feat(runtime): add dynamic workflow execution

[kingfs]

Summary

• add a restricted JavaScript dynamic-workflow engine with agent, parallel, pipeline, scoped phases, nested workflows, budgets, cancellation, replay, and real Git worktree isolation
• preserve legacy runtime behavior by keeping stateRoot unchanged and assigning isolated sessionRoot directories only to workflow child agents
• add UTF-8-safe incremental workflow event/result decoding, protocol validation, and typed failures to @chaitin/agent-compose-runtime-sdk
• persist nested workflow snapshots, validate linked worktrees before cache reuse, and interpret abandoned state as interrupted
• document the runtime contract, persistence layout, cache identity, failure semantics, and compatibility guarantees

Testing

• task lint
• task build
• npm test (runtime/javascript: 223 passed, 87.87% statement coverage)
• npm test (runtime/agent-compose-runtime-sdk: 41 passed)
• TEST_SHAPE=integration npm test (runtime/javascript: 61.11% statement coverage)
• TEST_SHAPE=e2e npm test (runtime/javascript: 62.07% statement coverage)
• task docs:build
• task test (workflow unit/integration/e2e shapes pass; the full command currently fails in the unrelated existing pkg/agentcompose/app/TestProjectAgentModelResolverUsesCurrentRevisionAndDaemonDefault, where seeded gemini-3.1-pro-preview overrides the test's expected daemon default)
• git diff --check

Checklist

• Documentation updated for the new runtime and SDK behavior.
• Unit, integration, E2E, compatibility, cancellation, replay, and worktree tests added.
• Existing non-workflow stateRoot and session behavior remains compatible.
• No secrets, private endpoints, certificates, or local runtime state included.

Sample

• https://github.com/kingfs/agent-compose-sample/tree/master/agents/dynamic-workflow

[monkeyscan]

评审任务执行失败。

[kingfs]

@monkeyscan review pr

[monkeyscan]

评审任务执行失败：任务执行超时。

[monkeyscan]

评审任务执行失败：任务执行超时。

[kingfs]

@monkeyscan review pr

[monkeyscan]

评审任务执行失败：任务执行超时。

[kingfs]

@monkeyscan review pr

[monkeyscan]

@kingfs wrote:
@monkeyscan review pr

处理失败，请稍后重试。

[eetoc]

已基于 PR #564 当前 HEAD（254f7358）完成本地回归验证。

验证时使用该 PR 代码实际构建 daemon 和 guest 镜像，通过真实 daemon、Docker guest、持久化存储、Git 仓库和 worktree 生命周期运行。LLM 调用边界使用协议兼容的可控 stub，以稳定控制返回结果和取消时序；本次验证不包含外部 LLM 服务端行为。

实际结果：

1. 连续在两个 scoped phase 中调用 agent，或在 scoped phase 后继续调用 agent，后续调用会重复生成 root/agent:0，工作流报错：

   duplicate workflow invocationKey: root/agent:0
   

2. 在 child controller 创建前取消父 signal，观察到：

   parentAborted=true
   childAborted=false
   

   在准备阶段触发该取消窗口后，provider 仍会继续执行，agent 最终可被记录为 done。provider 已启动后的常规取消路径可以正常停止。

3. 使用同一个 worktree agent 连续执行三个独立 run：

   Run A: 首次执行
   Run B: resume Run A
   Run C: resume Run B
   

   实际结果：

   provider calls = [1, 0, 1]
   statuses       = [done, cached, done]
   

   Run B 成功命中缓存，但其记录缺少 worktreeHead；Run C resume Run B 时缓存校验失败，重新执行 provider 并创建了新的 worktree。

基础测试结果：

Runtime SDK:        41/41 passed
Runtime JavaScript: 223/224 passed

以上三个问题均已通过实际运行或可控时序验证确认。基于当前结果，建议修复并补充对应回归测试后再合入。

[monkeyscan]

评审任务执行失败。

[monkeyscan]

评审任务执行失败：任务执行超时。

[monkeyscan]

评审任务执行失败：任务执行超时。

[eetoc]

LGTM