fix(quota): recover undeclared capabilities before fallback selection

[huangruiteng]

A runnable recommendation could suppress runtime-capability verification for blocked work before explicit Turn selection. For example, a P1 local task hid the re-entry route for a P0 network task even though network was merely undeclared. Preserve verification before selection; successful checks re-enter the same Turn, while failed checks retain the explicit fallback command. Committed Todo bindings remain immutable.

Move verification-target selection, authority exclusion and the recovery contract into the existing TS capability-gate owner. Python retains host/scheduler adaptation and shell rendering. Update the generated /loopx skill and the bilingual TS migration RFC. Runtime observations remain host-local; no shared-authority grant, lease or provider write is added.

Validation: 99 Python tests (re-entry, real CLI selection/settlement, compact projection, generated skill installation); 11 TS tests; control-plane typecheck; Ruff; diff check. A baseline-versus-patched fixture reproduces the missing route. LoopX check: zero errors, three pre-existing registry warnings.

User entry points: CLI and managed Turn reuse the existing runtime_capability_reentry and selection-command projection. There is no new setting, API schema or frontend control to expose; no frontend assets change. Generated Codex skill readback is covered. This PR does not claim live Lark delivery acceptance.

Migration economics:

Item	Result
Canonical owner	Python re-entry rules replaced by agents/capability_gate.ts
Removed Python	112 lines removed from re-entry, including target lookup/filtering and contract construction
New bridge	24 lines added in the compatibility facade
Runtime calls	Previously 0 for this Python rule; now 0 healthy / 1 per capability-gap interaction packet; nested CLI projection reuses the result
Product diff	+109 / -122, net -13 lines; tests and docs excluded
Exit condition	Remove the remaining facade when the outer interaction/host command builder moves to TS; no production dual-run

Scope: bounded host-local read-plan cutover, not complete quota/interaction migration or shared-authority promotion.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

我按 exact head 409484d43517715380f35ccd26b5ecca323b0227 重新审阅了整条 origin/main...HEAD。旧行为的问题是真实且会持续累积：当 action_portfolio 仍要求显式选择时，selected_todo 只是建议值，但 Python re-entry 投影会先用它过滤候选，因此更高优先级、缺少 agent-owned runtime capability 的 Todo 可能完全不出现在 agent_channel.next_task_action，agent 只看到 fallback selection。

我用同一 public CLI integration fixture 对 detached base bfd1ec8db846bca3af47e559aa5fe7e515e57370 和当前 head 做了对照：base 因缺少 next_task_action 失败，head 的 selection-required true/false 两条路径均通过。最小的单点 Python 条件虽能修眼前分支，但会继续让 capability eligibility、优先级、receipt binding 与 authority exclusion 跨 Python/TypeScript 双重维护；把这条规则放回既有 TypeScript capability gate 是更小的长期所有权变更。

改动思路

入口仍是现有 build_interaction_contract / quota should-run projection；权威输入仍来自 typed capability gate、action portfolio、Turn receipt、host-observed capabilities 与 scheduler execution context。新的 projectRuntimeCapabilityReentry 只做纯投影：receipt 优先于 advisory selection，primary_blocked_todo_id 优先于列表展示顺序，并排除 credentials、production_access 以及已观测能力。它不会写 Goal lease、永久 grant、settlement 或 advancement checkpoint。

Python runtime_capability_reentry.py 现在只负责 effect runtime 调用、argv 渲染及映射到既有 agent/CLI channel；interaction_contract.py 只构建一次共享结果。这样 CLI/managed Turn 继续共用同一 projection。前端没有对应的 capability 配置控件或第二套状态，因此本 PR 无需新增 frontend companion；Lark 继续消费既有 capability-gate projection，本 PR也没有把 Lark live delivery 冒充为已验证。相关 PR #4283 未修改这些 owner symbols；若它先合并，仍需 rebase 后重跑 exact-head evidence。

具体改动

整条 PR 共 9 个文件、+234/-146：核心生产变化集中在 capability_gate.ts +67/-1；runtime_capability_reentry.py +73/-112 与 interaction_contract.py +16/-29 合计删除了重复 Python policy，并把 follow-up 的 channel adapter 从 1549 行的 interaction contract 移到现有 re-entry facade，使该文件降至 1528 行，修复 maintainability ratchet，而非放宽阈值。其余为 public CLI/installer 回归、TypeScript negative cases，以及中英文迁移 RFC。

关键代码讲解

• capability_gate.ts:116 projectRuntimeCapabilityReentry 是唯一决策 owner。它从结构化 binding/blocked candidate 构建候选，receipt 绑定优先，primary binding 决定优先级；无 eligible candidate 时返回 null。
• runtime_capability_reentry.py:19 build_runtime_capability_reentry_packet 是薄兼容层。repair_missing 为空时提前返回，避免健康路径产生 TypeScript hop；非空时调用 typed effect，并仅把 command_argv 渲染为 shell-safe command。
• runtime_capability_reentry.py:106 apply_agent_channel_projection 将同一 typed result 映射为非 settlement、非 advancement 的 verification action；失败后仍保留显式 selection 作为 fallback。
• interaction_contract.py:1417 build_interaction_contract 只计算一次 re-entry packet，再注入既有 agent/CLI channel，避免重复运行或第二数据源。

对主干的风险

最强反例是跨 runtime 漂移：TypeScript 决策正确，但 Python adapter 可能遗漏 channel 字段；另一个风险是错误地把 owner-held capability 当作 agent 可修复权限，或让 advisory selection 覆盖已提交 receipt。当前 head 用 exact-string typed arrays/objects 和明确的 receipt precedence 防住这些路径，没有 substring/prose classification。

验证覆盖：TypeScript capability gate 11/11；受影响 Python 267 tests；ruff、control-plane typecheck、git diff --check、maintainability ratchet、public/private boundary scan；detached base/head public CLI regression；install-local-smoke 单独在约 99 秒内通过，并以 300 秒 per-check timeout 重跑完整 premerge；远端 Sign-off、4 个 test shard、pytest、merge-gate、Sonar、Windows 与其余 required checks 全部通过。无 blocking finding。剩余风险主要是将来同时修改 TS owner 与 Python adapter 时的 contract drift，以及 #4283 先落主干后的集成变化；任何 head/base 变化都应重跑本结论。

我的整体评价

结论为 APPROVE。该变更修复了真实的 priority inversion，决策所有权与仓库架构一致，机制是纯派生投影且不扩张权限；follow-up 还实质性消除了 maintainability failure。用户入口上，CLI 与 managed Turn 已验证，generated /loopx 安装说明有覆盖；没有新增配置，所以无需前端改造，Lark 仍沿用共享 projection 而不是另建 source of truth。

由于这是 author-owned PR，GitHub 不允许正式 self-approval，所以我以 COMMENTED review 记录 approval conclusion；它不构成 merge authority，也不替代 owner review。

English verdict: APPROVE at exact head 409484d43517715380f35ccd26b5ecca323b0227; no blocking findings, all required checks pass, and the change keeps TypeScript as the single decision owner without granting durable authority. This is an author-owned commented conclusion, not merge authorization.