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/run.go

@@ -9,10 +9,12 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
 )
 
-// Run drives the agent loop: inject relevant memory into the first-turn context,
-// then run up to MaxRounds steps. Each step runs the model (streaming through
-// OnStream when configured) and any requested tool calls are executed under a
-// per-invoke timeout and fed back as RoleTool messages so the model can continue.
+// Run drives the agent loop: assemble a fresh model-visible context every round
+// (re-inject relevant memory from the current conversation, optionally compacting
+// the older turns when the context exceeds the configured budget), then run up to
+// MaxRounds steps. Each step runs the model (streaming through OnStream when
+// configured) and any requested tool calls are executed under a per-invoke
+// timeout and fed back as RoleTool messages so the model can continue.
 // A checkpoint is persisted at every step, every step is traced as a
 // state_transition span (with an llm_call child and tool_call spans for each
 // invocation), and after each step the memory Extractor runs with errors logged
@@ -29,7 +31,7 @@ func (l *Loop) Run(ctx context.Context, sessionID string, messages []model.Messa
 		return nil, fmt.Errorf("loop: run: %w", ErrNilAgent)
 	}
 
-	msgs := l.injectMemory(ctx, messages)
+	msgs := messages
 	toolState := &toolLoopState{prevFP: make(map[string]bool)}
 	res := &Result{}
 	for step := 1; step <= l.maxRounds; step++ {
@@ -37,8 +39,14 @@ func (l *Loop) Run(ctx context.Context, sessionID string, messages []model.Messa
 			return nil, fmt.Errorf("loop: run: step %d: %w", step, err)
 		}
 
+		// Compact older turns when the conversation exceeds the context budget,
+		// then assemble the model-visible context for this round (fresh memory
+		// injection + transform hook), so every step sees an up-to-date context.
+		msgs = l.compactIfNeeded(ctx, msgs)
+		view := l.assembleContext(ctx, msgs, step)
+
 		rc, roundSpan := l.span(ctx, "loop.step", trace.SpanKindStateTransition)
-		outcome, err := l.step(rc, sessionID, msgs)
+		outcome, err := l.step(rc, sessionID, view)
 		if err != nil {
 			l.spanProp(roundSpan, "error", err.Error())
 			l.finishSpan(roundSpan, trace.SpanStatusError)