fix(scheduler): dispatch stranded peer handoffs

[huangruiteng]

Closes #4311

Motivation / 动机

An executor-excluded independent_handoff could remain globally unclaimed while the origin monitor kept polling. Delivery through a Codex message was not a durable dispatch or claim receipt.

被 origin Agent 排除的 independent_handoff 可能长期保持全局未认领，而 origin monitor 继续轮询；仅发送 Codex 消息不构成持久派发或认领回执。

What changed / 改动

• distinguish dispatchable and typed no_eligible_peer frontiers; both are non-quiet

• reuse the existing manager-context inbox for same-Goal peer dispatch, with restart-safe dispatched -> read -> claimed receipts and terminal stale-claim conflict

• require live canonical Todo eligibility before first dispatch and canonical claim readback before acknowledgement

• share manager.context_inbox between quota heartbeat and managed turn run-once; project the bounded receipt through the Turn contract capsule used by frontend/Lark consumers

• add a bilingual RFC and process-restart CLI acceptance test

• 区分 dispatchable 与类型化 no_eligible_peer frontier，二者均不可静默等待

• 复用现有 manager-context inbox 完成同 Goal peer 派发，并提供跨重启的 dispatched -> read -> claimed 回执及 stale-claim 冲突终态

• 首次派发前校验规范 Todo eligibility，确认前再次校验规范认领

• quota heartbeat 与托管 turn run-once 共享 manager.context_inbox；有界回执进入 frontend/Lark 共用的 Turn contract capsule

• 增加双语 RFC 与跨进程 CLI 验收

Product surfaces / 产品入口

CLI/quota and managed Turn changed. Lark consumes the same Turn-start inbox projection. No new frontend control is needed because this is internal routing rather than configurable state; the existing frontend consumes the shared Turn capsule.

Validation / 验证

• ruff check on changed Python and tests
• pytest -q tests/test_manager_context_handoff.py tests/control_plane/test_quota_selection.py tests/test_turn_envelope.py (55 passed)
• npm run typecheck:control-plane
• targeted TypeScript quota-selection and Turn-envelope tests (18 passed)

The full TypeScript suite was also attempted; Python-spawning legacy cases used the host Python 3.9 and failed on the repository's existing dataclass(slots=True) Python >=3.11 requirement. The targeted changed surfaces pass under the pinned Python 3.13 environment.

[huangruiteng]

结论：REQUEST_CHANGES（exact head 7ea69f114c9ed7c8d3bc315658e7e46c4ebca2aa）。

动机

这个 PR 解决了“quota/Turn 已经选出同一 Goal 的 peer handoff，但没有真正投递”的断路问题：它引入了带确定性 receipt 的 manager-context dispatch，并在创建 peer claim 前重新读取 canonical Todo。问题定义和同 Goal、已注册 peer 的边界是清楚的。

改动思路

整体流转是 typed quota-selection frontier → dispatch_from_quota_decision → 私有 0600 handoff receipt/inbox → canonical Todo claim/readback → quota/Turn payload。no_eligible_peer、未注册 source、replay 和 claim conflict 都有显式分支，没有按可见 agent 顺序猜 recipient。这个分层复用了 manager-context、Todo 与 agent-scope 的现有 owner。

具体改动

新增 loopx/capabilities/manager_context/agent_handoff.py，负责 receipt identity、同 Goal peer 过滤、canonical Todo 重读、claim/ack 和幂等 replay；quota_selection.ts/Python projection 增加 dispatchable 与 no_eligible_peer frontier；quota should-run 和 managed Turn 在 agent id 存在时接入 dispatch，并把 receipt 投影到 payload。对应的 55 个 focused tests（manager handoff、quota selection、Turn envelope）在 exact head 全部通过。

关键代码讲解

唯一必须修复的阻塞点在 loopx/cli_commands/quota.py:736-749：handle_quota_command 内联了 dispatch_from_quota_decision 的 import/call。这个函数本来已经是仓库维护性 ratchet 关注的决策热点；基线 58dbaeaec9bfad537f5b8ad2b1def9b2dcbd5197 的 ratchet 通过且没有 oversized decision finding，而 exact head 新增了 oversized_decision_function:loopx.cli_commands.quota:handle_quota_command。请把这段调用抽成现有 quota command seam 下的有界 helper（保持 agent-id/should-run guard 和同一 payload receipt），不要用 reviewed exception 掩盖增长。

对主干的风险

这是 P1：GitHub hosted run 34737896084 的 test-shard (3)、pytest 和 merge-gate 都因该 maintainability finding 失败，故当前 exact head 不具备可合并状态。功能性测试虽绿，但不能抵消仓库级门禁；若直接合入，后续 quota 控制面会继续在一个已超限的 orchestration function 上叠加逻辑。其余 handoff 风险（跨 Goal、未注册 recipient、stale claim）在本次代码和 negative tests 中已有明确拒绝路径。

我的整体评价

协议边界、canonical Todo authority 和 replay 设计是合理的，变更也落在正确的 manager-context owner；但当前提交必须先修复 quota handler 的 ratchet 回归，再重新跑完整 hosted checks。修复后请回读新的 exact head，并至少确认 maintainability、pytest、test-shard (3) 与 merge-gate 全部通过。

English verdict: REQUEST_CHANGES — extract the new inline handoff dispatch from handle_quota_command; the exact head fails the repository maintainability ratchet and therefore hosted pytest/test-shard/merge-gate.

[huangruiteng]

Rebased onto merged main 7eb4b7b and addressed the requested maintainability failure by extracting the heartbeat receipt reconciliation path from handle_quota_command.

Validation on exact head 1308566:

• tests/canary/test_maintainability_ratchet.py: 8 passed
• manager handoff/quota selection/turn envelope focus: 55 passed
• tests/control_plane/test_quota_settlement_cli.py: 45 passed
• Ruff and git diff --check: passed

The conflict resolution retained latest main reward-memory/periodic imports and todo presentation ordering while preserving this PR manager handoff dispatch semantics. Requesting a fresh independent review; this PR is not being self-merged.

[huangruiteng]

Closing this implementation as superseded by the merged capable-manager semantic-handoff RFC (#4330, commit ddf6efa).

The defect remains real, but this PR makes same-Goal, pre-existing Todo, source-excluded quota selection, and manager_context inbox ownership part of the handoff architecture. The accepted M2 direction requires one typed Core collaboration request/assessment/result contract that also supports pre-Todo and cross-Goal work; #4312 records are migration inputs, not the new universal gate.

保留并迁移的验收义务：dispatchable / no_eligible_peer 可观察性、确定性 legacy alias、重启后的 dispatched→read→claimed 关系、stale claim 冲突、canonical Todo claim readback、duplicate ingress 与 effect-after-crash 对账。它们将进入 M2 的 characterization / compatibility fixtures，并与 #4094 的 continuation adapter 一起验证；不会把当前分支直接并入新 owner。

No accepted request or Todo is being marked complete by closing this PR; the canonical implementation Todo will retain a supersedes link to #4312.