory summary
+// when present, and applies the optional TransformContext hook last so a caller
+// can project/filter what the model actually sees. It returns the input unchanged
+// when nothing is configured.
+func (l *Loop) assembleContext(ctx context.Context, messages []model.Message, round int) []model.Message {
+	var leading []model.Message
+	if l.working != nil {
+		if s := l.working.Summarize(); s != "" {
+			leading = append(leading, model.TextMessage(model.RoleSystem, "工作记忆: "+s))
+		}
+	}
+	leading = append(leading, l.injectMemory(ctx, messages)...)
+
+	out := leading
+	if l.transformContext != nil {
+		out = l.transformContext(ctx, leading, round)
+	}
+	return out
+}
+
+// compactIfNeeded compacts the conversation when its estimated context exceeds
+// the configured MaxContextTokens. It is a no-op when compaction is disabled
+// (non-positive MaxContextTokens), the compactor is nil, the estimate is within
+// budget, or the compactor returns an error — a compaction failure is logged and
+// the conversation is kept as-is rather than failing the run.
+func (l *Loop) compactIfNeeded(ctx context.Context, messages []model.Message) []model.Message {
+	if l.maxContextTokens <= 0 || l.compactor == nil {
+		return messages
+	}
+	compacted, err := l.compactor.Compact(ctx, messages, l.maxContextTokens)
+	if err != nil {
+		l.log.Warn("context compact", "error", err, "tokens", l.maxContextTokens)
+		return messages
+	}
+	if compacted == nil {
+		return messages
+	}
+	return compacted
+}
+
 // extractMemory runs the extraction hook (design §6) over the current
 // conversation and persists any derived items. Errors are logged, never fatal, so
 // a bad extraction cannot abort the run.