feat(runtime): support sequential agent graph epochs

[me2seeks]

Summary

• persist monotonic root-to-graph epoch bindings while preserving the existing deterministic graph as epoch 1
• advance a finished, quiescent graph before the next graph/swarm root Turn is durably admitted
• fence supervisor wakes and resolve stop, permission, revision, and retirement paths through durable graph ownership

A finished graph remains immutable. The next graph-capable Turn receives a new deterministic graph ID, and concurrent Hosts converge through a storage CAS instead of reopening or skipping the previous graph. Ordinary default Turns leave the finished graph untouched. Legacy epoch 1 stays virtual until the first real cutover, so Sessions do not acquire graph state just by being read.

This establishes the lifecycle boundary needed by #2588. Historical graph selection and cross-epoch result inputs remain follow-up work.

Refs #2588

Verification

• npm --workspace @maka/core run test — 539 passed
• npm --workspace @maka/storage run test — 777 passed, 16 skipped
• npm --workspace @maka/runtime run test — 2,788 passed, 6 skipped
• npm --workspace @maka/runtime-host run test — 910 passed
• builds for Core, Storage, Runtime, and Runtime Host
• Biome on all changed files
• git diff --check

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — a graph/swarm Turn can advance from a finished Agent Graph to a fresh graph epoch
• No

中文说明

摘要

• 持久化 root Session 到 Agent Graph 的单调 epoch 绑定，同时保持原有确定性 graph ID 作为 epoch 1
• 只有当前图已经结束且静默，才会在下一条 graph/swarm root Turn 持久化之前切换到新 epoch
• supervisor wake、停止、权限回答、会话修订与删除清理都通过 Runtime Host 的持久化 graph ownership 解析

已经 finish 的图仍然不可修改。下一条启用 Graph 的 Turn 使用新的确定性 graph ID；多个 Host 并发切换时由 Storage CAS 收敛，不会重开旧图或跳过 epoch。普通 default Turn 不会切换 graph epoch。第一次真正切换前，legacy epoch 1 只是只读兼容值，因此会话不会仅因查询而写入图状态。

本 PR 建立 #2588 所需的生命周期基础。历史图选择和跨 epoch 结果输入将在后续实现。

[likun666661]

LGTM. The Runtime Host-owned cutover, terminal/quiescent recheck, wake suppression, and storage CAS establish the right lifecycle boundary for sequential graph epochs.

Non-blocking follow-up: please add a recovery test for the crash window where advanceAgentGraphEpoch has committed but supersedeAgentGraphSupervisorWakes has not. On restart, any retryable wake belonging to the prior graph should be deterministically superseded rather than retained or redelivered against the new current epoch.

[me2seeks]

Added the crash-window recovery coverage and tightened recovery so only wakes from the stale graph epoch are superseded; a current-epoch wake for the same root still delivers normally. Also rebased onto current main.